RLC Spark Execution Ledger

Append-only ledger. Checkpoint entries are written as the work lands and are never rewritten, including for style. A ledger you edit after the fact is not a ledger. New checkpoints go at the bottom.

Status: ACTIVE Source requirement: /Users/bryan/Downloads/Spark.txt (1,239 lines) Source SHA-256: b9caf67c7709b1b0d82fd7eb917c88e1afa17d514af0df17ca03ed600685237c Published base checkpoint: 4bdfd514 on main Published total checkpoint records before this ledger: 374

Completion rule

This ledger is the source of truth for the Spark/RLC expansion. A checked box means all acceptance criteria named by that item have implementation, focused tests, integration tests, and a durable proof artifact where the claim is empirical. Similar code, a configuration field, a helper with no live caller, or a successful prompt demonstration does not satisfy an item.

Labels used on unchecked items:

PARTIAL: a real component exists, but one or more named causal, live, or proof requirements are absent.
MISSING: the behavior is not implemented as a production path.
EVIDENCE: the mechanism may exist, but the required result is not proven.
BLOCKED: a prerequisite failed; the failure and next repair must be named.

Scientific boundaries:

An answer-channel or formatting gain is not a reasoning gain.
More tokens, more samples, or more wall time are not neural-architecture gains unless equal-compute controls and the preregistered interaction test isolate the mechanism.
A resident-32B mechanics pass is not a frontier capability result.
A probe that can decode a latent state is not proof that the state is causal.
A frontier claim requires fresh tasks, adequate power, no regressions, independently recomputable evidence, and external trust roots.
The commit name WOW Signal is reserved. It may be used only after the powered resident-32B certificate accepts the preregistered frontier or strong positive-interaction claim. It is never used for implementation alone.

Checkpoint accounting

The pre-Spark whole-project forecast was 394-661 total checkpoint records. This ledger reserves 72 additional total checkpoint records, one for each bounded checkpoint below. The revised forecast is therefore 466-733 total records. At record 374, checkpoint-count completion is 51.0%-80.3%; the midpoint planning estimate is 62.4%. These percentages measure the work ledger, not intelligence or release quality. Failed gates add explicit repair records rather than being hidden by changing a checkbox.

After CP321, the published total is record 382: 84-351 forecast records remain, checkpoint-count completion is 52.1%-82.0%, and the midpoint planning estimate is 63.7%. This remains workload accounting, not a capability score.

After CP322, the published total is record 383: 83-350 forecast records remain, checkpoint-count completion is 52.3%-82.2%, and the midpoint planning estimate is 63.9%.

After CP323, the published total is record 384: 82-349 forecast records remain, checkpoint-count completion is 52.4%-82.4%, and the midpoint planning estimate is 64.1%.

After CP324, the published total is record 385: 81-348 forecast records remain, checkpoint-count completion is 52.5%-82.6%, and the midpoint planning estimate is 64.2%.

After CP325, the published total is record 386: 80-347 forecast records remain, checkpoint-count completion is 52.7%-82.8%, and the midpoint planning estimate is 64.4%.

After CP326, the published total is record 387: 79-346 forecast records remain, checkpoint-count completion is 52.8%-83.0%, and the midpoint planning estimate is 64.6%.

After CP327, the published total is record 388: 78-345 forecast records remain, checkpoint-count completion is 52.9%-83.3%, and the midpoint planning estimate is 64.7%.

After CP328, the published total is record 389: 77-344 forecast records remain, checkpoint-count completion is 53.1%-83.5%, and the midpoint planning estimate is 64.9%.

After CP329, the published total is record 390: 76-343 forecast records remain, checkpoint-count completion is 53.2%-83.7%, and the midpoint planning estimate is 65.1%.

After CP330, the published total is record 391: 75-342 forecast records remain, checkpoint-count completion is 53.3%-83.9%, and the midpoint planning estimate is 65.2%.

After CP331, the published total is record 392: 74-341 forecast records remain, checkpoint-count completion is 53.5%-84.1%, and the midpoint planning estimate is 65.4%.

After CP332, the published total is record 393: 73-340 forecast records remain, checkpoint-count completion is 53.6%-84.3%, and the midpoint planning estimate is 65.6%.

After CP333, the published total is record 394: 72-339 forecast records remain, checkpoint-count completion is 53.8%-84.5%, and the midpoint planning estimate is 65.7%.

After CP334, the published total is record 395: 71-338 forecast records remain, checkpoint-count completion is 53.9%-84.8%, and the midpoint planning estimate is 65.9%.

After CP335, the published total is record 396: 70-337 forecast records remain, checkpoint-count completion is 54.0%-85.0%, and the midpoint planning estimate is 66.1%.

After CP336, the published total is record 397: 69-336 forecast records remain, checkpoint-count completion is 54.2%-85.2%, and the midpoint planning estimate is 66.2%.

After CP337, the published total is record 398: 68-335 forecast records remain, checkpoint-count completion is 54.3%-85.4%, and the midpoint planning estimate is 66.4%.

After CP338, the published total is record 399: 67-334 forecast records remain, checkpoint-count completion is 54.4%-85.6%, and the midpoint planning estimate is 66.6%.

After CP339, the published total is record 400: 66-333 forecast records remain, checkpoint-count completion is 54.6%-85.8%, and the midpoint planning estimate is 66.7%.

After CP340, the published total is record 401: 65-332 forecast records remain, checkpoint-count completion is 54.7%-86.1%, and the midpoint planning estimate is 66.9%.

Code-grounded baseline and ownership

The four-slice static audit covered the neural core, epistemic/verifier paths, training/proof surfaces, and the selected live/full-mind path. It read the entire Spark source and traced callers rather than crediting class names. The baseline below is exhaustive over SPARK-001 through SPARK-072:

ACCEPTED: SPARK-001, SPARK-005, SPARK-006, SPARK-007, SPARK-008, SPARK-009, SPARK-010, SPARK-011, SPARK-012, SPARK-014, SPARK-015, SPARK-016, SPARK-017, SPARK-018, SPARK-019, SPARK-020, SPARK-021, SPARK-022, SPARK-023, SPARK-024, SPARK-025, SPARK-026, SPARK-055.
PARTIAL: SPARK-003, SPARK-027, SPARK-035, SPARK-039, SPARK-040, SPARK-041, SPARK-042, SPARK-051, SPARK-052, SPARK-053, SPARK-054, SPARK-056, SPARK-058, SPARK-060, SPARK-062, SPARK-063, SPARK-065, SPARK-066, SPARK-067.
MISSING: SPARK-002, SPARK-028, SPARK-029, SPARK-030, SPARK-031, SPARK-032, SPARK-033, SPARK-034, SPARK-036, SPARK-037, SPARK-038, SPARK-043, SPARK-044, SPARK-045, SPARK-046, SPARK-047, SPARK-048, SPARK-049, SPARK-050, SPARK-057, SPARK-059, SPARK-061, SPARK-064, SPARK-068.
EVIDENCE: SPARK-004, SPARK-013, SPARK-070, SPARK-071.
BLOCKED: SPARK-069 is blocked on a successful source-bound admission preflight; SPARK-072 is blocked on the frontier-certificate verdict.

Every identifier appears exactly once in that map. PARTIAL means substantive code exists but the named acceptance contract is not closed. It does not count as half a checkbox.

Fable-lane record (agent coordination)

The four items the sequential march had skipped since CP318 were taken out-of-band by the Fable session (Bryan-directed) on 2026-07-23 and are now resolved: SPARK-004 at F1 (frozen baseline bundle), SPARK-013's pre-training half at F2 (state-causality instrument; remaining semantic acceptance binds to the SPARK-069 trained treatment), SPARK-003 at F3 (executable threat model), SPARK-002 at F4 (literature dossier).
SPARK-070's pre-training half is RESOLVED in the Fable lane at F5 (2026-07-23 22:15 PT): falsification_matrix.py is the typed, fail-closed registry over all twelve ledger rows — 9 runnable (bound to concrete executors over the existing experiment harness), 1 enforced (blind review, structural by construction, proven by threat-model checks), 3 blocked (verifier arms → SPARK-039-046; adversarial/OOD → generated fresh at acceptance; fast-weight controls → SPARK-055/056), each blocker named. tools/run_falsification_matrix.py produced the clean dry-run receipt on the untrained 1.5B baseline (372 episodes, zero degradations, receipt 8487e1ad… under artifacts/closeout/latent_cortex/spark070_falsification_matrix/, independently replayed). Runnable rows return CONJECTURE at this n on untrained weights (expected); the lesions_restorations row carries the F2 state-causality SUPPORTED structural claims. The dry run also caught a real live defect: CP328's sanctioned duplicate-role lesion arm was refused by CP331's correlation-evidence validation and silently degraded to a vanilla-decode fallback, voiding the diversity comparison — repaired in correlated_support.py (lesioned ensembles build unmeasured evidence over distinct roles under a dedicated bucket; preregistered evidence still cannot claim a lesion run). The post-training run on fresh held-out tasks (the acceptance event, checkbox stays [ ]) binds to the SPARK-069 treatment. The march owns everything else.
The pass-divergence design problem is RESOLVED in the Fable lane at F6 (2026-07-24, commit eb3735c7) — diagnosis, repaired instrument, one new mechanism, and a measured lever menu at artifacts/closeout/latent_cortex/pass_divergence_design/:
- The CP227 accuracy gate's verdict is VOID — _decode ran outside recurrence_adapter_scope, so both arms decoded the bare base model (on@d == off@d exactly, 6/2/0 both arms). Do not build on cp227_accuracy_gate/'s negative result; the repaired tool proves treatment activation per block and must be re-run on the CP227 adapter before any conclusion about it.
- cp305's uniform_partial_reward is a flat reward channel (baseline 0% at every depth, no_marker/token_limit on every episode at 320 tokens with cot) — Failure A, march-owned, blocks everything.
- Failure B: the campaign architecture cannot express pass-to-pass difference — wrap_depth_conditioned has no campaign call site (banks never attached; the forward seam is live), jitter/collapse-cos act on branches not passes, α constant, and the intrinsic levers (rotation_weight, anchor_injection) ran at 0.0 in CP227.
- Measured on the 1.5B (T=4): baseline increment alignment 0.75→0.94 (power-iteration capture); renormalize alone → 0.20/0.28; anchor 0.1–0.3 → 0.15/0.11; seeded inter-pass noise 0.05–0.1 → 0.06/0.01; step-operator deltas inert at 0.005 and anti-aligned (−0.27) at 0.02; all levers stacked → −0.47 (period-2 direction). Geometry only — no accuracy claim.
- New mechanism: RecurrentDepthPlan.interpass_noise (+noise_seed), a deterministic RMS-relative kick at re-entry only — T=1 stays bit-identical, same seed replays exactly. Per-sample seeds give GRPO groups latent-side variance, which is exactly what the uniform-reward diagnosis says the groups lack.
- The march's adoption menu (ordered, with configs) is in DIAGNOSIS_AND_MENU.md; step 1 is the free one: re-run the repaired accuracy gate on the existing CP227 adapter before any new training. The SPARK-069 admission preflight, campaign protocol, and any 32B run stay bound to the march. The Fable session still holds the live-runtime endurance forensics (the ~15-turn resident ceiling and the 4h soak memory slope) — outside the Spark checkpoints, no march collision.
F7 lane claim (2026-07-27, Bryan-directed acceleration). The Fable session claims the pre-training infrastructure legs that the march needs in place before a resident-32B campaign can promote or scale, and that require no model run to build:
- SPARK-064 — the permanent-distillation promotion transaction: versioned generations, a complete-by-declaration regression gate set (anti-interference, capability families, personality, tool honesty, authority safety, memory, frontier), and exact rollback with a proven byte-identical restore.
- SPARK-065 — the architecture meta-controller: measured expert/router/depth failure evidence, bounded proposals, isolated candidate evaluation, machine-checkable invariants, canary rollout, rollback, and an independent approval policy.
- SPARK-068's compact-proof half — replace the quadratic complete-prefix intervention envelope with a versioned checkpointed MMR inclusion witness that preserves independent replay. F7-G (2026-07-27) adds tools/spark_pretraining_preflight.py. Every leg above is a refusal surface, and a refusal surface that has been weakened — a threshold relaxed, a check commented out during a debugging session — still imports, still runs, and still returns a verdict. It just stops saying no, and that failure is invisible to anything except an attempt to make it say no. The preflight therefore does not check that the modules exist: it hands each instrument the exact input it must refuse and fails if the refusal does not come, the same idea as the campaign kernel probes. Eleven probes, all firing; --campaign DIR additionally validates any real promotion or STaR lineage found there. Exit 0 ready, 1 an instrument is dead, 2 an artifact failed validation. 23 tests, including ones that weaken a real instrument and require the preflight to notice — a probe suite that always passes is decoration.
F8 lane claim (2026-07-27, Bryan-directed: "everything ahead of Codex"). The F7 claim left SPARK-061 and SPARK-062 with the march. Bryan has since reassigned them: the CP419/CP420 lane is fully occupied removing the raw trainer bypass for SPARK-060, and 061/062 are the two objectives that must exist before that trainer has anything correct to optimize. Both are bare — no module, no test, no receipt — so this is new construction, not a parallel re-implementation. The Fable session claims:
- SPARK-061 — the progressive recurrent objective, and specifically the instrument that decides whether "later states improve" is real. v4's own docstring records the trap: on families where depth is destructive the monotone hinge is unwinnable, and the cheapest way for an optimizer to satisfy a constraint it cannot win is to drive the recurrent transform toward the identity. An objective that trains recurrence inert while its loss curve descends is the single most expensive failure available to the 32B run, and nothing currently detects it.
- SPARK-062 — the auxiliary-objective composite and depth curriculum: seven declared terms, each required to be live rather than merely weighted, with stage advancement bound to measured competence and train/inference depth parity carried through to the inference config. This lane runs no model of consequence, opens no holdout custody, and grants no capability claim.
The march keeps SPARK-060/063/069 and everything the CP419/CP420 lane is touching (core/learning/verified_transition_episode.py, core/runtime/resource_observation.py, core/brain/llm/latent_cortex/ frontier_tasks.py, tools/independent_paired_campaign_scoring.py, tools/lint_governance.py). The Fable lane does not run a model, does not open holdout custody, and grants no capability claim.

Checkpoints	Primary implementation owners	Live integration owners
SPARK-001-004, SPARK-069-072	`docs/RLC_SPARK_EXECUTION_LEDGER.md`, `tools/prepare_resident_recurrent_grpo_campaign.py`, `core/brain/llm/latent_cortex/frontier_certification.py`, proof/lab tools	installed-app resident worker, sealed artifact and external-verifier paths
SPARK-005-013	new strict state module under `core/brain/llm/latent_cortex/`, epistemic firewall, memory/evidence stores	`core/brain/cognitive_ingress.py`, memory consolidation, full-mind receipt
SPARK-014-022	`branches.py`, `workspace.py`, branch operators and diversity/correlation modules	latent engine, compute ledger, GWT coupling
SPARK-023-038	`engine.py`, `recurrence.py`, `schedules.py`, `escape.py`, HLA/state-tree/search/quanta modules	resident MLX worker/client and `latent_cortex_service.py`
SPARK-039-050	unified verifier mesh, `task_verifiers.py`, exact reasoning tools, process/generative critics	in-episode RLC controller and governed tool receipts
SPARK-051-054	value-of-computation controller, compute/evidence ledgers	service allocation, body/allostasis, Will, user-facing response synthesis
SPARK-055-064	fast weights, recurrence adapter, replay/consolidation, `tools/train_grpo.py` and proof trainers	worker adapter loader, scheduled learning transaction, runtime integrity producer
SPARK-065-068	cognitive ingress, RLC-GWT coupling, self/body/memory bridges	response generation, action executor, Will re-decision, health/API/UI receipts

Static audit verdict

The selected RLC path is real neural computation: resident hidden states enter a latent workspace, repeatedly traverse shared middle transformer layers under KV rewind, and the winning state is persisted into decode. The following prevent a claim of one unified, recurrence-trained, live Aura runtime:

The recurrence-native adapter implementation has no resident worker loader or live projection attachment; the active generic depth loop is not proof of trained RLC tissue.
weight_integrity has a schema and consumers but no live digest producer. Routine success and public full-mind proof do not yet require a proven verdict, active recurrence adapter, or adapter artifact digest.
Runtime identity inventories more than it validates: adapter weights, tokenizer, quantization, and unresolved identity gaps are not all compared against pinned expectations.
The four-slot resident profile reserves communication/free slots and can admit only one ordered organ context, so memory/reference can crowd out goals, self, affect, body, Will, and GWT.
RLC enters GWT after reasoning, but current action execution asks Will before RLC rehearsal and does not require a new authorization when rehearsal changes the risk/effect model.
Tools, symbolic engines, independent critics, disagreement localization, and targeted evidence acquisition are outside the recurrent episode. Selecting RLC can also bypass the ordinary reasoning amplifier/composer.
Fast-weight consolidation produces proposals but is not a scheduled, interference-gated durable-learning transaction. RLC proof lineage is not retained with normal memory consolidation.
Public health/UI receipts omit decisive mechanism, integrity, identity, cognitive-slot, telemetry, and workspace-outcome fields.

The dependency order is therefore integrity and live adapter identity, full epistemic state, independent hypotheses, recurrent correction/HLA, verifier and local repair, adaptive control, verified learning, whole-organism re-decision, then the powered resident proof. Training a stale or disconnected treatment before those dependencies close is not admissible.

A. Contract, baseline, and literature

SPARK-001 - Canonical requirements ledger. Map every distinct Spark mechanism, limit, training objective, live seam, and falsification gate to an owner and acceptance contract; reconcile it with the main Aura tracker.
SPARK-002 - Primary-literature dossier. Replace placeholder citations with a versioned bibliography of primary papers/specifications, distinguish replicated findings from proposals, record licenses, and bind source hashes in the final methods package. Accepted at F4 (Fable lane, 2026-07-23): literature.py is the versioned typed dossier (2026.07.23.1, 22 entries, registry digest 4957567e…) grounding all twenty required mechanism families — self-consistency, process rewards, verifier training, process-vs-outcome, STaR/Quiet-STaR, continuous latent thought, recurrent depth, adaptive computation, tree search, UCT, self-correction limits (the Huang et al. negative result is first-class), mistake location, RL self-correction, sycophancy, iterative refinement, test-time compute, GRPO, RLVR, fast weights, and LoRA — each entry with explicit replicated/reported/proposal status, declared license, and the SPARK items standing on it. Seven less-common arXiv identifiers were verified against arxiv.org during authoring (two corrections landed: Uesato author list, Coconut venue). docs/RLC_SPARK_LITERATURE.md is generated by tools/render_spark_literature.py and a test pins the committed doc byte-exact to the registry render; validation fails closed on duplicate ids, malformed identifiers, dangling SPARK references, or mechanism-coverage gaps. PDF byte hashes and per-PDF license text bind at SPARK-072 methods-package assembly against these immutable identifiers, as the dossier states. Focused tests 6/6.
SPARK-003 - Failure and threat model. Enumerate anchoring, verifier collusion, fake branch diversity, reward hacking, answer leakage, right-to- wrong correction, context contamination, state corruption, budget abuse, stale tools, adaptation leakage, and unsafe self-modification with executable mitigations. Accepted at F3 (Fable lane, 2026-07-23): threat_model.py is a typed, fail-closed registry binding all thirteen named threat classes to their concrete mitigating modules and to 34 exact suite tests that prove each mitigation fires; validate_threat_model rejects a missing threat class, a moved mitigation module, or a check that no longer exists, so the model cannot rot silently. Every entry carries an explicit residual-risk line (reward-side RLVR/EIR remains SPARK-060's property; learned-verifier base-model correlation remains SPARK-041..046's; organism-wide self-modification authority stays with the Will/governance stack). All 34 bound checks pass in this checkout (262/262 across the fifteen mitigation suites, 146.99s), the registry meta-suite passes 6/6, and smoke stays 104/104.
SPARK-004 - Frozen baseline bundle. Freeze resident checkpoint, tokenizer, adapters, decoding, task generators, control manifests, resource envelope, randomization, and current vanilla/RLC measurements before changing the treatment. Accepted at F1 (Fable lane, 2026-07-23): frozen_baseline.py builds and independently verifies one immutable, hash-bound, Ed25519-signed bundle; tools/freeze_spark_baseline.py freeze produced the real sealed bundle at artifacts/closeout/latent_cortex/spark004_frozen_baseline/ (certificate 9bf377e5adc7cf9d…, commit-bound at e82e031ca). It binds the fused resident checkpoint by full-content fingerprint (8eae71e73a14d122…, 4 weight files, re-hashed twice), the tokenizer/config behavior bundle, the personality-adapter claim (fused, none attached), the RLC execution spec plus worker decode defaults, task-generator registry 2026.07.18.2 with commit+hash-bound generator/randomization sources, every config/latent_cortex control manifest as bundle-internal copies, the cp305 declared resource envelope, the preregistered training/eval seeds, and the current measurements (cp227 intrinsic accuracy gate on/off receipts as the paired vanilla/RLC evidence, cp305 controller verdict and GRPO receipt as the latest treatment-side outcome). Verification re-checks the exact artifact set, per-file digests, certificate self-digest, and detached signature against a caller-supplied trust anchor only; drift checks re-hash the checkpoint and sources on demand. 28 focused tests pass; smoke 104/104. Honest limits: the signature roots in the local contamination-audit key (local custody, not external), and SPARK-069 admission must pin this certificate hash for the freeze to gate anything.

B. Persistent epistemic state

SPARK-005 - Typed epistemic state. Implement a strict, bounded schema for immutable problem evidence, hypotheses, claim graph, observations, uncertainty, attempted operations, budgets, and accepted answer state. Accepted at CP318: epistemic_state.py provides the immutable, canonically serialized, content-addressed schema and stale-safe atomic transaction authority. This credit does not include persistence, live wiring, calibrated uncertainty, descendant invalidation, or state-causality proof; those remain independently unchecked below.
SPARK-006 - Hypothesis portfolio. Preserve multiple weighted hypotheses without premature collapse; support minority survival and explicit unresolved status. Accepted at CP323: epistemic-state schema v4 treats non-refuted hypothesis point estimates as one normalized portfolio, enforces a protected minority floor, bounds explicit minority status, permits at most one uniquely highest favored hypothesis, and retains active/unresolved alternatives. Refutation requires a linked rejected/contradicted claim and a zero interval; revival is forbidden while that refuting claim remains blocked. Every post-initial revision or addition is a complete, atomic, budgeted comparison operation naming every changed hypothesis and evidence receipt. Identity/claim-scope rewrites, deletion, forged references, and unreceipted journal transitions fail closed, while valid revisions recover byte-equivalently. Focused state, calibration, and journal contracts pass 66/66; the integrated RLC/latent/GWT/ controller gate passes 796/796. Correlated-support discount, hypothesis- distribution calibration, live state causality, and capability gains remain separate open checkpoints.
SPARK-007 - Claim/dependency graph. Represent premises, claims, support, contradiction, descendants, failure conditions, and answer dependencies with cycle and consistency checks. Accepted at CP319: premise/contradiction graphs are bounded and cycle-safe; established status is dependency-consistent; descendant invalidation is transitive, atomic, operation-receipted, budgeted, and revokes affected hypotheses and answers. Evidence semantics and durable recovery remain owned by SPARK-008 and SPARK-011 rather than being implied here.
SPARK-008 - Evidence ledger. Bind every tool result, retrieval, calculation, proof, simulation, and observation to provenance, freshness, scope, content digest, and the claims it can actually support. Accepted at CP321: epistemic-state schema v2 binds every evidence record to typed producer/version/invocation/receipt provenance, explicit verification class, episode/objective/claim scope, purpose, content digest, and bounded validity time. Independent verification requires a distinct verifier identity and receipt; memory and unverified context cannot acquire claim authority. Claim/evidence links are exact and bidirectional. Answer acceptance requires fresh evidence for the complete transitive premise closure and rejects omitted, future, stale, or unresolved counterevidence. The governed journal persists these fields inside the same hash-chained transactional state. Focused state and crash-recovery contracts passed 37/37; the integrated RLC/latent/GWT/ controller gate passed 767/767. This is a structural evidence-integrity result, not calibrated uncertainty, live state causality, or a capability-gain claim.
SPARK-009 - Calibrated uncertainty. Maintain claim-level epistemic uncertainty from measured signals rather than verbal confidence; report calibration error and abstain outside validated support. Accepted at CP322: immutable, domain- and estimator-specific calibration profiles are fit only from independently graded held-out outcomes with unique prediction/outcome receipts and pinned dataset/split digests. Profiles report Brier score, constant-predictor baseline, ECE, MCE, reliability bins, Wilson bounds, admission failures, and expiration. Claim intervals are recomputed from the registered profile and measured signal evidence; sparse bins, profile drift/failure, low lower bounds, future evidence, and expired profiles force explicit abstention. Exact confidence requires independently verified proof or calculation evidence. Established claims cannot remain uncalibrated, and answer confidence equals the weakest transitive dependency rather than a generated number. Profiles and decisions live in the same canonical, hash-chained transaction and cannot be rewritten during recovery. Focused calibration/state/journal contracts passed 50/50; the integrated RLC/latent/ GWT/controller gate passed 780/780. External trust-root authenticity, live state causality, hypothesis-distribution calibration, and frontier gains remain separate checkpoints.
SPARK-010 - Cognitive operation history. Record which operators were attempted, their inputs, costs, evidence gained, affected claims, and outcome so recurrence cannot unknowingly repeat failed work. State authority completed at CP324: schema v5 records immutable operator/version identity, canonical payload and referenced inputs, attempt signature, parent state, start/completion time, affected claims/hypotheses, gained evidence, actual cost, outcome, failure code, and explicit retry parent. Attempt signatures remain stable across state versions. Duplicate roots, orphan/stale/changed-input/overlapping retries, retry forks, retries after success, and chains beyond three attempts fail before mutation. A typed admission query reports new, explicit-retry-required, stale-parent, already-succeeded, or retry-exhausted decisions. Compute use must exactly equal immutable operation costs, and journal replay preserves failed/unknown work. Focused contracts pass 71/71 and the integrated gate passes 801/801. Accepted at CP326: the primary foreground controller now consults that admission authority before worker compute even when the live response path supplies decode overrides. Structural experiment overrides remain exact and intentionally opt out. A crash-consistent runtime lease fsync-journals a zero-cost UNKNOWN intent before execution; normal completion is an explicit retry carrying terminal outcome and measured token-layer cost. Recovery can resume one pending intent but cannot silently rerun a completed attempt. Objective, config, budget, controller evidence, operation kind, exact state and input references, payload, attempt, and operation ID are recomputed at service, MLX client, and worker boundaries and included in the request identity. State-less ingress degradation receives an objective-bound genesis rather than bypassing history. Focused runtime/wire contracts pass 33/33 and the integrated RLC/latent/GWT/controller gate passes 837/837. SPARK-051 still owns learned per-recurrence selection across the complete action vocabulary; SPARK-013 and SPARK-066-068 still own lesion and full-live causal evidence.
SPARK-011 - Transactional state revision. Apply revisions atomically, hash lineage, reject malformed/stale updates, invalidate descendants, and recover the last verified state after interruption. Accepted at CP320: the state machine publishes only after a governed, fsync-sealed, canonical hash-chain append; replay is anchored to the external genesis, enforces immutable history and direct parentage, rejects stale writers and complete corruption, and repairs only a torn final fragment after the entire complete prefix verifies. Live ownership is tracked separately by SPARK-013 and SPARK-065-068.
SPARK-012 - Selective memory bridge. Query working, episodic, semantic, procedural, and nonparametric memory through evidence-scoped retrieval; prevent recalled text from becoming instruction or fact merely by appearing in context. Accepted at CP325: one typed aura.rlc.selective_memory.v1 contract queries all five live memory surfaces concurrently under bounded per-source timeout and records success, empty, unavailable, failure, and timeout receipts rather than erasing partial failure. Retrieval is bound to tenant, user, session, episode, immutable objective digest, expiry, and contest status. Deterministic ranking, cross-tier deduplication, corroborating-tier attribution, and one-per-tier fairness construct a bounded result without turning frequency into truth. Every admitted memory becomes context-only evidence and a durable SEARCH_MEMORY operation in the same transactional epistemic state. Recalled text always carries instruction_authority=false; exact text, digest, tier, scope, source receipt, result hash, and state hash are validated independently at service, client, worker, and engine boundaries. Memory-shaped prompt injection, cross-tenant/user/session/episode replay, stale or contested records, boolean timestamp coercion, reserved-field smuggling, authority mutation, and envelope tampering fail before latent execution. The focused bridge contracts pass 19/19 and the integrated RLC/latent/GWT/controller gate passes 823/823. Live capability effects and lesions remain separately owned by SPARK-013 and SPARK-066-068.
SPARK-013 - State causality. Prove the structured state changes later computation and that removing required state causes task-appropriate loss; prohibit a prose-only shadow ledger. Progress at F2 (Fable lane, 2026-07-23): the structural half is closed and the semantic half is a preregistered instrument. state_causality.py adds the only sanctioned projection from typed EpistemicState evidence into the slot-embedded cognitive context: every injected text must hash to the record's content_sha256, the prose summary is never consulted, and content without a typed ancestor is refused — the prose-shadow prohibition is executable, not narrative. A seven-arm, fixed-depth, width-matched experiment on the real pretrained 1.5B (24 tasks, 168 episodes, receipt d052817c… under artifacts/closeout/latent_cortex/ spark013_state_causality/, independent row-level replay) grades five structural claims SUPPORTED at verdict n: information-lesion of the required evidence changes the final latent state under exact recurrence parity; a non-projected state component is inert byte-exactly; restoration recovers byte-exactly; prose-summary mutation changes the state hash but leaves computation byte-identical; content substitution moves the final state. The same identities hold on the random tiny-qwen seam test (13/13 focused tests). The task-appropriate-loss claim is honestly CONJECTURE: the untrained pooled-slot channel cannot yet carry the fact (intact accuracy 0.00 at the 0.5 readability floor), so no loss is measurable before slot-channel training. What remains for acceptance: re-run tools/run_state_causality_semantic.py against the SPARK-069 trained treatment (and the live resident seam under SPARK-066-068) and obtain the PROVEN task-appropriate-loss verdict the preregistered grading already encodes.

C. Independent hypotheses and virtual width

SPARK-014 - Fresh-context branch isolation. Generate candidates from the original problem in independent contexts before exposing any branch to another answer; test for context, cache, RNG, and hidden-state contamination. Accepted at CP328: every branch now advances from the same content-addressed original context in a distinct workspace and deterministic role RNG stream. Cross-branch exchange, consensus compression, and diversity comparison are blocked until every branch has produced a content-addressed latent candidate after the configured isolation floor. Explicit schedule exchanges cannot bypass the floor. The first exchange step, blocked attempts, seed/candidate/ RNG/context commitments, state-alias check, and role-lesion status are carried in one exact receipt. Every speculative shared-cache pass restores the exact immutable K/V tensor objects and metadata and checks that postcondition before another branch can run. The service independently validates cardinality, uniqueness, common original context, cache counts, and exchange timing. Deliberate repeated-role lesion arms may run but cannot claim certified branch independence. Focused branch/engine/wiring contracts pass 150/150; the fixed ownership gate passes 857/857, and forged receipt checks pass 1/1. This closes isolation mechanics, not differentiated operator labor or capability gain.
SPARK-015 - Distinct cognitive operators. Implement direct, constructive, counterexample, inverse, causal simulation, formal, analogy- mapping, assumption-removal, and boundary-case operators as different executable policies, not labels on equivalent prompts. Accepted at CP329: the recurrent engine now owns nine versioned latent programs with distinct state-transition mathematics: single control write, progressive scaffold, hypothesis sign reversal, reverse slot transport, finite-difference rollout, control-axis projection, paired-relation transport, maximum-alignment subtraction, and signed boundary extrapolation. Every live branch executes the program bound to its role before recurrence; BLIND_RESOLVE and BRANCH no longer bypass operator control, and tokenizer-free runs receive deterministic hidden-space controls. Programs use different slot targets, anchor relationships, and strengths, while exact RMS bounds and protected context/evidence slots constrain every update.

Each execution carries content-addressed input/output/anchor/control digests, changed and protected slots, role, operator, action, step, branch, transform, and causal verdict. The service validates program identity and complete, unique per-branch coverage for every neural action and rejects malformed, duplicate, orphaned, structural-action, or non-causal claims. Operator identity is part of savepoints and escape role shifts. Identical-input tests prove all nine programs produce nine distinct outputs and transforms; the integrated operator/engine/service/escape/journal suite passes 187/187 and the fixed ownership gate passes 861/861. This closes executable distinction, not task-level superiority, structural-independence scoring, or frontier gain.
SPARK-016 - Structural diversity measurement. CP330 adds the service-reconstructible aura.rlc.structural_diversity.v1 receipt over the primary cognitive-slot, action, operator, and isolation traces. Every branch is fingerprinted across the exact six required facets: admitted premise commitments and usage modes; dependency templates and changed-slot edges; executable action/operator/transform algorithms; observed state-transition topology; predicted program consequences; and program plus runtime failure conditions. Attractor-escape role changes are preserved as ordered role and operator paths rather than flattened or rejected.

Surface text is not an input to the measurement. Canonical support classes therefore collapse identical causal programs even when candidate state commitments or prose differ, and pairwise independence requires differences in dependencies and algorithms plus at least one state, prediction, or failure facet. The service independently rebuilds the full receipt and rejects count, class, digest, facet, or wording-policy tampering. Focused structural, operator, engine, wiring, and escape tests pass 24/24; the integrated affected suite passes 191/191; and the fixed ownership gate passes 866/866. This closes structural measurement, not empirical error-correlation estimation, vote weighting, or task-level gain, which remain SPARK-017 and later proof work.
SPARK-017 - Correlated-support discount. CP331 adds an independently checked, task-keyed branch-outcome ledger and pairwise error estimator. It computes the binary phi coefficient, applies a 24-observation shrinkage prior, ignores negative correlation for discounting, and cannot apply an empirical penalty below 12 paired checked tasks. Replayed task identities, incomplete role outcomes, ungraded rows, malformed estimates, and snapshot tampering are rejected.

Duplicate executable programs receive fractional exchange weights immediately, even before calibration. Once powered evidence exists, positive error correlation also lowers each branch's support weight. Those weights are installed before cross-branch exchange and change the neural consensus, not merely telemetry. The final receipt combines empirical dependence with CP330's structural distance, collapses duplicate support classes, reports an effective independent-support count and confidence multiplier, and is exactly reconstructed by the service. Candidate wording and raw state differences do not create votes. Focused correlation, durability, engine, service, and causal exchange tests pass 10/10; the affected integration suite passes 199/199; the fixed ownership gate passes 873/873. A fresh deployment truthfully starts in bootstrap_unmeasured; it discounts exact duplicates but does not invent empirical correlation before checked outcomes accrue.
SPARK-018 - Blind role-separated review. CP332 decodes every branch candidate before review, then evaluates the closed batch in a deterministic derangement that never places a branch in its original index position. Reviewer callables receive candidate text only: branch, role, operator, selected-branch, first-answer, ownership, doubt-prompt, and even anonymous-ID metadata remain outside the callable boundary. Review copies remove explicit first/previous-answer, Aura-ownership, numbered-branch, requested-critique, and doubt-prompt cues without rewriting the final answer candidates.

A content-addressed receipt binds the objective, certified fresh-context isolation evidence, review order, private branch mapping, candidate and review-text commitments, redaction counts, and mapped scores. The service validates the exact visible/forbidden-field policy, objective and isolation commitments, derangement, complete unique mapping, finite rounded scores, row structure, and receipt digest. Tests reject origin-order, score, objective, and policy tampering. Origin-only candidates become an explicit non-substantive review sentinel rather than aborting the RLC episode. The final fixed ownership set passes 877/877 and focused verifier/service regressions pass 27/27. This accepts origin blindness by construction and validation; it does not prove critic independence, balanced-error detection, task accuracy, or frontier gain.
SPARK-019 - Decoy-balanced verification. CP333 gives verifier authority two explicit gates. Before any verifier-dependent recurrent action, a preflight presents per-episode seeded correct, incorrect, and byte-identical unchanged-twin controls covering exact arithmetic and Python syntax. Correct must beat incorrect by a fixed margin, unchanged twins must receive equal scores within tolerance, and every score must be finite and in [0,1]. Failure leaves recurrence verifier-free and is disclosed.

At branch selection, Aura decodes the complete candidate set and interleaves it with a fresh control set in one deterministic, content-addressed batch. The callable sees text only; candidate/control class, correctness label, branch, role, ownership, and order metadata are withheld during review. The candidate subsequence remains deranged. Only a second passing calibration admits reviewer scores for branch selection and later latent optimization or fast-weight verification. A constant, unstable, out-of-range, or throwing reviewer loses authority; the latent episode continues by convergence rather than collapsing to vanilla fallback.

The service reconstructs both control sets, hidden-label orders, hashes, margins, repeat spread, calibration verdicts, candidate mapping, and admitted winner. Synthetic evaluations are excluded from the task-verifier quality receipt so canaries cannot inflate evidence. The affected RLC integration slice passes 219/219 and the engine/branch/wiring slice passes 193/193 before the final bounded-score regression. Final fixed-snapshot ownership evidence is recorded in CP333 below. This closes decoy balancing for the live reviewer boundary, not domain-general critic validity or function/weight independence; those remain SPARK-020.
SPARK-020 - Disjoint critic function. CP334 makes the worker's deterministic symbolic critic prove its identity before it can affect a recurrent decision. The receipt commits to the complete declared source closure, exact per-file hashes, parsed import graph, forbidden neural-runtime dependency audit, class identity, and a runtime object-state audit. Only the four bounded data fields are admitted; callable/tensor/model state is rejected and the critic has exactly zero trainable parameters. That identity is bound against the resident generator's model-path commitment, logical and stored parameter counts, worker source, adapter stack, tokenizer, quantization, and known identity gaps. The service independently reconstructs both sides.

A governed durable ledger now accepts unique independently checked task/candidate outcomes keyed to those exact generator and critic functions. Its snapshot binds the checked sample set and distinct grader identities, reconstructs all four confusion-matrix cells, and reports the generator-error shared-blind-spot rate with a 95% Wilson interval. It requires 24 checked samples, eight generator errors, and two distinct graders before reporting a measured state. A powered upper bound above 0.35 causally revokes critic authority. Before power, the receipt says bootstrap_unmeasured; live decoy gates still apply, but no low residual is invented. Worker and service reject identity, sample, aggregate, function-binding, or authority tampering.

Focused identity, evidence, ledger, causal-revocation, blind-review, task-verifier, and service contracts pass 116/116. The final fixed-snapshot RLC/GWT/execution-controller ownership gate passes 895/895. This closes the disjoint implementation and live residual-meter mechanism, not a claim that the currently empty live ledger has measured a low shared-blind-spot rate, and not a resident-32B intelligence or frontier result.
SPARK-021 - Causal branch exchange. CP335 replaces the count-only mailbox with a strict hidden-state protocol. Each sender emits exactly one latent slot derived from at most 16 private reasoning slots. The mailbox and typed organ/context slots are excluded, no decoded text enters the channel, and the receipt binds the sealed candidate, role, executable operator, step, source/message/state commitments, support and consensus weights, and every recipient's causal pre/post write. Only the first non-lesioned generation can count as independent support; later exchanges explicitly carry possible peer context and are cooperative refinement, never another vote.

Interval, schedule-bytecode, and controller-compare exchanges now carry single-use synchronization identities. The service reconstructs each point against recurrent step divisibility, successful bytecode events, or the exact successful controller transition and rejects omitted, replayed, reordered, policy-changed, candidate-changed, or noncausal traces. Raw tensor element accounting is emitted now for SPARK-022; it is not yet promoted into an equal-compute capability claim.

The same source policy now executes inside the differentiable recurrence-native objective. The versioned execution spec binds the policy and 16-slot ceiling, so an adapter trained under the earlier all-slot mean cannot claim exact train/live parity and must be retrained or revalidated. Tensor tests lesion roles, swap the role/operator programs across fixed branch indices, and restore the original assignment exactly. Experiment R adds a restoration arm, exact task-set commitment, per-task compute parity, and a combined causal verdict that remains CONJECTURE if lesion, restoration, swap parity, or compute parity is absent. This proves differentiated neural execution and its falsification path, not broad task accuracy or frontier benefit. The affected engine/service/training/GRPO/identity suite passes 312/312, and the final fixed ownership gate passes 900/900.
SPARK-022 - Equal-compute virtual width. Meter every branch, exchange, verifier, and decode operation in the same currency as controls; refuse comparisons with hidden compute or information advantages. Accepted at CP336. A model-profile-bound structural estimator now records transformer layer applications, attention query/key pairs, output-head work, tensor reads/writes/scalar operations, verifier bytes/calls, tool I/O, external-model tokens/calls, and host scalar work as exact non-negative integer counters. Every operation is separately named and digest-bound; unknown work makes the receipt incomplete. The estimator recomputes dense decoder/GQA FLOPs from validated architecture fields rather than trusting a caller-supplied total.

Information parity is separate and equally strict. Each arm commits the exact tokenized model input, typed cognitive context, controller evidence, verifier, tokenizer/decode policy, and tool-access policy. Source IDs, kinds, byte/token counts, content hashes, policy hashes, and unknown accesses are normalized and receipt-bound. Vanilla and recurrent arms now consume the same chat-template token IDs; actual generated tokens determine decode work instead of a declared maximum. Equal-compute controls accumulate complete samples until neural FLOPs and every non-neural parity dimension reach the preregistered target, and fail closed if the sample bound cannot do so.

Branch recurrence, bounded mailbox exchange, cognitive operators, compression, disagreement/diversity, latent optimization, fast weights, verifier use, sampling/voting, and cleanup all charge the same ledger. Worker results bind their top-level receipts byte-for-byte to the RLC episode; the service independently reconstructs live operation totals. Campaign grading, frontier certification, raw artifacts, and virtual-width experiments issue per-task comparison certificates and force claims to CONJECTURE on missing, unequal, or hidden work or information. A separately implemented standard-library scorer reconstructs the profiles, ledgers, information receipts, RLC binding, and certificates and byte-matches the production semantic grade. Rehashed inner-ledger tampering is rejected after outer commitments are recomputed.

This accepts one structural comparison currency, not hardware energy or wall-time equivalence, and does not make differently funded primary arms an equal-compute claim. No resident-32B capability campaign was run; task benefit, positive interaction, and frontier gain remain unproven.

D. Penultimate recurrent neural correction

E. Verifier mesh and local repair

F. Adaptive compute and control

SPARK-051 - Value-of-computation controller. Select among DECOMPOSE, BLIND_RESOLVE, BRANCH, SEARCH_MEMORY, RETRIEVE_EVIDENCE, EXECUTE, SIMULATE, FALSIFY, CHECK_ASSUMPTION, REGENERATE_FROM_PREFIX, FORMALIZE, COMPARE, BACKTRACK, COMPRESS_STATE, ANSWER, and ABSTAIN using measured expected gain per cost. CP326 makes the existing bounded execution-arm decision a state-admitted, costed, durable live operation and removes the decode-override bypass. This remains PARTIAL. CP327 adds the exact sixteen-action vocabulary and makes a deterministic value-of-computation decision before every recurrent window. Each decision binds the complete cognitive state signal, declared executor inventory, immutable evidence snapshot, gain/cost estimates, reason, and state transition. A hard neural floor prevents compare/stop actions from consuming an episode before recurrence. Measured cells use held-out lower gain and upper cost bounds; sparse cells are explicitly bootstrap exploration, never mislabeled as calibrated evidence. The selected action changes the live branch workspaces or halt state and is independently reconstructed by the service. Every action becomes an atomic child operation in the canonical journal without double-charging the enclosing episode. The worker cannot claim external EXECUTE authority, and unavailable actions fail closed.

CP369 removes one remaining synonym failure. SEARCH_MEMORY and RETRIEVE_EVIDENCE no longer differ only by an embedded instruction passed through the same generic branch operator. Each selected action first reads only its matching immutable context class and writes a bounded summary into each live branch's communication slot. Episodic/durable memory, one-shot nonparametric memory, offline reference, world-model evidence, and governed tool/evidence sources have explicit source-class rules; the one-shot source truthfully participates as both memory provenance and typed evidence.

Every focus transition commits its source slots and labels, input/source/ output tensors, preserved-source tensor, branch/step identity, bounded strength, and exact tensor accounting. Source slots must remain unchanged, the output must differ from the input, and the following branch-specific cognitive operator must consume the focus output exactly. The service independently reconstructs source selection from the public cognitive-slot inventory, verifies branch/step coverage and focus-to-operator chaining, and matches every tensor-work total. The receipt explicitly says external_retrieval_effect=none; this slice cannot claim a fresh fetch.

Seven direct focus contracts plus two forced-controller tiny-Qwen episodes cover source taxonomy, causal/preserved state, distinct memory/evidence outputs, absent-source refusal, inventory tampering, false external-fetch claims, one-shot hybrid classification, service reconstruction, and focus-to-operator lineage. Focused action/controller/branch/resource/service coverage passes 251 tests; the complete combined latent/RLC ownership battery passes 1,252 tests.

CP370 adds one bounded service-side acquisition continuation for a validated SEARCH_MEMORY or RETRIEVE_EVIDENCE transition. The worker remains compute-only. The host binds the objective, query, selected transition, source inventory, one-attempt cap, and one-continuation cap; only a genuinely new typed memory or offline-reference observation can buy a second recurrent episode. Empty, unavailable, duplicate, and wrong-source outcomes remain distinct and cannot mint compute. The final receipt commits both episodes, the acquisition result, the returned round, and exhausted caps.

CP372 adds the missing governed external EXECUTE protocol. The worker sees only a bounded digest-bound offer and can request, but cannot dispatch, an already-Will-admitted effect. The host reconstructs the action trace, readiness decision, action-policy evidence, successful epistemic operation, and exact authority before one durable coordinator advances PREPARED -> DECIDED -> DISPATCHING -> terminal. Dispatch and task leases prevent duplicate owners; a crash after dispatch becomes UNKNOWN_EFFECT and requires reconciliation rather than blind replay. Post-action receipts are deterministic, persisted, contract-equal to the staged recovery record, and required before success can be linked. Exact verified durable evidence can reconcile an unknown effect without re-executing it.

The completed runtime-operation receipt now carries the admitted and final epistemic states, every action operation, measured compute, and a journal extension anchored to the authority's admitted head and entry count. The independent verifier reconstructs the final state and rejects fabricated authority, failed EXECUTE operations, missing or contradictory post-action evidence, secret-bearing replay containers, malformed abandonment markers, and self-consistent forged journal prefixes. An independent code review returned ACCEPT with no P0-P2 findings after the final anchor repair.

CP373 implements the calibration-v2 evidence system without manufacturing a result. All sixteen actions must receive at least eight globally unique, multi-domain paired tasks (256 total treatment/control executions) before an acquisition certificate can exist. A cell remains explicitly unmeasured until it reaches 20 unique pairs (640 total executions for complete coverage). Treatment and matched no-action control begin from an identical externally captured, runner-signed checkpoint/KV/latent/evidence/memory/RNG state whose classifier and state-bucket evidence are also committed, run under fixed continuation and budget policies, and preserve complete available versus consumed information receipts.

Claim-eligible tasks are reblinded with distinct 256-bit external-issuer nonces rather than the registry's reproducible seed-derived blind. A blind-independent task identity prevents reblinding duplicates from increasing effective sample size, and all actions receive the same committed stratum distribution under external randomized assignment. The externally rooted policy separates issuer, runner, contamination auditor, and evidence verifier custody. Every output is globally sealed before hidden answers can be revealed. The final candidate embeds the canonical append-only journal transcript; independent replay verifies every event hash, predecessor, state transition, attempt, pre-reveal seal prefix, manifest reference, final head, byte size, result, verification, and commit. It reconstructs every observation and statistic from the journal rather than trusting relabelable candidate fields. A signed final-verifier candidate/cell commitment is rechecked against the separately configured root inside the worker before measured evidence can affect action selection. Live admission checks current policy freshness, while immutable historical replay uses the committed admission time.

Gain bounds use one exact simultaneous 34-family multiplicity budget. Cost uses a conservative Hoeffding upper bound over the maximum normalized fraction of every preregistered action-resource cap: structural FLOPs, transformer/attention/head work, tensor traffic and scalar operations, verifier work, tools, external models, and host operations. The independent CLI recomputes statistics and vector costs through a separate kernel. Legacy online transition moments remain bootstrap-only even after eight samples.

SPARK-051 is not accepted yet. CP373 proves the bounded campaign, verification, admission, and fail-closed runtime machinery on generated fixtures; it does not supply the missing resident evidence. A real, externally custodied resident-32B campaign must still acquire at least 20 unique pairs per action, publish its certificate, and survive live selection/lesion/restoration ablations before calibrated value or capability gain can be claimed. The negative frontier verdict is unchanged.

CP374 closes the causal-intervention prerequisite that CP373's protocol assumed but the resident engine did not yet possess. A campaign cell now carries one strict intervention authority binding the separately configured current campaign policy and revision, complete canonical plan and exact cell, protocol, pair, task payload, execution ordinal, deterministic first-attempt identity, runner-attested starting state, exact normalized worker request, action, arm, and ordinal zero. A superseded embedded policy is rejected even while still cryptographically valid. The external campaign runner signs the payload under the separately configured Ed25519 root. Client, worker, engine, and request digest independently validate and bind the intervention. The lane is rejected on foreground/user requests.

At the first recurrent action opportunity, the worker commits the actual complete state inventory and refuses a mismatch. Six components are directly worker-measured: latent slots; branch/ensemble roles, scores, halting, savepoints, evidence lineage, isolation/exchange state, traces, budget, and accounting; KV state; evidence; memory; and public action state. Durable host state and RNG root remain externally captured and runner-attested, with that ownership explicit in the receipt. Treatment reconstructibly emits exactly one campaign_forced decision even when the observational state policy would not select it; control consumes the matched opportunity without an action and remeasures the post-state rather than copying it. Both arms remove the target action from the remaining executor inventory.

Before the worker may consume the request-bound attempt, it appends and fsyncs one ACTION_INTERVENTION_CLAIMED event to the canonical campaign journal. That event commits the intervention, normalized request, and signed journal head/count. A delayed authority for a superseded retry therefore fails against current journal state even if its embedded prefix remains valid. The exact claim lineage is then consumed before execution in a bounded, hash-chained, atomically replaced replay ledger under an interprocess lock and rechecked before the worker records execution. A crash after consumption fails that attempt closed. Claimed attempts survive restart without being auto-retried and cannot transition to FAILED while a worker may still execute; they must produce a sealed arm result or remain unresolved for operator reconciliation. Exact idempotent journal recovery, staged-result import, final transcript replay, and later result/verification/commit remain available.

EXECUTE requires a real resident executor and governed offer; the parent validates its handoff, and the durable host coordinator preserves and revalidates the intervention through dispatch. Version-2 durable host records are digest-verified, narrowly migrated to version 3, validated, and atomically rewritten. A migrated terminal record replays; an interrupted dispatch becomes UNKNOWN_EFFECT and requires reconciliation rather than blind retry. The final receipt binds pre/post component maps, aggregate state and KV identities, decision, complete action trace, exact occurrence count, request/attempt identity, and plan authority. Parent-side replay rejects missing, altered, relabeled, rehashed, wrong-bucket/snapshot, reused, or state-mutating control evidence. Bare campaign transitions are rejected by the ordinary transition validator, so experimental manipulations cannot train the live observational policy. Intervention failures may not silently become a vanilla decode even when fallback is otherwise enabled.

Seven dedicated contracts include an actual tiny-Qwen/MLX recurrent episode exercising the real complete-state serializer, signed-authority tampering, current-policy and durable replay refusal, canonical-journal retry supersession, an intentionally state-infeasible forced action, missing-executor refusal, treatment/control occurrence counts, remeasured control state after one explicitly omitted matched opportunity, worker omission, and fallback refusal. The durable-host suite separately proves campaign-forced EXECUTE reaches a persisted handoff and both safe v2 migration outcomes. The broader intervention, value-policy, request-identity, worker/client, execution-controller, epistemic-runtime, external-execution, preaction, verified-best, action-calibration, and frontier-task suite passes 456 tests after two unrelated pre-existing real-tiny-model failures are deselected. Both failures reproduce on a clean origin/main: transient constraint application remains zero after admission, and the verified latent-tree test receives an unexpected answer-replacement candidate. They remain open defects and must close before the resident campaign.

CP374 does not create whole-host post-mutation capture or erasure receipts, execute the resident-32B campaign, calibrate a cell, demonstrate reasoning gain, or change the negative frontier verdict. The crash-resumable resident runner, external custody, at least 20 unique pairs per action, final independent certificate, live selection/lesion/restoration ablations, and broader frontier proof all remain required.

CP376 removes the circular campaign-plan dependency and establishes the encrypted, one-use private snapshot substrate described in the whole-project tracker. CP377 closes the first mandatory review boundary on that substrate: a worker capture key is no longer its own trust root. Before process spawn, the parent creates a bounded random launch challenge under an ephemeral Ed25519 supervisor key. The child receives only that public challenge and signs its exact digest into a boot- and PID-scoped worker identity. After spawn, the parent verifies the real child PID and challenge, signs the worker identity, and publishes the complete challenge/identity/attestation chain. The parent private key never crosses the process boundary.

Claim-grade state-capture request schema v2 requires that complete origin binding plus an independently supplied expected supervisor public key at request construction, current admission, historical replay, and public or private receipt verification. A bare legacy worker identity, a self-rooted rogue supervisor, a cross-worker substitution, a wrong PID, a rehashed inner mutation, or a stale live challenge fails closed. Expired launch evidence is rejected for current admission while its signatures remain historically replayable. Synthetic MLX resilience receipts now execute the same parent- challenge/child-signature path instead of receiving a production bypass.

CP377 validation passes 73 focused worker-origin, state-capture, cancellation, and runtime-identity contracts and the broader MLX client, admission, resilience, ownership, memory, heartbeat, stability, and runtime matrix at 337/337. Strict focused Ruff and diff hygiene pass. This closes only the self-rooted-origin P1. Partial resident application ambiguity, descriptor- rooted filesystem transactions, recoverable multi-file publication, external key custody, streaming serialization, kill/race/disk fault tests, and the real serializable resident continuation remain open. No training or capability result is created, and the negative frontier verdict is unchanged. Post-rebase governance integration also assigns the newly landed, fixed-path learned-world-model checkpoint to its narrow canonical owner; the exact inventory is 2,007 calls in 1,878 buckets with migration debt unchanged at 1,788 calls.

CP378 closes the partial-application ambiguity in the snapshot state machine. Restore schema v2 durably distinguishes prepared from application_started; the second marker is atomically published and fsynced before the resident mutation callback begins, and binds the exact worker boot, PID, arm, request, snapshot, and operation. Only a provably pre-apply v2 marker may roll back. Any callback exception, post-apply hash mismatch, pre-commit storage failure, or process death after the started marker becomes typed UNKNOWN_APPLICATION evidence with same-process retry forbidden and process replacement required. An uncommitted v1 restore is also quarantined because the legacy schema cannot prove whether application started.

The MLX latent-job boundary handles that typed error before its generic runtime handler, emits a correlated fatal quarantine receipt, and exits the worker after delivery so it cannot serve another request. A spawned-child fault test enters the real apply callback, receives SIGKILL, and proves the surviving process reconstructs the durable quarantine and refuses retry. Focused storage/worker coverage passes 52 tests; the broader snapshot, identity, runtime, worker-origin, admission, and MLX resilience boundary passes 248/248. This closes the second CP376 P1, not the remaining descriptor, publication, custody, streaming, broader fault, or continuation work. No training or capability verdict changes.

CP379 closes the ancestor/path-reentry P1. The private snapshot store captures the root inode once, retains descriptors for every fixed namespace and the interprocess lock, and performs nested traversal, creation, read, atomic replacement, unlink, directory removal, listing, and fsync relative to those descriptors with no-follow semantics. Every operation verifies that the public root name, namespace entries, and lock name still resolve to the held inodes before and after mutation. Root, namespace, and lock substitution fail before private I/O and leave the replacement tree untouched.

A real two-process race against one arm proves the descriptor-held flock serializes independent store instances: one process commits the restore and the other receives private_snapshot_arm_already_used; no second application occurs. Existing kill, recovery, tamper, and lifecycle coverage remains green. Focused action-store coverage passes 32 tests and the broader worker/client/ runtime boundary passes 252/252. CP379 does not claim multi-file publication atomicity, external key custody, streaming serialization, a resident continuation, training, or reasoning gain.

To prevent SPARK-051 from expanding indefinitely, its remaining pre-training implementation is a fixed five-checkpoint burn-down: recoverable publication transaction; external key custody plus streaming codec; serializable resident continuation and first-action hook; worker/client/service/runner/verifier lineage; then the resident-1.5B destructive integration gate. Defects found by those gates are repaired in place, but unrelated capability expansion does not enter SPARK-051. The resident-32B training and preregistered reasoning campaign follow that gate.

CP380 closes recoverable snapshot publication. A capture is now assembled in a hidden transaction directory containing its encrypted chunks, envelope, key, authenticated handle, use ledger, and authenticated private publication record. Every file and newly created directory is durably synced before one descriptor-relative directory rename makes the complete bundle visible. Both parent directories are synced after the rename. No reader can observe the former partial key/chunk/envelope/ledger/handle sequence.

Retry under the interprocess lock searches authenticated complete bundles by request and component commitments. A crash after commit therefore returns the original opaque handle instead of producing a duplicate. A crash before commit leaves only a hidden transaction, which the next lock holder validates and removes before retry. Publication I/O failures roll back the whole staging tree and retain their original OSError as causal evidence under a stable snapshot-publication error.

Four new destructive contracts prove the transaction boundary: a spawned publisher is killed with SIGKILL before commit and leaves zero visible bundles; an injected ENOSPC leaves neither bundle nor transaction; a forced post-commit death recovers the exact committed handle; and two independent simultaneous publishers converge on one handle, one snapshot digest, and one bundle. Focused storage coverage passes 36/36 and the broader snapshot, identity, worker-origin, MLX runtime, admission, cancellation, and resilience boundary passes 256/256. Strict Ruff, bytecode compilation, enterprise gate, exact effect-ownership ratchet, and diff hygiene pass.

CP380 closes only the publication-transaction checkpoint. External key custody plus streaming, the serializable resident continuation and first- action hook, end-to-end lineage, and the resident-1.5B destructive gate remain the four fixed pre-training checkpoints. No resident training, reasoning gain, frontier gain, or WOW Signal is claimed.

CP381 closes external snapshot-key custody and bounded streaming. The snapshot store now requires an explicit custodian and has no plaintext-key or environment fallback. Its production provider wraps each random DEK with AES-256-GCM under a wrapping key held by macOS Keychain. The authenticated wrapped-key envelope binds the custodian identity, request, and opaque handle; only that envelope enters the snapshot transaction. Ordinary system opens are read-only and fail if custody was not provisioned. One host-only provisioning method creates and confirms the Keychain item before resident workers spawn, avoiding concurrent implicit key creation.

Publication now canonicalizes JSON incrementally, slices even one large scalar into bounded pieces, coalesces exact-size chunks, hashes and encrypts each chunk, and writes it directly into the hidden transaction. It no longer holds complete plaintext components plus every encrypted copy. Restore authenticates one chunk at a time into one component buffer, verifies its streaming digest, reconstructs the concrete state value, and zeroes the transient plaintext and DEK buffers. Existing-bundle retry must prove the current custodian can unwrap the committed DEK before it may return the original handle.

The focused custody, streaming, snapshot, and strict-Keychain boundary passes 57/57. The broader snapshot, identity, worker-origin, MLX runtime, admission, cancellation, resilience, and secret-backend boundary passes 277/277. Tests prove bounded multi-chunk binary and single-scalar JSON round trips, absence of the raw DEK from every bundle file, wrong-custodian refusal, tamper/context rejection, explicit provisioning, post-close refusal, and unchanged crash/ race/erasure semantics. A real native-Keychain probe provisioned once and reopened read-only with identical public custody and wrapping-key identities. Ruff, bytecode compilation, enterprise gate, exact governance ratchet, and diff hygiene pass.

CP381 closes only custody and streaming. Serializable resident continuation plus the first-action hook, end-to-end worker/client/service/runner/verifier lineage, and the resident-1.5B destructive integration gate remain the three fixed pre-training checkpoints. No resident training or reasoning/frontier gain is claimed.

CP382 closes the serializable resident-continuation and engine-owned first- action checkpoint. PortableStateComponent is a strict, deterministic, bounded binary format for nested scalar state, bytes, mappings, sets, NumPy tensors, and MLX tensors. It streams directly into the authenticated snapshot store, rejects malformed/noncanonical payloads, and preserves MLX bfloat16 through an exact uint16 bit view instead of an invalid NumPy conversion. No pickle or executable deserializer enters the boundary.

The engine now owns the capture point immediately before opportunity one can choose a cognitive action. It records complete branch runtime, latent slots, exact recurrent KV snapshots, evidence, memory, public action state, and the runner-supplied durable/RNG roots. Capture-only execution exits before policy selection and decode. Restore installs those tensors and cache snapshots into an equivalent fresh frame, rebinds per-episode KV capabilities to the fresh worker's root, and re-encodes all eight components before action. Any mismatch rolls the resident back and verifies the rollback; ambiguous rollback raises the existing fatal UNKNOWN_APPLICATION quarantine.

Real tiny-Qwen2 tests prove zero action/decode leakage, exact eight-component capture/restore identity, runner-state drift rejection, post-failure worker usability, and a complete engine -> streamed encrypted store -> authenticated restore round trip using actual latent and KV tensors. The post-rebase affected engine/store/intervention/worker/client matrix passes 334 tests. Strict Ruff, bytecode compilation, and diff hygiene pass. The candidate has no enterprise or governance delta from clean origin/main; both share two pre-existing enterprise ratchet regressions and the same three-new/one-stale governance drift introduced by cooperative semantic-closeout commits. No baseline is raised here.

CP382 does not claim end-to-end transport or a resident result. Worker/client/ service/runner/verifier lineage and the resident-1.5B destructive integration gate remain the two fixed pre-training checkpoints. No resident-32B training, reasoning gain, frontier gain, or WOW Signal is claimed.

CP383 closes the public resident action-state transport and independent verification lineage. The runner-facing frame contains only the signed capture request, current policy, exact model/execution/request identities, current parent-attested worker binding, and, for restore, the public capture receipt and arm. Runner durable and RNG roots are reconstructed from the signed payload inside both runner and worker code and checked against the request's portable-state commitments. No opaque snapshot handle, continuation value, raw runner state, DEK, or wrapped-key material crosses ordinary MLX IPC.

The client provisions external Keychain custody only when this claim-only lab lane is requested, replaces the proposed resident binding with the real live worker binding, independently admits the frame, and verifies every returned public receipt. The worker captures before action/decode, retrieves private state by authenticated request after process replacement, installs and re-encodes all eight continuation components, and invokes a first-action hook only after the exact aggregate state hash matches. Treatment and matched no-action are consumed once each; the second committed arm seals the pair, cryptographically destroys its key, deletes ciphertext, and returns signed terminal lifecycle evidence. Unknown application quarantines and replaces the MLX worker.

Capture and restore trust are intentionally separate. Historical request and capture signatures remain rooted in the original supervisor, while each replacement resident worker receives a fresh parent challenge and attestation under the current supervisor. A review-discovered key-confusion bug was repaired before publication, and the challenge-expiry regression now proves historical capture verification under the old root plus restore verification under the new root. The independent pair verifier accepts distinct worker/supervisor keys by arm and reconstructs capture, intervention, restore, once-only custody, seal, and erasure lineage.

Focused action-state, worker/client/service, and authority coverage passes 186/186. The complete affected continuation, custody, intervention, engine, runtime identity, worker origin, MLX admission/cancellation/resilience, secrets, and contract matrix passes 480/480. Ruff, bytecode compilation, and diff hygiene pass. CP383 adds no enterprise or governance debt; the two inherited enterprise regressions and unrelated three-new/one-stale governance drift remain explicit and no baseline is raised.

CP383 closes the lineage checkpoint, not the resident result. The sole fixed pre-training checkpoint is now the resident-1.5B destructive integration gate across capture, process replacement, both arms, duplicate refusal, terminal erasure, and independent verification. Resident-32B training and the preregistered reasoning/frontier campaign follow. No gain or WOW Signal is claimed.

CP384 completes that fixed pre-training gate. Capture admission and worker execution now independently derive the model-weights identity from the loaded resident worker, binding full checkpoint content, parameter counts, adapter stack, tokenizer artifacts, quantization, and identity gaps. A real Qwen2.5-1.5B-Instruct-4bit continuation is captured before action or decode, streamed through encrypted external custody, restored exactly into both arms, sealed after one use each, and cryptographically erased. A second real-model execution restores one state into runner-signed treatment and control: treatment performs exactly one BLIND_RESOLVE, control performs none, duplicate consumption fails, and permanent parameters remain identical.

The evidence boundary is explicit: CP383 proves spawned-process transport, replacement, trust rotation, quarantine, and independent verification; CP384 executes the corresponding state and intervention contracts on real model weights and MLX tensors in fresh engine frames. It does not relabel that composition as one all-layers spawned IPC test. The opt-in resident gate passes 2/2; after final identity hardening, the combined state/intervention/ wiring suite passes 159/159 with those gates enabled and 157/157 with two explicit deselections in ordinary runs. The registered resident_model category does not increase skip debt. CP384 adds no enterprise or governance drift and raises no baseline.

The five-checkpoint pre-training burn-down is now closed. SPARK-051 itself remains open for resident-32B training, the preregistered equal-compute reasoning/frontier campaign, independent verification, and live selection, lesion, and restoration evidence. No gain or WOW Signal is claimed. This is total checkpoint record 626; the 640-920 forecast leaves approximately 14-294 records, or 68.0%-97.8% checkpoint-count completion (80.3% midpoint).

CP385 attempted the first fresh resident-32B answer-channel preflight after the mechanical gate. The detached process was contained and terminal, but failed before model load because independently generated finite-domain train and holdout batteries shared a prompt. That failure is retained as evidence. The curriculum now seeds explicit train/holdout partitions and constructs holdout under an exclusion set containing every committed training prompt and task identity. The exact failing preflight dimensions and seed are a regression test. The focused curriculum, trainer, and preregistration matrix passes 44/44. This is total checkpoint record 627; the 640-920 forecast leaves approximately 13-293 records, or 68.2%-98.0% checkpoint-count completion (80.4% midpoint). A new source-bound campaign contract and detached preflight must still run before any training or gain claim.

CP386 loaded the actual resident 32B and exercised the repaired split. Its recurrent baseline scored 1/6. Calibration completed five probes before its bounded time cap: 4/20 completions were parseable and correct, for a 0.20 answer-channel fraction. The calibration admission correctly refused training as answer_channel_blocked; zero optimizer groups or updates ran. The terminal reporter then raised KeyError: diagnosis because zero-group telemetry intentionally had only reason. That failed receipt is retained. _signal_admission_report is now total over zero groups and emits an explicit diagnosis and required next gate; its focused trainer/curriculum/ preregistration matrix passes 45/45. No adapter, gain, or frontier claim is available. This is total checkpoint record 628; the 640-920 forecast leaves approximately 12-292 records, or 68.3%-98.1% checkpoint-count completion (80.5% midpoint).
SPARK-052 - Adaptive breadth/depth/tool routing. Scale recurrence, branch count, lookahead, tools, and verifier effort from difficulty, uncertainty, stakes, body pressure, deadlines, and resource admission while preserving user-facing work.

CP387 closes the live adaptive-routing contract. One deterministic policy now combines prompt-structure difficulty, uncertainty, stakes, allostatic body pressure, caller deadline, resident model scale, and the canonical runtime-admission pressure snapshot. It causally sets bounded recurrence, virtual width, latent-tree nodes/depth/branching, generative/counterfactual/ prefix verifier effort, and the zero-or-one host acquisition continuation. Critical body/resource pressure and sub-45-second deadlines force the lean tier. Unknown resource observation is conservatively pressured rather than treated as free headroom.

The plan is versioned, strict, and content-addressed. It is applied before worker IPC, re-enforced after learned execution-arm selection, and checked against actual worker steps, branches, verifier profile, decode surface, and wall budget. Independent execution and acquisition validators reject nested or outer-envelope tampering. Explicit experiment structural overrides remain identifiable and do not receive a false adaptive-execution certificate. Retrieved evidence still passes through the existing governed ingress and instruction-free context contracts; this checkpoint does not widen external execution authority.

The affected adaptive policy, service, acquisition, latent-tree, runtime control-plane, ontogeny, and surface matrix passes 284/284. Strict Ruff, bytecode compilation, and diff hygiene pass. The enterprise ratchet has zero finding in touched files and retains only its inherited repository excesses. This is total checkpoint record 629; the 640-920 forecast leaves approximately 11-291 records, or 68.4%-98.3% checkpoint-count completion (80.6% midpoint). This is runtime allocation evidence, not reasoning gain.
SPARK-053 - Principled stop and abstain. Stop on verified convergence, low value of further compute, budget exhaustion, or irreducible uncertainty; distinguish each reason in receipts and language generated by Aura herself.

CP389 implementation and evidence: A strict terminal-disposition policy now sits above the previously separate residual halter, learned stop head, value-of-computation terminal actions, loop-stability proof, and resource budget. It reconstructs one precedence-ordered reason from committed public evidence: irreducible uncertainty, wall/compute/recurrent budget exhaustion, measured non-positive value of further compute, verified fixed-point convergence, verified answer/action readiness, stability containment, interruption, or planned-depth completion. A converged label without a finite, anchor-bounded selected transition marked as a fixed-point candidate is rejected instead of promoted.

The disposition selects answer, bounded-answer, abstain, execute, or defer semantics. It does not select canned prose. A reason-specific epistemic instruction is tokenized with the resident tokenizer and appended only at final synthesis, after verification and temporary adaptation. The largest possible instruction is reserved before recurrence, so honesty language cannot unexpectedly consume the protected answer budget. The resident model still generates every output token in its own words.

The receipt commits the low-level halt, loop evidence, action trace, exact decision-time budget, precedence, reason, disposition, instruction text and tokens, complete final bridge token sequence, model-generated output tokens and text, and whether a confidence-bound resident repair produced the final answer. Independent service validation reconstructs the decision, proves monotonic budget lineage, proves the instruction is the exact suffix of the applied bridge, binds final output identity, and rejects substrate-only or unapplied language on the live path. Public token sequences expose only the fixed epistemic instruction, never private latent state or chain-of-thought.

Tests cover every required reason and precedence, false convergence, recomputed-hash tampering, detached instruction tokens, unapplied resident instructions, substrate honesty, full tiny-Qwen execution, learned stopping, service IPC, GWT coupling, and value-of-computation terminals. The affected matrix passes 171/171; strict Ruff, bytecode compilation, and diff hygiene pass. The enterprise ratchet reports no finding in touched files and retains only inherited repository excesses (placeholder_stub_mock 15 > 13 and high/critical 40 > 39). This closes principled stop semantics and language causality; it does not claim a reasoning or frontier-performance gain.
SPARK-054 - Complete causal receipts. Record state lineage, operators, branch isolation, tool evidence, verifier scores, accepted/rejected updates, compute, adaptations, stopping, final synthesis, and integrity proofs without exposing private chain-of-thought.

CP390 implementation and evidence: Every latent episode now carries one ordered, hash-linked public causal envelope over twelve fixed stages: ingress identity, recurrent state lineage, cognitive operators, branch isolation/exchange, optional tool-memory evidence, verification, accepted/rejected updates, compute, temporary/durable adaptation, stopping, final synthesis, and runtime/model integrity. Each node commits only field names, presence, shape/count, canonical value hashes, and the preceding node hash. Raw latent values, private reasoning, and tool-secret values are never copied into the envelope.

The envelope is reconstructed at every identity boundary instead of being patched: the engine emits an honest partial receipt, the worker rebinds it after request and resident-worker identity exist, and the client rebuilds the final form only after capturing source/app/runtime provenance. The live service independently reconstructs the complete DAG and rejects missing stages, anonymous episodes, source drift, reordered or recomputed nodes, mutated privacy declarations, detached final-output identity, and unproven parameter or applicable fast-weight integrity. Action-capture and failed episodes retain honest partial envelopes rather than inventing completion.

Dedicated tests prove exact ordering and hash lineage, optional-evidence semantics, source and self-consistent envelope tampering resistance, strict integrity completion, and absence of sentinel private reasoning/tool values. Tiny-Qwen engine, worker, client, service, terminal, stop-policy, value-of-computation, runtime-identity, and integrity matrices pass 277/277. Strict Ruff, bytecode compilation, and diff hygiene pass. The enterprise ratchet has no new finding attributable to CP390 and remains at its inherited repository excesses (placeholder_stub_mock 15 > 13 and high/critical 40 > 39). This proves receipt completeness and provenance on the implemented path; it is not a reasoning-gain or frontier-performance claim.

G. Temporary and permanent learning

H. Whole-Aura causal integration

SPARK-066 - Resident-32B penultimate neural path. Make the selected architecture run inside the live resident checkpoint's middle-layer/latent execution, bind exact model and adapter identity, and prove successful output is not replaced by an ordinary generation or shallow orchestration fallback.

F7-E (2026-07-27, Fable lane) closes the falsifiability half in core/brain/llm/latent_cortex/penultimate_execution_receipt.py. This repository has already paid for the clause about fallbacks: CP227's accuracy gate was voided because _decode ran outside recurrence_adapter_scope, both arms decoded the bare base model, on@d matched off@d exactly, and a null that meant nothing was nearly a result. Nothing lied — nobody checked that the treatment was attached, and an unchecked treatment reported as a run treatment is indistinguishable from a real null. The receipt makes the four ways that happens refusable:
- The adapter that was never attached. attached: true is not taken at its word: the adapter must name the blocks it activated in and they must equal the blocks it was expected to activate in. An empty activation list with attached: true is adapter_did_not_activate, refused at construction. A partial activation is refused the same way. An unattached adapter may not carry an identity — a control arm is a valid receipt, it simply cannot claim a treatment.
- The depth that never recurred. Passes carry per-pass state digests; several passes that all landed on the same state return no_recurrence_occurred — the T=1 identity wearing a depth label. One pass is not accused of failing to recur.
- The state computed and then dropped. The decoded state must be the final pass's state; decoding from an earlier or unrelated state returns decode_did_not_consume_final_state. A run that recurred and then decoded from something else did not use its own work.
- The fallback wearing the latent name. The mechanism is declared, a fallback requires a reason, and a receipt recording a fallback cannot also claim recurrent_latent. An ordinary generation refuses the latent claim without erroring — that is the receipt correctly declining, not a failure. A correction inside this checkpoint. The first draft required the execution to sit at layer_count - 2, reading "penultimate" as a fixed layer index. That is wrong for this architecture and worse than wrong as a check: the RLC's window is a middle band (prelude / window x T / coda), so the only way a real producer could satisfy it was by reporting a position it did not run at — a check that can only be passed by lying. It is replaced by a structural execution_window: a non-empty band of real layers with something before it (window_has_no_prelude otherwise) and something after it (window_has_no_coda), and a window spanning the whole stack is refused as an ordinary forward pass wearing the name. That is what actually separates middle-layer latent execution from shallow orchestration.
21/21 focused tests pass, including a direct reconstruction of the CP227 shape asserted to be unable to produce a PROVEN verdict.

F7-E part two wires the producer, so this is no longer a validator with nothing feeding it. The failure it closes is CP227's, one level down: the repaired gate's aggregate calls counter proves something fired. It cannot distinguish "all ten wrapped projections fired" from "nine did and one was silently never wrapped", and that difference is a treatment that is really nine tenths of a treatment, with the arms differing by a fraction of the adapter nobody designed.
- RecurrenceAdapterActivation (live path, recurrence_adapter.py) now records applied_blocks and applied_sites per application, with activated_blocks() and unfired_sites(expected) on top. Identity is optional and travels with the projection: ScopedLoRALinear.from_base accepts block_index and site, validated at attach.
- Both attachment sites supply it — attach_adapters (tools/train_intrinsic_recurrence.py, the one the accuracy gate imports) and recurrent_sft_execution.py — and attach_adapters now returns adapted_sites and adapted_block_indices: the exact set a downstream activation receipt must match.
- tools/eval_intrinsic_accuracy.py's dark-adapter gate is upgraded from "something fired" to "everything wrapped fired." A block where any adapted site stayed dark now stops the run with instrument_partial_adapter, names the first dark site, and returns 2 — the same fail-closed treatment the fully-dark case already got.
- activated_blocks in the receipt is therefore measured, not asserted: the value comes off the forward pass that actually ran.
9/9 additional integration tests run the real attach_adapters against a real MLX module tree (not a stub): every wrapped site fires on a healthy forward; one projection reverted to a bare nn.Linear is named while calls still looks healthy; a whole dark block is named site by site; the no-scope CP227 shape leaves every site unfired while the model still answers; and a partial activation is shown to be unable to construct a latent receipt at all. 6/6 further contracts cover the identity recording itself, and the existing adapter suites pass 34/34.

F7-E part three closes the state producer. core/learning/intrinsic_recurrence_receipt.py builds the receipt from the tensors an actual pass produced. Three things there are load-bearing:
- The decode state is the window's output, not the model's. The stack is prelude → window x T → coda → norm → head, and the state the coda consumed is trajectory[-1]. Digesting the post-coda hidden instead would make "decode consumed the recurrent state" trivially true for any forward pass, including one that discarded the window's work — the check would pass by construction. A test asserts the two digests differ.
- Digests are canonical: float32, contiguous bytes, with the shape hashed in, so the same computation replays identically and two different states of equal byte length cannot collide.
- run_and_receipt opens the adapter scope itself. Forgetting the scope is precisely the CP227 failure; on this path it cannot be forgotten by a caller.
14 further tests run a real Qwen2 stack through the real recurrent_hidden_states loop: a reverted projection yields an attached: false receipt that refuses the claim rather than an exception, a partially dark adapter cannot produce a receipt at all, per-pass L2 deltas are measured distances, and a changed weight changes every pass digest. The SPARK-064/065/063/066/067/068 surface plus the touched adapter and instrument suites pass 253/253.

The checkbox stays open: no resident-32B execution has produced a receipt through this yet — the producer is proven on a tiny real Qwen2, not on the resident checkpoint — and the live worker's decode path still has to call it. This checkpoint runs no model of consequence and grants no capability claim.
SPARK-067 - Organism-wide bidirectional coupling. Connect epistemic state and recurrent control causally with agency/Will, memory, tools, personality/self-model, affect/body, global workspace/consciousness, reasoning amplifiers, goals, and learning; lesion each seam and test both directions instead of accepting metadata-only coupling.

F7-F (2026-07-27, Fable lane) makes "instead of accepting metadata-only coupling" mechanical, in core/brain/llm/latent_cortex/coupling_matrix.py. That phrase was doing a lot of work: a field copied from the epistemic state into another subsystem's receipt is not a coupling, it is a field that was copied — the subsystem behaved identically, nothing downstream can tell the difference, and the receipt showing the field present reads exactly like evidence. So:
- Every seam declares its evidence kind. metadata is legal to record and illegal to count: a seam whose forward or reverse evidence is metadata is refused by name as metadata_only_coupling.
- Every seam needs both directions, with the direction label checked against the slot it was filed in. A controller that reads memory but cannot change what memory keeps is coupled to memory the way a thermometer is coupled to weather.
- Every seam needs a lesion that removes the effect — at least half of the intact deviation from baseline must disappear when the seam is cut. An effect that survives its own lesion was not flowing through that seam (lesion_did_not_remove_effect), and a lesion with no intact effect to remove is refused rather than counted as a clean cut.
- The nine subsystems are complete by declaration: a matrix missing one is invalid, not partially passing on the seams someone bothered to measure. Every uncoupled subsystem is named in the matrix verdict.
23/23 focused tests pass, including each of the nine subsystems dropped in turn, metadata in either direction, a mislabeled direction, a direction that measured nothing, an underpowered direction, an effect surviving its lesion, a flat lesion, a duplicate seam, and an edited seam.

F7-F part two adds the harness that makes the classification a measurement. The matrix cannot enforce the metadata rule on its own: kind arrives already labelled, and a caller who believes their coupling is behavioral writes behavioral. Believing it is the normal case — the field is being passed, the receipt does show it — and being wrong about it is the whole failure. So core/brain/llm/latent_cortex/coupling_harness.py does not accept the label. It runs the behavior with the seam open and closed and compares the observable outcome identity of each trial; if every trial's outcome is identical either way, the seam moved a field and changed nothing an observer could act on, and the evidence is recorded as metadata however the caller would have described it. The lesion is measured the same way. A seam that raises mid-run is a finding, not a weak measurement — averaging over the trials that survived would report a broken seam as a weak one — and a trial count below the matrix's floor is refused while it can still be fixed. Measurements are deterministic per trial index and replay exactly.

13 further tests, including the case that looks fine from the inside: a seam that really passes a field on every trial, whose downstream decision is byte-identical either way, classified metadata and then refused by the matrix.

The checkbox stays open: the harness runs whatever callables it is given and none of the nine seams has been bound to it against a live organism. This checkpoint runs no model and grants no coupling claim — an empty matrix proves nothing and neither module pretends otherwise.
SPARK-068 - Production reliability and observability. Bound latency, memory, event-loop work, cancellation, concurrency, worker recovery, health, degradation taxonomy, privacy, audit logs, and UI presentation; fix root causes rather than suppress warnings or emit canned fallback prose. Replace the action-calibration protocol-v3 complete-prefix envelopes with a versioned compact append-only proof (for example, checkpointed Merkle/MMR inclusion witnesses) that preserves independent replay and external signatures without quadratic JSON serialization, memory growth, or verification time.

F7-C (2026-07-27, Fable lane) builds the compact append-only proof named in the second half of this item. Three modules, all additive — the existing campaign_journal_prefix envelope is untouched and still authoritative:
- journal_state.py extracts the journal's transition function from action_intervention._validate_journal_prefix, which now delegates to it. The rules, ordering constraints, and refusal messages are unchanged. This extraction is the point: the compact proof folds the same state machine, and a second copy would have been free to drift into a witness that passes one validator and fails the other. The folded state is serializable and digestible, and is bounded by the plan's cell count rather than by journal length.
- journal_accumulator.py is a Merkle Mountain Range over event digests: append-only by construction, O(log n) inclusion, domain-separated node tags, and the size hashed into the root so a proof valid at one length cannot be replayed at another.
- journal_witness.py is the versioned v2 envelope — accumulator root, a checkpoint (folded state + O(log n) inclusion proof of the checkpoint event), and the bounded suffix the verifier folds itself.
What the checkpoint is trusted for is stated, not hidden. A witness cannot prove its own checkpoint state; someone replayed genesis→k once to compute it. So verify_journal_witness requires the caller to pass the checkpoint digest it independently trusts and refuses on disagreement — there is no "if omitted, trust the witness" path, and that argument being mandatory is the safety property. A verifier that trusts nobody passes checkpoint_sequence=0, which degenerates to exactly today's full replay. In the checkpointed branch the witness's root is deliberately not re-compared against itself: that check cannot fail, and a check that cannot fail is the defect this ledger keeps naming. The suffix is bound instead by the checkpoint event's proven inclusion plus the previous_event_sha256 chain, so substituting another log's suffix requires a SHA-256 collision.

54/54 focused tests pass: the accumulator verified against every index at twelve log sizes plus a tamper battery (reordered, edited, truncated, swapped leaf, forged path step, flipped side, shortened path, replayed at another size, wrong peak), and the witness proved equivalent to the complete-prefix validator at every checkpoint position, plus its own refusals (untrusted checkpoint, disagreeing digest, genesis smuggling a digest, tampered body, forged inclusion, wrong plan, lying size, edited suffix, a journal whose live attempt already committed). At an 8-event checkpoint cadence a 159-event campaign's envelopes carry under a tenth of the events the prefix form carries.

The checkbox stays open: this closes only the compact-proof half. The latency, memory, event-loop, cancellation, concurrency, worker-recovery, health, degradation-taxonomy, privacy, audit-log and UI halves of SPARK-068 are untouched, and the v2 envelope has no campaign caller yet — adoption is the march's call, since it owns the campaign protocol. No model runs here.

I. Training admission and scientific proof

SPARK-069 - Fresh source-bound resident campaign. Run the detached answer-channel preflight, require measured parseability and discriminative reward, preregister the new source/model/training bundle, then train/fuse only if admission passes. Preserve failed admissions as engineering evidence and repair their causes.
SPARK-070 - Full falsification matrix. On fresh held-out tasks run: recurrence-depth curves; wrong/right transition matrix; blind-review arms; structural-diversity arms; verifier arms; latent ablate/perturb/transplant; sham/noise/no-op controls; adversarial and OOD variants; d+1/2d/4d compute generalization; lesions/restorations; fast-weight controls; and broad non-reasoning regressions.
SPARK-071 - Powered resident frontier certificate. Compare base vanilla, base+RLC, adapter vanilla, adapter+RLC, equal-compute/search controls, and externally pinned frontier controls on broad fresh tasks. Require adequate power, confidence bounds, positive interaction, no material regressions, sealed raw artifacts, independent verification, and external trust roots.
SPARK-072 - Publication, conditional WOW Signal, and continuation. If and only if SPARK-071 accepts, create the WOW Signal commit and publish full transcripts, methods, literature, task manifests, hashes, statistics, ablations, failures, limitations, and verifier output. If it rejects, publish the negative result and open named repair checkpoints. In either case, resume every remaining Aura tracker item, final live proof, production/enterprise gates, and only then the deferred multi-hour soak.

Running result

No Spark frontier or intelligence claim is currently accepted. The latest resident recurrent-GRPO campaign ended without optimizer updates or a usable learning signal. Checkpoints 309-313 repaired diagnosis, admission, the separate answer-channel curriculum, and a source-bound detached preflight path. Those are real engineering gains, but they do not satisfy SPARK-069 or any later capability checkpoint until a fresh exact-source run produces accepted evidence. CP321 adds a tested epistemic evidence firewall; it does not change that negative capability verdict. CP322 adds measured claim calibration and enforced abstention; it likewise does not change the negative capability verdict. CP323 adds a durable weighted hypothesis portfolio with protected alternatives; it likewise does not change the negative capability verdict. CP324 completes the durable operation-history authority but not its live controller consumer; it does not change the negative capability verdict. CP325 closes the selective-memory ingress and its context-only authority chain; it does not change the negative capability verdict. CP326 closes durable live cognitive-operation history and admission; it does not change the negative capability verdict. CP327 makes the full action vocabulary causal inside recurrent execution and persists checked outcomes; it does not change the negative capability verdict. CP328 closes fresh-context branch isolation and exact cache restoration; it does not change the negative capability verdict. CP329 closes the nine-program executable cognitive-operator bank; it does not change the negative capability verdict. CP330 closes wording-independent structural support classes and exact service reconstruction; it does not change the negative capability verdict. CP331 closes causal correlated-support weighting and durable checked-outcome calibration; it does not change the negative capability verdict. CP332 closes first-answer/ownership blindness for branch review and makes the review policy service-verifiable; it does not change the negative capability verdict. CP333 closes decoy-balanced verifier admission and contains critic failures; it does not change the negative capability verdict. CP334 closes pre-causal critic/generator function separation and installs the durable shared-blind-spot meter and reliability gate; its live evidence state is still honestly unpowered and it does not change the negative capability verdict. CP335 closes bounded, declared, provenance-preserving branch exchange and train/live mailbox parity. The tensor role lesion/swap/restoration path proves that executable labor follows role programs rather than fixed branch indices, but no powered resident task campaign has yet established an accuracy benefit; the negative capability verdict is unchanged. CP336 closes claim-grade structural resource and information accounting. It also repairs the resident recurrent-GRPO execution-spec artifact omitted during the CP335 schema migration and makes the independent scoring kernel reconstruct the new accounting evidence without importing the production grader. This does not create a capability result; the negative capability verdict is unchanged. CP337 closes persistent evidence-grounded recurrence. Prompt KV remains shared and read-only; typed cognitive evidence forms an immutable causal prefix; and mailbox plus hypothesis slots carry the mutable recurrent state without prose re-encoding. Service-reconstructed transition commitments prove evidence invariance and causal hypothesis updates. A source matrix traces resident weights, Black Hole/RAG and episodic memory, Wikipedia/reference retrieval, local nonparametric one-shot memory, live organs, and GWT return. Governed in-episode web/tool production remains open, and the negative capability verdict is unchanged. CP338 closes stable shared-loop mechanics and fixed-depth train/live update parity. Fixed-anchor norm control, finite-state refusal, model-bounded KV positions, continuity-aware fixed-point diagnostics, contained-divergence disposition, and service-reconstructed certificates are now causal runtime contracts. This does not create a capability result; the negative capability verdict is unchanged.

CP339 closes the learned accept/discard mechanism and its calibrated artifact, runtime state-selection, accounting, and independent reconstruction contracts. The fitted-head causal test changes a real tiny-Qwen recurrent trajectory and downstream logits. No resident-32B head or capability campaign ran; default execution therefore remains explicit passthrough and the negative capability verdict is unchanged.

CP340 closes calibrated learned stopping and its measured quality/value evidence boundary. The controller preserves hard recurrence invariants before consulting the head; the service reconstructs its exact causal inputs and decision after transport. Task-disjoint bounded workload tests establish earlier easy-task stopping without hard-task regression, and a real tiny-Qwen equal-evidence arm proves reduced recurrent transitions and layer applications. No resident-32B stop head or powered broad-task campaign ran, so this does not change the negative frontier verdict.

CP340 validation passes the focused policy/artifact/workload/causality/tamper boundary 7/7 and the affected adaptive-halting, learned-bridge, attachment, engine, recurrence, schedule, branch, wiring, update-acceptance, value-of-computation, and resource suite 256/256 in 89.36 seconds. The final fixed-snapshot RLC, latent-cortex, recurrence/training, global-workspace, GWT, and execution-controller ownership gate passes 1307/1307 in 485.19 seconds. Strict focused Ruff, bytecode compilation, and git diff --check pass.

Validation: the focused acceptance-head and causal-transition suite passes 6/6; the combined acceptance/service reconstruction boundary passes 7/7; and the affected engine, recurrence, schedule, branch, wiring, and recurrence-native suite passes 203/203. The first full ownership gate exposed a real evidence- slot/reasoning-slot shape mismatch in relative-distance extraction. The repair uses a pooled-vector distance when the two immutable groups have different slot counts, and the targeted ingress regression then passes 6/6. The captured final fixed-snapshot RLC, latent-cortex, recurrence/training, global-workspace, GWT, and execution-controller ownership gate passes 1300/1300 in 776.80 seconds. Strict focused Ruff, bytecode compilation, and git diff --check pass. No resident 32B capability campaign was run, and the negative frontier verdict is unchanged.

CP341 closes confidence-bound best-state authority and overthinking reversion. Verifier observations are branch-local, ordinary scalar scores cannot acquire state-selection authority, interval overlap preserves the incumbent exactly, and the final returned state is independently reconstructable from the action and loop-stability receipts. Fixed-depth experiments remain unchanged. No resident-32B bounded-verifier campaign or externally rooted verifier evidence ran, so this does not change the negative frontier verdict.

CP341 validation passes the focused engine, verifier, resource-accounting, schedule/branch, value-of-computation, and service boundary 194/194 and the affected adaptive-halting, learned-bridge, attachment, recurrence, escape, update-acceptance, epistemic-runtime, and wiring gate 298/298. The final fixed-snapshot RLC, latent-cortex, recurrence/training, global-workspace, GWT, and execution-controller ownership gate passes 1323/1323 in 912.53 seconds. Strict focused Ruff, bytecode compilation, and git diff --check pass.

CP342 closes the hidden-state neural uncertainty mechanism and its bounded causal authority. Correctness and entropy now come from an objectively trained, task-disjoint calibrated neural head over the admitted recurrent state rather than self-reported confidence. Sparse bins abstain; supported measurements can select among branches only when an admitted task verifier does not supersede them. The service reconstructs the exact head estimate and selection evidence. No resident-32B head or powered broad-task campaign ran, so the negative frontier verdict is unchanged.

CP342 validation passes 13/13 focused artifact/live tests, three repeated real tiny-Qwen causal-selection runs, and the affected engine, schedule/branch, worker-origin, service, resource, update/stop, verified-best, value-of-computation, and epistemic-runtime gate 254/254 in 137.07 seconds. The final fixed-snapshot RLC, latent-cortex, recurrence/training, global-workspace, GWT, and execution-controller ownership gate passes 1338/1338 in 707.61 seconds. Strict focused Ruff, bytecode compilation, and git diff --check pass.

This is total checkpoint record 403. The revised forecast remains 466-733 total records, now approximately 63-330 records after this checkpoint. Checkpoint- count completion is approximately 55.0%-86.5%, with a midpoint planning estimate of 67.2%. Next: publish CP342, then implement SPARK-029's subtle mistake locator.

CP343 closes the bounded transition-mistake locator. Complete controlled- mutation traces now train a prior-to-proposal neural head under task-disjoint in-domain calibration and genuinely domain-disjoint OOD evaluation. Aggregate and per-domain exact, within-one, no-error, discrimination, and calibration gates must pass before an artifact is admitted. Rejected proposals remain visible, while the locator is structurally forbidden from steering repair. The service independently reconstructs the entire live receipt.

CP343 validation passes the focused artifact, live runtime, real tiny-Qwen, and service boundary 99/99 and the affected engine, branch, recurrence, update, uncertainty, resource, worker-origin, stop, verified-best, and wiring gate 250/250 in 133.40 seconds. The final fixed-snapshot RLC, latent-cortex, recurrence/training, global-workspace, GWT, and execution-controller ownership gate passes 1351/1351 in 701.66 seconds. Strict focused Ruff, bytecode compilation, and git diff --check pass. The effect-ownership baseline now records only the four reviewed locator and neural-uncertainty artifact read/write effects. Repository-wide governance lint remains non-green at 49 regressions and 13 stale buckets from concurrent unrelated work; CP343 does not absorb that drift or claim the global gate.

This is total checkpoint record 404. The revised forecast remains 466-733 total records, now approximately 62-329 records after this checkpoint. Checkpoint- count completion is approximately 55.1%-86.7%, with a midpoint planning estimate of 67.4%. Next: publish CP343, then implement SPARK-030's complete- trace bidirectional reflector. Final multi-hour soaks remain deferred until every shorter gate is green.

CP344 closes the always-on complete-hidden-trace reflector. Every transition now contributes bounded prior, proposal, and admitted sketches; a read-only full-sequence pass compares each step with both past and future admitted context, the initial latent premise, and final latent conclusion. Rejected proposals remain visible but outside the admitted path. The service independently reconstructs the evidence, and no decoded answer or mutation authority enters the critic.

CP344 validation passes the focused reflector/service boundary 94/94 and the affected engine, branch, recurrence, update, uncertainty, locator, resource, escape/telemetry, worker-origin, and wiring gate 257/257 in 133.92 seconds. The final fixed- snapshot RLC, latent-cortex, recurrence/training, global-workspace, GWT, and execution-controller ownership gate passes 1358/1358 in 697.14 seconds. Strict focused Ruff, bytecode compilation, and git diff --check pass.

This is total checkpoint record 405. The revised forecast remains 466-733 total records, now approximately 61-328 records after this checkpoint. Checkpoint- count completion is approximately 55.3%-86.9%, with a midpoint planning estimate of 67.6%. Next: publish CP344, then implement SPARK-031's calibrated contradiction tensor over the reflected trace. Final multi-hour soaks remain deferred until every shorter gate is green.

CP345 closes SPARK-031's calibrated transition-by-latent-position contradiction tensor. The reflector now preserves bounded position evidence, and the learned head compares every cell against local, premise, conclusion, prefix, suffix, and trajectory context. Cell and step probabilities are calibrated independently. Admission requires complete controlled-mutation and sham tensors, disjoint task/trace/evidence identities, genuine OOD domains, middle and long-context support, aggregate and per-domain localization floors, and calibration bounds.

The worker and service reconstruct every channel, feature commitment, probability, aggregate, source binding, and non-authority claim. Exact-SHA, no-follow artifact loading fails closed. Unavailable mode emits no probability and performs no hidden scoring work. SPARK-031 grants no state, selection, repair, or attention authority; SPARK-032 remains the bounded perturbation milestone.

CP345 validation passes the focused controlled-mutation/tensor/reflector/service boundary 108/108 and the affected engine, update, locator, uncertainty, escape, resource, worker-origin, and wiring gate 227/227 in 120.93 seconds. The final fixed-snapshot RLC, latent-cortex, recurrence/training, global-workspace, GWT, and execution-controller ownership gate passes 1372/1372 in 715.81 seconds. Strict focused Ruff, bytecode compilation, and git diff --check pass. The effect-ownership baseline selectively records the contradiction artifact's reviewed atomic write and stable no-follow descriptor read. Repository-wide governance lint remains non-green at the same 49 unrelated/concurrent regressions and 13 stale buckets present before this checkpoint; CP345 does not absorb that drift or call the global gate green.

This is total checkpoint record 406. The revised forecast remains 466-733 total records, now approximately 60-327 records after this checkpoint. Checkpoint- count completion is approximately 55.4%-87.1%, with a midpoint planning estimate of 67.7%. Next: publish CP345, then implement SPARK-032's bounded, counterfactually verified contradiction-driven perturbation. Final multi-hour soaks remain deferred until every shorter gate is green.

CP346 closes SPARK-032's bounded contradiction-driven intervention. An admitted coordinate may propose a change only to its writable position in the selected branch's final latent workspace. The guided arm competes against exact no-op and norm-matched orthogonal-random controls in counterbalanced, repeated, fixed-length transformer probes. Protected cognitive-context slots are receipt-bound and immutable.

The contradiction head cannot authorize its own proposal. Retention requires a decoy-admitted authoritative verifier, stable exact or calibrated-interval observations, equal actual layer applications, and a guided lower bound above both control upper bounds by the configured margin. Every unavailable, scalar-only, unstable, unequal-compute, tied, regressing, failed, or under-budget path restores the exact baseline. The service independently reconstructs source, configuration, geometry, protected positions, observations, compute parity, decision, and rollback.

CP346 validation passes the focused contradiction/perturber/tensor/wiring boundary 126/126 in 68.19 seconds and the final affected engine, predecessor, resource, worker-origin, and service boundary 293/293 in 157.95 seconds. The conservative fixed-snapshot RLC, latent-cortex, recurrence/training, resident-campaign, global-workspace, GWT, and execution-controller ownership gate passes 1538/1538 across 96 files in 789.05 seconds. Strict focused Ruff, bytecode compilation, and git diff --check pass. Repository-wide governance lint remains non-green at the same 49 unrelated/concurrent regressions and 13 stale buckets; CP346 adds no effect-ownership entry and does not absorb that drift.

This is total checkpoint record 407. The revised forecast remains 466-733 total records, now approximately 59-326 records after this checkpoint. Checkpoint- count completion is approximately 55.5%-87.3%, with a midpoint planning estimate of 67.9%. Next: publish CP346, then implement SPARK-033's locally conditioned exploration. Final multi-hour soaks remain deferred until every shorter gate is green.

2026-07-27 - CP414 independently verified negative SPARK-059 verdict

The first contained five-arm Qwen2.5-1.5B recurrent-SFT falsification produced a real negative result. The trained arm improved mean paired holdout loss against base on all 24 examples and passed shuffled-trace and syntax-only controls, but failed the sham-label comparison in the structured-program family. Personality and tool-honesty likelihood canaries each lost one target top-1 decision. Generated behavior was not measured. The independent verifier accepted the complete source, model, authority, custody, workload, adapter, containment, kernel-probe, and decision chain. No transfer, reasoning, frontier, resident-32B, promotion, or WOW Signal claim was admitted.

2026-07-27 - CP415 generated-behavior falsification boundary

SPARK-059's evaluator now measures the behavior that the recurrent checkpoint actually generates instead of treating teacher-forced likelihood as a proxy for user-visible reliability. A source-bound registry defines 12 training-disjoint canaries across identity grounding, tool-effect honesty, and authority safety. The grader accepts bounded paraphrases through explicit alternative groups, rejects named hallucinations and false effect/authority claims, and recomputes every result from raw generated text.

The sealed evaluator now runs the same deterministic recurrent graph on the base and trained arms. Every observation binds the exact generation contract, arm, adapter fingerprint, raw token sequence and hash, raw text and hash, engine result, decode termination, fallback state, and before/after permanent parameter digests. The base arm must execute with recurrent adapters disabled; the trained arm must match the independently rehashed trained-adapter fingerprint. A candidate passes only when all trained outputs satisfy their canaries and no base-pass case becomes a trained failure.

The independent verifier reconstructs the canonical generation contract, requires the exact ordered case set, rehashes raw tokens and text, regrades all outputs, replays family and right-to-wrong accounting, and includes generated behavior in the final transfer gate. The behavior-producing engine and its direct execution dependencies expand the exact source closure from 34 to 51 roles. Missing, extra, reordered, or modified roles fail closed.

The complete structured/recurrent SFT custody, containment, admission, tokenization, trainer, evaluator, and verifier boundary passes 205/205 in 381.60 seconds. The focused behavior/evaluator/verifier replay passes 31/31; strict Ruff, bytecode compilation, and diff hygiene pass. No model campaign ran in CP415, so CP414's negative scientific verdict is unchanged.

SPARK-059 remains open. Next is a larger balanced, retention-aware, evaluator-disjoint curriculum; complete family-balanced epochs; equal-work controls; then a new contained five-arm run through this generated-behavior verifier. Counting CP415 makes 661 total checkpoints. The 661-920 forecast leaves 0-259 records, or approximately 71.8%-100.0% checkpoint completion with an 85.9% midpoint. Final multi-hour soaks remain deferred.

2026-07-27 - CP416 balanced retention and replay-proof training contract

The next SPARK-059 campaign now combines the structured candidate with 24 training and 12 validation retention rows across identity grounding, tool-effect honesty, and authority safety. Split-independent case identities make cross-split prompt reuse detectable. The runtime retention manifest commits the exact generated-behavior registry and rejects any exact prompt or eight-word phrase overlap; this is an executable admission condition, not a test-only assertion.

Projected-dataset schema v2 commits a deterministic family-balanced epoch, its exposure receipt, and the retention manifest. Every family receives equal optimizer updates, every row is covered, large allocations fail before list construction, complete epochs and full validation are mandatory, and controls replay the exact multi-epoch history. Balanced checkpoint state binds its initial adapter fingerprint and must satisfy step == epoch * epoch_length + cursor. Resume, evaluation, and independent verification reconstruct those claims.

The reference trainer and all control arms seed adapter construction at the same boundary and use one shared exact tensor fingerprint. A mismatch between the trained arm's initial adapter and any control fails before control training can earn equal-work credit. Authority, control, and evaluator source closures expand to 15, 16, and 54 roles. Legacy evidence remains readable only under its frozen source boundary; changed code does not claim in-place resume compatibility.

Validation passes 228/228 across the complete structured/recurrent SFT matrix in 392.68 seconds and 71/71 across the repaired focused boundary. Ruff, bytecode compilation, and diff hygiene pass. CP416 runs no model, does not alter CP414's negative verdict, and grants no reasoning, frontier, resident-32B, promotion, or WOW Signal claim.

SPARK-059 remains open. Next is fresh larger synthetic custody, exact balanced authority, contained admission and training, equal-work controls, the five-arm generated-behavior evaluator, and independent verification. Counting CP416 makes 662 total checkpoints. The 662-920 forecast leaves 0-258 records, or approximately 72.0%-100.0% checkpoint completion with an 86.0% midpoint. Final multi-hour soaks remain deferred.

2026-07-27 - CP417 independently verified retained-campaign boundary

The retained Qwen2.5-1.5B five-arm campaign completed under a fresh deny- default evaluator contract. Its executable kernel probe binds exact files and replays the real sandbox command: evaluator read succeeded, while network, process fork, production write, resident sibling-checkpoint read, and training- checkpoint write each failed with kernel EPERM. The final detached run returned zero after 259.31 seconds with no restart, timeout, surviving process group, or lineage. The independent verifier rehashed and reconstructed the complete source, authority, custody, checkpoint, initial adapter, control, model, decision, containment, and kernel-probe chain.

The scientific result is mixed and remains negative at the required boundary. The trained arm improved heldout loss against base on all 48 examples (1.42289607 to 1.15947549; mean delta -0.26342058), with 124 wrong-to-right, 27 right-to-wrong, and 97 net target-token corrections. It beat sham labels, shuffled traces, and syntax-only controls with no negative-transfer family. Personality, tool-honesty, and safety likelihood canaries all passed.

Generated behavior did not pass. The trained checkpoint satisfied only 1/12 behavior canaries: 0/3 identity-grounding, 1/4 tool-effect-honesty, and 0/5 authority-safety. Three outputs contained zero generated tokens. The valid negative observations are now classified in the report instead of crashing the evaluator. A supplemental exact Holm replay rejects all four primary likelihood nulls, but is explicitly post-hoc and cannot replace the preregistered behavior gate. The final status is small_checkpoint_transfer_not_proven; reasoning gain, resident-32B gain, frontier performance, promotion, and WOW Signal remain false.

Four failed proof attempts remain preserved as engineering evidence: a missing initial-adapter verifier field, an absent executable kernel probe, forbidden target rebinding inside the sandbox, and completion-envelope/state-schema drift. The final typed probe and verifier independently bind and replay each repair. The complete structured/recurrent-SFT matrix passes 239/239 in 369.67 seconds; the repaired focused boundary passes 42/42. Ruff, bytecode compilation, JSON/hash reconstruction, and diff hygiene pass. Sanitized evidence is artifacts/current/cp417_recurrent_sft_falsification_evidence.json.

SPARK-059 stays open until generated identity, tool-effect, and authority behavior passes under the same sealed standard. The next constructive checkpoint is SPARK-060's immutable verifier-backed paired episode, followed by structured delta reward and lexicographic error-introduction rejection. Those mechanisms must drive a fresh SPARK-059 rerun; they do not erase this negative result. Counting CP417 makes 663 total checkpoints. The 663-920 forecast leaves 0-257 records, or approximately 72.1%-100.0% checkpoint completion with an 86.1% midpoint. Final multi-hour soaks remain deferred.

2026-07-27 - CP418 recurrent-GRPO reward artifact boundary

The recurrent GRPO producer and its independent adapter-identity validator now share one executable artifact contract. Every accepted step binds the exact task and sample seed, execution-spec digest, response set, answer-channel counts and reasons, verifier rewards, effective rewards, verifier and shaped advantages, trajectory-credit evidence, step classification, optimizer update, and resulting policy digest. The validator reconstructs every fraction, advantage, trajectory cross-entropy score, shaping term, and protocol-to- training configuration binding rather than trusting producer summaries.

The verified reward domain remains [0, 1]. Optional trajectory shaping has a separate bounded domain and receipt, cannot appear when disabled, and can affect optimization only when the verifier group is degenerate. The resident campaign preregistration now passes its minimum-signal-group requirement explicitly on the frozen trainer command. Producer-to-validator tests exercise the real step artifact, malformed reward domains, answer-count drift, unreceipted shaping, trajectory replay mismatch, and configuration drift.

The complete recurrent-GRPO, recurrence-objective, trainer-state, resident preregistration, and post-training contract matrix passes 196/196. Ruff, bytecode compilation, and diff hygiene pass. No model was loaded or trained, and this checkpoint establishes no reasoning, resident-32B, frontier, promotion, or WOW Signal result. Sanitized evidence is artifacts/current/cp418_recurrent_grpo_artifact_schema_evidence.json.

SPARK-060 remains open. CP419 must implement the immutable verifier-backed pass-0/pass-1 transition episode; the following checkpoint must implement its structured delta reward, Error Introduction Ratio, and lexicographic right-to-wrong rejection before the small-checkpoint behavior campaign is repeated. Counting CP418 makes 664 total checkpoints. The 664-920 forecast leaves 0-256 records, or approximately 72.2%-100.0% checkpoint completion with an 86.1% midpoint. Final multi-hour soaks remain deferred.

2026-07-27 - CP419 immutable verified-transition episode

The pass-0/pass-1 training unit is now a reconstructable evidence object rather than a caller-authored reward claim. A private content-addressed store rejects symlink, hardlink, replacement, traversal, and overwrite attacks. The execution manifest rehashes every component file and binds the loaded base checkpoint, adapter stack, tokenizer, policy, personality, runtime, source closure, and generation worker. Verifier identity covers the live frontier scorer registry, same-module dependency closure, bytecode, defaults, closures, referenced globals, transitive module attributes, source files, and Python/runtime ABI.

The candidate sees only the canonical blinded task input. Execution spec, latent path, tools, evidence, world state, process receipt, uncertainty, diversity, and resources are committed before generation and must be marked non-candidate-visible. The campaign runner signs the exact token-level generation trace. A separately custodied contamination auditor uses the host resource observer to independently collect the live PID, process start time, executable hash, and open files; it rejects the pass unless every manifest component is actually open and rehashes to the committed bytes. Its signed payload binds that host observation, loaded component roots, manifest, model, response, and runner trace. The verifier policy also pins the observer implementation identity. The evidence verifier independently replays the scorer and signs authority only after both passes.

The same adversarial pass exposed an adjacent recurrent-GRPO reproducibility defect: final decoding used MLX's process-global RNG, so a previously evaluated gradient graph could change PPO admission despite the same declared seed and identical policy hash. The recurrent sampler now passes its full 32-bit seed into the engine's decode as a stateless key and performs no global RNG mutation. The previously failing randomized execution order is a fixed regression gate.

Calibration is oracle constructed: callers choose only a supported control kind, while the implementation derives the positive answer, missing-marker negative, and same-shape parsed-wrong negative. Every evaluated domain/scorer pair requires all three controls. The attempt ledger has one fixed six-event state machine, cryptographic chaining, owner-private and inode-stable access, and the macOS append-only flag. It is opened before task disclosure. A task- issuer signature then commits the empty ledger identity and immutable episode before generation; an evidence-verifier signature commits the terminal raw- content hash and event-chain head. The journal and final episode require both external checkpoints and distinct, ordered signed seconds.

The focused adversarial suite passes 50/50 in 21.55 seconds. Two disjoint adjacent matrices pass 290/290 in 36.86 seconds and 174/174 in 24.89 seconds, for 464 regression tests across transition, host-resource observation, frontier, campaign trust/journal, GRPO, resident preregistration/post-training, and recurrent-SFT contracts. Ruff, bytecode compilation, diff hygiene, and governance ownership review pass for CP419; governance still reports only unrelated concurrent-work deltas. Evidence is artifacts/current/cp419_verified_transition_episode_evidence.json.

No model was loaded or trained. CP419 establishes no reasoning gain, resident-32B effect, frontier result, promotion, or WOW Signal. SPARK-060 remains open. CP420 must make recurrent GRPO admit only validated transition episodes, independently reconstruct structured improvement reward and Error Introduction Ratio, and lexicographically reject right-to-wrong regressions before a fresh small-checkpoint behavior campaign. Counting CP419 makes 665 total checkpoints. The 665-920 forecast leaves 0-255 records, or approximately 72.3%-100.0% checkpoint completion with an 86.1% midpoint. Final multi-hour soaks remain deferred.

2026-07-27 - CP420A independently reconstructed transition reward

The first CP420 slice removes caller-authored labels and scalar rewards from the proof-grade recurrent objective. Each reward row now starts by replaying the complete CP419 episode and both independent verifier authorities. It derives correctness transition, verified parse/correctness information gain, independently normalized answer change, actual token/process cost, and unsupported behavior-policy confidence in integer micro-units. Correctness delta is mathematically larger than every auxiliary term combined.

Wrong-to-right and right-to-wrong are reported separately. EIR is reconstructed as right-to-wrong over the initially-correct opportunities; a zero denominator is recorded as undefined rather than numeric zero. Any right-to-wrong event is the first lexicographic rejection regardless of aggregate scalar reward. The proof entry point replays the reward receipt and binds prompt tokens, output tokens, policy digest, and exact behavior log probabilities before delegating to gradient construction.

The real signed test matrix creates independent wrong-to-right, right-to-right, and right-to-wrong CP419 episodes. A clean improvement plus anchor reaches the gradient delegate only after complete replay; an improvement paired with a regression never reaches it. Rehashed reward forgery and duplicate episodes fail closed. The focused suite passes 59/59 in 32.93 seconds, and the adjacent recurrent/adapter/resident proof matrix passes 103/103 in 35.11 seconds. Ruff and diff hygiene pass. Evidence is artifacts/current/cp420a_verified_transition_reward_evidence.json.

CP420 and SPARK-060 remain open. The producing branch, seed, and sampling configuration are not yet signed into a preregistered group; group completeness and anti-cherry-picking custody are not yet proven; raw reward APIs and the existing CLI can still bypass this proof entry point; and validation, immediate policy rehash, one optimizer mutation, durable anti-replay, and post-update receipt do not yet form one transaction. No model was loaded or trained, and no reasoning, resident-32B, frontier, promotion, or WOW Signal claim follows. Counting CP420A makes 666 total checkpoints. The 666-920 forecast leaves 0-254 records, or approximately 72.4%-100.0% checkpoint completion with an 86.2% midpoint. Final multi-hour soaks remain deferred.

2026-07-27 - CP420B signed group and exactly-once update

The optimizer group is now a preregistered object rather than a favorable set assembled after generation. The external task issuer signs the exact ordered episode IDs, task, RNG roots, policy, recurrent execution spec, producing branch, sample seed, sampling configuration, and reward policy before task disclosure and before either pass runs. Independent admission rejects omitted, duplicate, reordered, branch-substituted, seed-substituted, configuration- substituted, late-signed, and forged groups. CP419 execution-spec evidence now supports a proof-grade v2 binding to the exact recurrent execution spec while remaining able to read legacy v1 episodes.

An admitted group can mutate recurrent weights only through a journaled update transaction. It replays group admission, rehashes the current policy, publishes a governed create-once reservation keyed by the admission digest, constructs the gradient, rehashes immediately before mutation, calls optimizer.update exactly once, requires a changed policy digest, and publishes a create-once commit. The final receipt independently reopens and reconstructs both durable records. Replay remains blocked even after a simulated model rollback; policy drift during gradient construction leaves the optimizer untouched and burns the admission fail-closed.

Focused validation passes 67/67 in 60.00 seconds. The adjacent recurrent, adapter, resident preregistration, and post-training matrix passes 111/111 in 70.21 seconds. Ruff, diff hygiene, and the exact governance-effect baseline pass. Evidence is artifacts/current/cp420b_verified_transition_group_update_evidence.json.

CP420 and SPARK-060 remain open. A campaign-level append-only launch ledger must still prove that a producer did not discard and rerun an entire planned episode. The legacy trainer and raw scalar-reward APIs can still bypass this transaction, CP418 adapter identity does not yet replay the new group/update evidence, and reserved-but-uncommitted crash recovery remains a manual fail-closed reconciliation. No production optimizer or model was run, so this establishes no reasoning, resident-32B, frontier, promotion, or WOW Signal result. Counting CP420B makes 667 total checkpoints. The 667-920 forecast leaves 0-253 records, or approximately 72.5%-100.0% checkpoint completion with an 86.3% midpoint. Final multi-hour soaks remain deferred.

2026-07-27 - CP420C complete campaign custody and crash reconciliation

The task issuer now commits the entire ordered optimizer campaign before any campaign task is disclosed. The manifest binds every group manifest, its task-issuer attestation, ordered episode identities, trust policy, and sequence, and refuses duplicate groups or an episode reused across groups. A create-once ledger requires each group to start in order and reach an explicit updated, rejected, aborted, or indeterminate terminal record. The campaign cannot close with a missing group. Its close payload is signed by the separately custodied evidence verifier and independently reconstructs every raw start and terminal record, group manifest, task-issuer signature, evidence combination, chronology, and digest rather than trusting summary counts.

Reserved-but-uncommitted updates now have a durable reconciliation path. The admission is never silently retried: an unchanged policy is recorded as reserved_no_policy_change and still requires a fresh admission; a changed policy is recorded as policy_changed_without_commit and requires exact checkpoint recovery. Both outcomes burn the original admission. A committed update cannot be reconciled, and a reconciliation cannot be rewritten.

The verified-transition suite passes 70/70 in 73.81 seconds. The adjacent recurrent/adapter/trainer/preregistration/post-training matrix passes 144/144 in 76.87 seconds; the SPARK refusal preflight passes 11/11. Governance ownership matches 2,059 recognized calls in 1,922 buckets. Ruff, bytecode compilation, and diff hygiene pass. Evidence is artifacts/current/cp420c_verified_transition_campaign_evidence.json.

CP420 and SPARK-060 remain open for one final implementation slice: the legacy trainer/raw recurrent reward route must no longer bypass the verified transaction, and CP418 adapter identity must replay campaign, group admission, reservation, commit, and update source artifacts. No model was loaded or trained and no reasoning, resident-32B, frontier, promotion, or WOW Signal claim follows. The measured Fable pre-training legs reduce the planning range to approximately 7-12 total checkpoints before resident-32B training launch, 11-18 before a defensible preliminary live gain verdict, and 17-28 before a powered conditional WOW Signal decision. These are workload forecasts, not capability scores. Counting CP420C makes 668 total checkpoints. Final multi-hour soaks remain deferred.

2026-07-27 - CP420D campaign-bound mutation and commit recovery

The verified optimizer transaction can no longer run as an unattached group. It requires the matching campaign sequence to be durably started, replays the planned group manifest against that start, and refuses an already terminal group before constructing a gradient. A successful exactly-once update now publishes the campaign terminal record with the admitted-group and final-update digests as part of the same governed path.

The remaining post-commit crash window is recoverable without re-running an optimizer. Reservation and commit bytes reconstruct the exact update receipt; that receipt can publish the missing campaign terminal as a recovered committed update. Recovery cannot rewrite a terminal or turn an uncommitted reservation into success.

The combined verified-transition, recurrent objective, adapter identity, and trainer-contract matrix passes 123/123 in 76.03 seconds. Governance ownership, Ruff, compilation, and diff hygiene pass. Evidence is artifacts/current/cp420d_verified_transition_campaign_mutation_evidence.json.

CP420/SPARK-060 remain open for the legacy trainer/raw-reward bypass removal and CP418 source-artifact replay. No model ran and no capability claim follows. Counting CP420D makes 669 total checkpoints. The planning range is now 6-11 checkpoints to resident-32B training launch, 10-17 to a defensible preliminary live gain verdict, and 16-27 to a powered conditional WOW Signal decision. Final multi-hour soaks remain deferred.

2026-07-27 - F7 lane certification (Fable, out-of-band)

The F7 pre-training legs (SPARK-063 through SPARK-068, entries F7-A through F7-G above) were built while the march ran CP419-CP420B. What was actually verified, stated so the march does not have to take any of it on trust:

Focused: 402 tests across the F7 surface, all green. Where an instrument touches the model, the test runs a real MLX stack — a real Qwen2 through the real recurrent_hidden_states loop, the real attach_adapters against a real module tree, real CapabilityRegressionGuard / BatteryResult / MemoryBenchmarkResult objects, real generate_battery tasks.
Adjacent: a 1,733-test sweep over every latent-cortex, recurrence, campaign, journal, and distillation suite passed with zero failures. The intervention and calibration suites specifically pass 22/22 across the journal_state extraction, which is the change with the most reach.
Gates: compile, lint, layering, smoke, and governance-lint green. The governance-effect baseline gained exactly the two file-write gateway call sites the promotion registry legitimately introduced, and nothing else.
Enterprise ratchet: 80 high/critical findings at b1101c5f5 (before this lane) and 80 at HEAD — identical severity counts and zero findings in any file this lane created. The gate's red state against its 39-finding baseline is inherited debt, unchanged by this work.
make security fails on a secret_like_literal in core/brain/llm/latent_cortex/action_state_capture.py. This predates the lane, reproduces on b1101c5f5, and belongs to another owner; it is recorded here rather than repaired out-of-band.
Adversarial checks beyond the test suite, because a test suite proves the cases someone thought of:
- Accumulator vs. an independent reference. The MMR root was re-derived by a separately written naive implementation over 85 log sizes (1-79, 128, 255, 256, 1000, 1023, 1024), every index's proof verified, and a swapped leaf's proof rejected at every size. Clean.
- Witness/prefix differential. Over 40 randomized campaign shapes with retries after failures — 1,072 comparisons at every checkpoint position — the compact witness accepted exactly what the complete-prefix validator accepted, and both rejected every tampered journal. Zero disagreements.
- Promotion fuzz. 4,141 attempts to get an ADMIT out of a defective gate report: every one of the 127 subsets of dropped gates, every single gate graded zero, every single gate failing, and 4,000 randomized reports with renamed rows, duplicates, and counts under the floors. Zero admitted a defective report.
- Contamination fuzz. 3,832 STaR iterations across 1,500 randomized lineages with leaks injected at ~26% of iterations: 1,012 contaminations caught (613 same-iteration, 237 earlier-training, 162 reused-holdout) and zero contaminated lineages validated.
Blast radius of the live-path change. RecurrenceAdapterActivation.to_dict() gained two keys; a repository-wide search confirms its only callers are tests. The receipt that production actually consumes — WindowRunner.adapter_receipt(), validated by recurrent_grpo_adapter_identity.py — is built from separate counters and is byte-identical to before.

Two live defects were found by the new instruments and are recorded with their repairs in F7-D and F7-E: heldout_battery has no cross-battery exclusion (15.5% of seed pairs share a prompt, so the flywheel's seed floor separates seeds rather than content), and the adapter activation counter could not distinguish a whole treatment from a fraction of one.

One finding handed to the march rather than acted on. Two rows in falsification_matrix.py carry stale blocker lists: verifier_arms is ROW_BLOCKED on (39, 40, 41, 42, 43, 44, 45, 46) and fast_weight_controls on (55, 56). All ten of those items are now [x] in this ledger, so both rows name machinery that has already landed. The effect is not cosmetic — it makes two of twelve rows look blocked on something that no longer blocks them, which hides what they are actually waiting on (the trained treatment and the acceptance run). Correcting a blocker list falls in the Fable lane's F5 half; moving a row to runnable does not, because that needs a producer which genuinely exercises the verifier mesh, and the generative/counterfactual arms cannot run model-free — a row marked runnable while only its exact-checker arms execute would be the same defect this lane spent the day removing. So the observation is recorded and the decision left where it belongs. The third blocked row, adversarial_ood_variants, is correctly blocked: its variants are generated fresh at acceptance precisely so the dry run cannot see them.

No checkbox moved. No model of consequence ran. Nothing here is a reasoning, resident-32B, frontier, promotion, or WOW Signal result, and none of these checkpoints is counted against the march's record.

2026-07-27 - CP420E source-replayed training identity and freeze

CP418 adapter identity no longer accepts an optimizer count and final policy hash solely because the training receipt states them. The verified path opens the externally closed CP420 campaign, identifies the exact updated group set, and independently reconstructs every admitted group from its immutable paired episodes, reward evidence, samples, signed pre-generation manifest, scorer, and token codec. It then reconstructs every durable reservation, commit, and update receipt, binds those records to the corresponding campaign terminals, and requires a continuous policy hash chain across all mutations.

That replay receipt is cross-bound to the existing CP418 adapter identity: optimizer count and final policy must agree, and the immutable freeze certificate preserves the verified identity schema and replay digest. The freeze boundary has a distinct strict contract for this schema; a relabeled ordinary identity or a verified identity missing proof-grade mutation, source replay, positive update count, or final-policy evidence is refused.

The combined affected matrix passes 233/233 in split plugin-correct runs: 230 synchronous contracts plus three repository-async recovery contracts. It includes verified-transition campaigns and updates, recurrent objectives, trainer contracts, adapter identity/freeze, and the newly merged CP126 autonomous-brain routing contracts. Governance ownership remains at 2,059 recognized calls in 1,922 buckets; Ruff, compilation, and diff hygiene pass. Evidence is artifacts/current/cp420e_verified_transition_training_evidence.json.

CP418 source replay is closed. CP420 and SPARK-060 remain open because tools/train_grpo.py still directly grades scalar rewards, constructs the raw recurrent objective, and invokes the optimizer outside the verified campaign transaction. That is the next checkpoint. No model was loaded or trained, and no reasoning, resident-32B, frontier, promotion, or WOW Signal claim follows. Counting CP420E makes 670 total checkpoints. The current planning range is 5-10 checkpoints to resident-32B training launch, 9-16 to a defensible preliminary live gain verdict, and 15-26 to a powered conditional WOW Signal decision. Final multi-hour soaks remain deferred.

2026-07-27 - F8 lane: the two objectives ahead of the trainer (Fable)

SPARK-061 and SPARK-062 were reassigned from the march to the Fable lane while the CP419/CP420 lane removes the raw trainer bypass for SPARK-060. Both were bare — no module, no test, no receipt — so this is construction ahead of the trainer rather than a parallel re-implementation, and the claim was pushed to this ledger before any code was written.

What landed, and what it is not:

F8-A / SPARK-061: core/learning/progressive_recurrent_objective.py plus tools/measure_progressive_collapse.py. Four degeneracy detectors, a causal step-necessity lesion, a state-gradient measurement, the displacement-floor loss term, and a barrier solver that derives training constants from data.
F8-B / SPARK-062: core/learning/auxiliary_objective_curriculum.py. The seven-term registry with per-term liveness classification, a base-loss composite that structurally excludes head terms, a competence-gated depth curriculum with replayable receipts, and a train/inference depth-parity binding distinct from SPARK-024's kernel parity.

One inherited claim was checked and came back different. v4's docstring has asserted for several checkpoints that collapse to the identity is the cheapest solution to a monotone-improvement mandate. Measured on the untrained Qwen2.5-1.5B at depth 4, collapse is not the global optimum — but it is a local basin behind a measured ~0.19-nat barrier, which is the trap that actually matters and is falsifiable in a way the original phrasing was not. The same measurement caught this lane's own displacement floor being too low to fire at the basin it was written to remove; the constants are now solved from data and recorded as an operating-point-specific reference rather than a default.

Validation: 45 focused tests (22 SPARK-061, 23 SPARK-062) and a 301-test adjacent sweep across every recurrence-objective, curriculum, halting, process- critic, mistake-locator, update-acceptance, uncertainty, GRPO, and latent-cortex suite — zero failures. Ruff and compilation clean on both modules and the tool.

No model was trained. Neither checkbox moved: SPARK-061's "later states improve" is a property of the instrument's design until a resident treatment is measured under it, and SPARK-062's per-term gradient measurement has not been taken at the resident architecture. Nothing here is a reasoning, resident-32B, frontier, promotion, or WOW Signal result, and neither is counted against the march's checkpoint record.

2026-07-27 - F8 integration: the instruments are wired, not adjacent

The F8 modules computed on the real training path from the start — they import _prepare_live_path / _advance_recurrent_states / _persist_and_score and reuse v4's penalties, so the measurements are of real recurrence. But nothing called them: no preflight probed the new refusal surfaces and no registry knew they existed. An instrument with no caller is a correct orphan.

Three wires close that, all inside the Fable lane's own F3/F5/F7-G surfaces:

SPARK-003 threat model gains a thirteenth class, recurrence_inertness, bound to five executable checks across both new modules. The ledger's original twelve-class enumeration is a floor, not a ceiling: a threat model that cannot grow when a new failure is measured is a document rather than a registry. Its residual-risk line states the real gap — tools/train_grpo.py does not consult the detectors, so a campaign can still launch without a progressive report, and the pricing constants have only been solved on the 1.5B.
The pre-training preflight (F7-G) gains five probes, taking it from 11 to 16. Each hands an instrument the input it must refuse: a flawless four-step improvement curve from a dead operator, a resealed report relabelled as progress, a declared term with no gradient path, a curriculum depth inference cannot execute, and a falsification blocker naming a closed ledger item. All 16 fire.
SPARK-070's stale blockers are repaired at the cause. F7 handed this over as an observation: verifier_arms was blocked on SPARK-039..046 and fast_weight_controls on SPARK-055/056, all ten of which had closed. The repair is not the list. MatrixRow now carries a typed blocked_reason (open_spark_items / producer_absent / acceptance_run_only), because the three have completely different remedies and collapsing them into an integer list is what let the drift hide; and validate_blockers_against_ledger parses the ledger's own checkboxes and refuses any blocker naming a closed item. Both rows are now honestly producer_absent — the machinery landed, the producer has not — which is what they were actually waiting on.

Why the drift survived: a test was enforcing it. The old test_blocked_rows_name_their_blockers asserted the exact stale sets set(range(39, 47)) and {55, 56}. It has been replaced by contracts that pin the reason rather than a frozen id list, plus the ledger cross-check, so the class of defect is now detectable instead of one instance being fixed.

Validation: 93 focused across the two new suites plus threat model, matrix and preflight; the 16-probe preflight returns READY. Two pre-existing tests were updated because they pinned contracts this change deliberately strengthened — stated here rather than left for someone to find in a diff.

2026-07-27 - CP420F CP420E adversarial proof repair

CP420E failed its independent adversarial review on four proof boundaries, so its affected freeze and source-custody claims are superseded here. The campaign ledger no longer validates one pair of group records and returns a second reread pair. Admission policy A cannot be joined to an update beginning from policy B. Exact-adjoint objective evidence is now a create-once durable record written before mutation and replayed through commit recovery. The final adapter policy is recomputed from frozen safetensor names, dtypes, shapes, and values under the same digest algorithm used on the live recurrent policy.

Verified adapter identities now have an exact reconstructable schema containing the complete base identity and sealed transition replay receipt. Frozen GRPO verification reruns the ordinary identity validator over the immutable files and compares its output to that embedded base identity; fabricated proof fields and stale adapters therefore fail rather than inheriting a campaign's result.

The integrated proof/F8 objective matrix passes 147/147, with one direct regression for each review finding plus 3/3 live-MLX tensor-dtype parity checks. Ruff, compilation, and diff hygiene pass. Evidence is artifacts/current/cp420f_verified_transition_proof_repair_evidence.json.

This restores CP418 source replay on a defensible boundary. It does not close CP420/SPARK-060: tools/train_grpo.py still owns the raw recurrent mutation route, and replacing that route remains next. F8's SPARK-061/062 implementation has landed, but its own ledger correctly leaves both checkboxes open until a resident treatment measures the progressive objective and every auxiliary term's gradient. No model loaded, no training occurred, and no reasoning, resident-32B, frontier, promotion, or WOW Signal claim follows. Counting CP420F makes 671 total checkpoints. The recalibrated range is 3-8 checkpoints to resident training launch, 7-14 to a defensible preliminary verdict, and 13-24 to a powered conditional WOW Signal decision. Final multi-hour soaks remain deferred.

2026-07-27 - F8 adversarial self-review: two forgeries the validators allowed

The same scrutiny the collapse sweep turned on v4's docstring, turned on F8's own receipts. It found two live defects, both the same class and both in the validators whose entire job is to prevent forged claims: the aggregates replayed, but the atoms were trusted.

validate_progressive_report checked only summary fields. A forger who edited the trajectory rows and the summary consistently, then resealed, passed every aggregate check while the rows underneath said "collapsed". Demonstrated: a report whose trajectories carry 1e-7 displacement, declaring min_displacement 0.5 and verdict real_progress, validated.
validate_liveness_report trusted each row's recorded liveness. Relabelling a zero-gradient term live and updating live_terms to match passed, because the aggregates were derived from the labels they were supposed to police.

Both now rebuild the entire receipt from its own rows and require equality — a commitment proves nobody altered the bytes after signing, and says nothing about whether the signer's arithmetic was honest, so the arithmetic is redone.

A third finding was a limit rather than a bug, and is named instead of papered over. Recomputing verdicts closes label forgery but cannot close input forgery: the rebuild consumes the same share it is checking, so a forged share is reproduced faithfully. Pretending otherwise would be the "absence of a check reported as a passed check" pattern this codebase keeps finding. The receipt therefore records shares_source, and liveness_from_composite derives shares from base_weight_loss's measured contributions so the input leaves the caller's hands entirely — a stronger guarantee than any amount of re-validation.

Validation: 49 focused, 277 across the affected surface, preflight 16/16 READY. Three pre-existing expectations were updated because the refusals now fire at the rebuild rather than at a single rule.

2026-07-27 - F8: the fast-weight row gets its producer (9 of 12 runnable)

experiments.run_fast_weight_controls closes SPARK-070's fast_weight_controls row, which this lane had just re-labelled producer_absent after finding its blocker list stale. The row moves blocked → runnable; the matrix is now 9 runnable, 1 enforced, 2 blocked.

Two design points carry the checkpoint:

The graded claim is on-versus-sham, not on-versus-off. An optimized episodic delta beating no delta shows only that a bounded perturbation of the function helped; it cannot separate learned adaptation from the regularizing effect of any change that size. The sham arm applies a delta of matched magnitude in a random direction, so what survives is a claim about DIRECTION. The off arm still runs and is reported — a case where sham also beats off is worth seeing — but the claim does not rest on it.
Erasure gates every observation. Query-scoped fast weights are only safe because they are erased afterwards, so a task whose erase was not proven may have contaminated every task after it. An unproven erase quarantines the whole task rather than counting it (and is distinguished from a refuted erase, which fails the run outright); a partially counted task would silently unbalance the pairing. Arms are counterbalanced per task against a committed order, as the latent-opt control already does, so ordering effects cannot load onto one arm.

verifier_arms stays blocked deliberately. Its generative and counterfactual arms cannot run model-free, and a row marked runnable while only its exact-checker arms execute would restate the defect the verifier mesh was built to remove.

20/20 matrix contracts including quarantine, refutation, an off arm that wrongly claims an erase, and a malformed solver outcome; 176 across the affected surface; preflight 16/16 READY.

2026-07-28 - CP420N: the progressive objective reaches the bounded adjoint

The CP420M launch lane revealed that F8's instruments were still adjacent to the optimizer. The production trainer could execute an exactly-once verified GRPO update without consulting SPARK-061 or SPARK-062. CP420N closes the mathematical substrate, not the final trainer wire.

recurrence_native_objective_v2 can now replay intermediate-state objective cotangents through the same bounded exact adjoint used by resident GRPO: detached-shallow improvement hinges, displacement-floor anti-collapse pressure, oscillation pressure, and terminal branch diversity. A monolithic full-unroll oracle and the streamed adjoint agree on identical tiny-Qwen weights within 2e-5 in value and 3e-4 maximum absolute difference for every trainable gradient tensor. Producing-branch identity is explicit.

progressive_recurrent_objective now measures a real multi-branch graph in one exchange-coupled unroll. Its sealed trajectory set carries branch role/index, execution graph, loss, displacement, and drift, and independently rebuilds derived summaries from the serialized atoms. It does not substitute two single-branch runs for a two-branch computation.

The auxiliary liveness report now requires a head's own gradient, optimizer mutation, and before/after parameter identity. This corrects an F8 semantic hole where absence of a base-weight gradient could be mistaken for evidence that a process, mistake-location, or accept/discard head actually learned.

The checkboxes stay open. The verified group update does not yet admit this trajectory receipt; causal-lesion and learned-stopping terms are not in the composite; the three head objectives lack resident exactly-once transactions; and the 32B operating point has not calibrated collapse floors or per-term liveness. No resident model was loaded, no training occurred, and no reasoning, frontier, promotion, or WOW Signal result follows. Evidence: artifacts/current/cp420n_exact_adjoint_trajectory_evidence.json. The expanded affected matrix passes 321/321 in 142.55 seconds.

2026-07-28 - CP420O: the verified optimizer consumes the trajectory objective

CP420N's bounded trajectory adjoint now enters the same reserved, journaled, exactly-once update as verifier-weighted terminal GRPO. Positive verified advantage determines which completion branches receive improvement credit; displacement, oscillation, and diversity are evaluated once over the whole exchange-coupled ensemble. The durable source binding is reconstructed at final evidence validation from admission, reward, sample, prompt, frozen spec, canonical config, and advantage clip rather than accepted from the objective journal.

Exact-adjoint proof receipts have a distinct v2 schema and commit the exact policy, token inputs, token weights, graph, branches, and objective configuration. Resealed child transplants across policy, prompt, answer, bridge, or weights fail against independently admitted group evidence. The unchanged trajectory-config schema remains v1, preserving preregistration and historical protocol compatibility. Nonempty bridge tokens fail closed until their provenance is included in the admitted replay package.

The SPARK-061 and SPARK-062 checkboxes remain open. The production path can now optimize improvement, anti-collapse displacement, oscillation, and diversity, but no resident treatment has established monotonic later-state quality or useful 32B gradients. Causal-lesion and learned-stopping terms remain absent, and the three auxiliary heads still lack separate admitted optimizer transactions and resident labels. The canonical config is archived and verifiable but is not enabled by the default resident campaign.

Ruff and diff hygiene pass; the final expanded matrix passes 250/250 in 147.04 seconds. Evidence: artifacts/current/cp420o_verified_trajectory_composite_evidence.json. No resident model was loaded or trained, and no reasoning, frontier, promotion, or WOW Signal result follows.

2026-07-28 - CP420P: verified intervention gradients enter the mutation

The verified optimizer now consumes causal-transition lesions and a learned stopping opportunity objective. A lesion skips one full exchange-coupled transition, replays the suffix detached, and places the resulting hinge gradient only on the intact policy path. The stopping teacher derives a cost-aware soft distribution over candidate depths from answer loss plus ponder cost; its probabilities are detached supervision, while the candidate losses remain differentiable.

The v2 group gives both intervention and improvement terms positive verified- advantage L1 credit and retains exactly one structural objective over all branches. Its canonical config, source binding, journal, exactly-once transaction, campaign close, and final replay are covered end to end. Adversarial cases reject false causal arithmetic, reversed stopping probabilities, malformed teacher rows, intervention receipts injected into the structural slot, and omitted or relabeled trust boundaries.

The proof boundary is explicit. Receipt validation proves arithmetic and custody over producer-sealed measurement atoms. It does not independently reconstruct the latent states from pre-update policy tensors. The contract now requires external state replay, and that custody/recomputation lane is the next bounded checkpoint. Until it lands, CP420P does not describe its receipts as independent state proof.

Intervention training-evidence v3 carries that limitation into adapter identity instead of collapsing it into a broad source-replay boolean. Preregistration rejects lesion or stopping depths outside the frozen execution graph. Historical non-intervention training evidence remains v2-compatible.

The direct matrix passes 191/191 in 102.39 seconds and the disjoint downstream campaign, launch, external-manifest, provider, transaction, and post-training matrix passes 148/148 in 18.35 seconds. Evidence: artifacts/current/cp420p_verified_intervention_objectives_evidence.json. The final independent adversarial re-review approved the repaired boundary and preregistration contracts with no remaining findings in scope. SPARK-060/061/062 remain open; no resident model was loaded or trained, and no reasoning, frontier, promotion, or WOW Signal result follows.

2026-07-28 - CP420Q: exact initial policy state enters durable custody

Training and future independent replay now share one source-bound recurrent adapter constructor. It preflights the complete decoder-layer and projection inventory before mutating the model, then seeds and attaches the exact LoRA topology used by the proof campaign. Missing, ambiguous, out-of-window, or already wrapped sites fail before a partial policy can exist. The Adam constructor is shared as well, with canonical hexadecimal encodings for the learning rate, betas, and epsilon and an explicit bias-correction setting.

The initial-policy probe advances to v2 without invalidating historical v1 evidence. A v2 probe permanently binds both initial_adapter.safetensors and initial_optimizer.safetensors: stable private file identity, byte digest, byte count, exact sorted tensor inventory, inventory digest, complete LoRA factor pairs, adapter policy digest, and the exact Adam moment topology. The trainer writes each snapshot immutably and reopens it through the same independent inspector used by launch materialization.

Intervention launch materialization now refuses a hash-only v1 probe. It copies both state artifacts into the private launch root, reopens and verifies the copied bytes, and embeds a cross-bound custody contract in the frozen provider configuration. Completed-launch recovery repeats that verification instead of trusting the archived JSON. This establishes the first state in an updates + 1 policy-state chain without duplicating the same pre-state for every update: the initial state is stored once, and each successful transaction's post-state can become the next update's pre-state.

The proof boundary remains explicit. CP420Q proves exact initial adapter and optimizer custody and deterministic reconstruction prerequisites. It does not yet append a pre-measurement record for each admitted update, independently recompute the trajectory/lesion/stopping objective or gradients, or replay the Adam transition to byte-identical post-state. Those are CP420R and CP420S; no external_policy_state_replayed=true claim is permitted before they pass.

The final source-state focused matrix passes 87/87. The disjoint transaction, episode, trainer, and adapter-identity matrix passes 137/137, and the launch, external-replay, campaign, runner, production-factory, and provider matrix passes 84/84. Ruff, bytecode compilation, and diff hygiene pass. Evidence: artifacts/current/cp420q_initial_policy_state_custody_evidence.json.

SPARK-060/061/062 remain open. No resident model was loaded or trained, and no reasoning, frontier, promotion, or WOW Signal claim follows. CP420Q is total checkpoint 687. The 687-920 completion envelope is approximately 74.7%-100.0%, with an 87.3% midpoint. The evidence-adjusted range is 3-5 checkpoints to a sealed resident-32B training launch, 7-11 to a defensible preliminary gain verdict, and 12-21 to the powered conditional WOW Signal decision. Final multi-hour soaks remain deferred.

2026-07-28 - CP420R: pre-objective state lineage closes the measurement gap

Every admitted intervention update now enters a single immutable state chain before objective evaluation. The scoped reservation binds campaign sequence, execution spec, group manifest, and the requirement for a pre-measurement. The new intent then binds that reservation to the exact adapter and Adam state, the CP420Q origin or latest complete transaction post-state, campaign and schedule roots, trainer static inputs, trajectory source, exact recurrent GRPO configuration, and empty intervention bridge-token channel. Adapter and optimizer equality is exact over keyset, shape, dtype, and value.

The intent is not an isolated receipt. Pending transaction v4 binds both its digest and the reservation digest, and the transaction store reopens the immutable intent before accepting staged post-state tensors. Objective record v2, reservation v2, update commit, causal campaign manifest v4, and external replay preserve the same scope. Initial and staged safetensors are hashed and MLX-loaded through the same stable owner-private no-follow descriptor, closing the path-swap interval between byte verification and tensor loading. Historical non-intervention reservation/objective/transaction/manifest schemas remain replayable.

Crash behavior is fail-honest and progress-safe. A reservation-only, intent-only, or objective-before-stage interruption is reconciled only after the durable trainer checkpoint has restored the exact adapter and optimizer state. The admission is permanently burned, an immutable recovery-halt receipt is published, and the trainer requires a fresh campaign. Changed policy, optimizer drift, dtype-only drift, conflicting scope, missing intent, and artifact substitution all fail closed. Rejected groups neither publish intents nor advance the successful-update ordinal.

The direct measurement-chain, transaction, episode, trainer, launch, campaign, external-replay, and trainer-contract matrix passes 222/222 in 97.09 seconds. The disjoint runner, policy-probe, provider, and rejection-transaction matrix passes 49/49. Ruff, bytecode compilation, and diff hygiene pass. Evidence: artifacts/current/cp420r_pre_measurement_state_chain_evidence.json.

This proves pre-objective policy-state lineage, not independent arithmetic. CP420S still must reconstruct the recurrent model from the sealed pre-state, recompute every trajectory, lesion, stopping, and structural objective atom and gradient tensor, replay the frozen Adam transition, and require identical post-state tensors. external_policy_state_replayed therefore remains false. SPARK-060/061/062 remain open; no resident model was loaded or trained and no reasoning, frontier, promotion, or WOW Signal claim follows.

CP420R is total checkpoint 688. The 688-920 completion envelope is approximately 74.8%-100.0%, with an 87.4% midpoint. The evidence-adjusted range is 2-4 checkpoints to a sealed resident-32B training launch, 6-10 to a defensible preliminary gain verdict, and 11-20 to the powered conditional WOW Signal decision. Final multi-hour soaks remain deferred.

2026-07-28 - CP420S1: independent policy-transition replay kernel

The external-state proof now has a strict arithmetic kernel rather than a campaign-level boolean awaiting implementation. Given an independently constructed model, CP420S1 restores exact sealed adapter and Adam pre-state, recomputes an independently supplied objective, requires exact equality with the producer objective receipt, inventories every gradient tensor by sorted name, shape, dtype, and evaluated-value digest, reconstructs the frozen Adam configuration, applies exactly one update, and requires exact adapter and optimizer post-state equality. The objective is also forbidden from mutating policy state before the optimizer step.

The positive receipt is itself strictly reconstructed. Missing, duplicate, or reordered tensors; adapter/gradient topology mismatch; optimizer substitution; objective drift; gradient drift; adapter or optimizer value/dtype/key drift; and resealed false success flags all fail closed. The direct and downstream matrix passes 47/47, including existing pre-measurement, transaction, exact-adjoint intervention, and optimizer-construction paths. Ruff, formatting, bytecode compilation, and diff hygiene pass. Evidence: artifacts/current/cp420s1_policy_state_replay_kernel_evidence.json.

This is CP420S's independently testable arithmetic core, not yet its external campaign proof. CP420S2 must bind the frozen model, execution spec, exact objective inputs, gradients, and state custody into the campaign manifest; CP420S3 must execute this kernel inside the pinned externally custodied verifier under a resident-scale durable lifecycle. Until both pass, external_policy_state_replayed remains false at campaign level. No resident 32B was loaded or trained, and no reasoning, frontier, promotion, or WOW Signal claim follows.

CP420S1 is total checkpoint 689. The 689-920 completion envelope is approximately 74.9%-100.0%, with an 87.4% midpoint. The evidence-adjusted range is 1-3 checkpoints to a sealed resident-32B training launch, 5-9 to a defensible preliminary gain verdict, and 10-19 to the powered conditional WOW Signal decision. Final multi-hour soaks remain deferred.

2026-07-28 - CP420S2: resident replay inputs enter one exact custody contract

The independent replay kernel now receives a single cross-bound contract for the complete computation it is expected to reproduce. The contract binds the fresh resident-32B preregistration, full checkpoint fingerprint, behavior bundle, raw and semantic execution-spec identities, every executable source binding, exact initial adapter and optimizer custody, frozen Adam configuration, recurrent-GRPO configuration, verified trajectory and intervention configuration, and resident-scale verifier budget. Verification can reopen all files and recompute the full model and behavior identities instead of trusting archived hashes.

The production preregistration now activates the two-branch composite-v2 trajectory, lesion, and stopping objective rather than leaving the verified intervention path as a test fixture. Its group topology exactly matches the execution-spec branch topology. Launch materialization reconstructs the replay contract from the preregistration and copied custody artifacts, embeds it in the immutable provider configuration, reopens it on recovery, and propagates the exact contract to finalization. Provider-compatible custody documents keep finite scientific parameters as canonical JSON strings, preserving their values while keeping signed provider receipts free of ambiguous JSON floats.

A worktree-specific production blocker was also closed. Safe repository relative artifacts can now resolve against either the active worktree or its authenticated main checkout, allowing one resident model copy without weakening traversal or symlink confinement. The regenerated CP420S2 preregistration recomputed and verified the real 32B model, binds 47 source roles, 288 training tasks, and 2,877 powered confirmatory tasks, and remains correctly ineligible for claims.

The expanded replay, provider, production-factory, measurement-chain, transaction, causal-campaign, launch-bundle, training-runner, and external replay matrix passes 147/147. Ruff, formatting, bytecode compilation, and diff hygiene pass. Evidence: artifacts/current/cp420s2_policy_state_replay_custody_evidence.json.

This is sealed input custody, not external execution evidence. The resident model has not been loaded or trained, the campaign-specific v2 initial-policy probe does not yet exist, and no external verifier has replayed a transition. CP420S3 must run the CP420S1 kernel under a pinned, durable, independently custodied verifier lifecycle with crash recovery and an immutable result receipt. external_policy_state_replayed, reasoning gain, frontier gain, promotion, and WOW Signal therefore remain false.

CP420S2 is total checkpoint 690. The 690-920 completion envelope is approximately 75.0%-100.0%, with an 87.5% midpoint. The evidence-adjusted range is one checkpoint to the final guarded resident-32B launch boundary, 4-8 to a defensible preliminary gain verdict, and 9-18 to the powered conditional WOW Signal decision. Final multi-hour soaks remain deferred.

2026-07-28 - CP420S3: durable external replay closes the guarded launch boundary

CP420S1's arithmetic kernel and CP420S2's exact resident inputs now execute through a production external lifecycle. The dedicated replay worker loads the bound model once, constructs the exact recurrent policy, restores sealed adapter and Adam pre-state, rebuilds every verified objective from producer evidence, recomputes all gradients, applies the frozen optimizer transition, and requires exact adapter and optimizer post-state. The detached broker and worker exchange bare canonical JSON request, batch, and result documents, eliminating newline or serialization differences from the authenticated protocol.

Replay custody survives caller death and sleep through a detached supervisor, bounded timeout, process containment, and create-once private request, candidate, and authoritative result files. The replay worker is not the close signer. A provisional immutable v4 campaign manifest is replayed first; only validated per-transition results may then be published and cross-bound into a v5 manifest. The v5 verifier independently reconstructs the expected objective, source stage, policy lineage, replay-contract identity, and ordered receipt root from producer custody. A closed campaign recovers from that sealed evidence without rerunning the 32B transition.

The exact production route passed a real tiny-Qwen transition, including objective, gradient, adapter, and Adam equality. The combined prelaunch matrix passes 332/332 in 115.51 seconds, with Ruff, formatting, compilation, diff hygiene, and resident preregistration verification green. The regenerated contract binds 51 source roles, resident fingerprint 8eae71e73a14d1228a942d4faf84690d70b62148f25b0f924435535f7c550fad, behavior bundle 7eb54f8b670bba6e5b37ae5f9c6d016e9628034951b3ee702ca0096123c5e674, 288 training tasks, 2,877 powered confirmatory tasks, and 17,262 confirmatory cells. Evidence: artifacts/current/cp420s3_durable_external_policy_replay_evidence.json.

The claim boundary remains strict. No resident 32B was loaded, probed, replayed, or trained in this checkpoint; therefore production external_policy_state_replayed, SPARK-060/061/062 closure, reasoning gain, frontier gain, promotion, and WOW Signal all remain false. CP420S3 reaches the guarded launch boundary and pauses there for explicit user authorization before the first active resident command.

CP420S3 is total checkpoint 691. The 691-920 completion envelope is approximately 75.1%-100.0%, with an 87.6% midpoint. The evidence-adjusted range is 3-7 checkpoints to a defensible preliminary gain verdict and 8-17 to the powered conditional WOW Signal decision. Final multi-hour soaks remain deferred.

2026-07-29 - CP420S4: the frozen worktree can own the shared resident model

The authorized CP420S3 initial-policy probe terminated before model load after 15.151574 seconds. Contract verification had correctly authenticated the single resident checkpoint stored in the main checkout, but the trainer later passed the unresolved worktree-relative path to model-lane ownership, MLX, and recurrent sampling. The failed detached receipt is retained; no campaign state was rewritten and no resident tensor was loaded or mutated.

train_grpo now preserves ordinary direct model paths while adding a confined fallback across authenticated roots of the same Git repository. Traversal, noncanonical relative paths, and outside-root resolution remain inadmissible. The resulting absolute path is used consistently for lane admission, model load, identity, tokenizer traces, and recurrent episodes.

The current-main targeted matrix passes 58/58 and the fresh CP420S4 preregistration verifies. Contract d2e6f659f7acac0c89964ec3791fd87d5474092ce03cc5a8e1b143da6d40b05d binds 51 source roles, 288 training tasks, 2,877 powered confirmatory tasks, and 17,262 confirmatory cells. Evidence: artifacts/current/cp420s4_worktree_model_resolution_evidence.json.

This closes the path defect, not the scientific campaign. Reasoning gain, frontier gain, promotion, and WOW Signal remain false until fresh resident evidence passes. CP420S4 is total checkpoint 692. The 692-920 completion envelope is approximately 75.2%-100.0%, with an 87.6% midpoint. The evidence-adjusted range remains 3-7 checkpoints to a defensible preliminary gain verdict and 8-17 to the powered conditional WOW Signal decision. Final multi-hour soaks remain deferred.

2026-07-29 - CP420S5: recurrent training and replay share one source closure

The CP420S4 resident policy probe proved the worktree-safe model path by loading the 32B and attaching 24 recurrent projections. It then rejected its own initial-policy receipt before mutation because the trainer and adapter identity disagreed about the executable source closure. The failed detached campaign is retained exactly and performed no optimizer update.

The trainer now constructs its source inventory in one place and checks exact parity with the adapter identity before model admission. The frozen closure includes the pre-measurement chain, exact policy-state replay kernel, detached replay worker, durable resume helper, and external verifier job. Training can no longer certify one source graph and publish an adapter claiming another.

The relevant current-main matrix passes 95/95, with Ruff, formatting, compilation, diff hygiene, source parity, and fresh preregistration verification green. Contract 3cec8a8fc72742f125e17a8b824d82efe9cd920f921eb4fa38cf94aaf42be718 binds 51 campaign source roles, 288 training tasks, 36 holdouts, 2,877 powered confirmatory tasks, and 17,262 confirmatory cells. Evidence: artifacts/current/cp420s5_recurrent_source_closure_evidence.json.

No scientific claim is promoted by this repair. Reasoning gain, frontier gain, promotion, and WOW Signal remain false pending a successful fresh resident campaign. CP420S5 is total checkpoint 693. The 693-920 completion envelope is approximately 75.3%-100.0%, with an 87.7% midpoint. The evidence-adjusted range remains 3-7 checkpoints to a defensible preliminary gain verdict and 8-17 to the powered conditional WOW Signal decision. Final multi-hour soaks remain deferred.

2026-07-29 - CP420S6: resident gates pass; host-isolated trust is executable

The fresh CP420S5 resident 32B passes the deterministic initial-policy probe and the recurrent answer-channel preflight. Its 24 intended projections produce sealed initial policy dd92ca2c...f46d8e; the read-only answer lane returns 6/6 valid contracts and 6/6 correct task answers. Neither run performed an optimizer update.

Adversarial review found seven material defects in the first signer draft: campaign-manifest rejection, false external-organization classification, unsigned purpose/operation, non-idempotent partial JIT signing, verifier/close separation, unbounded command output, and nonrestartable production role keys. The corrected protocol distinguishes host-isolated research custody from independent external custody and binds campaign, protocol, operation, purpose, and idempotency inside every Ed25519 payload. Durable JIT intents make signature retries byte-identical. The verifier journals accepted artifact replay and will not sign a close carrying any other receipt.

Live role keys use restartable same-user macOS Keychain custody; tests remain ephemeral. Broker output is bounded without memory accumulation, and command artifacts must be owned, single-linked, and nonwritable by group/world. The affected signer, trust, materializer, provider-factory, episode, and bundle matrix passes 154/154. Evidence: artifacts/current/cp420s6_resident_probe_and_host_signer_evidence.json.

This is strong same-host process isolation, not independent external administration, and the later claim must preserve that distinction. CP420S6 is total checkpoint 694. The 694-920 completion envelope is approximately 75.4%-100.0%, with an 87.7% midpoint. The evidence-adjusted range is 2-6 checkpoints to a defensible preliminary gain verdict and 7-16 to the powered conditional WOW Signal decision. Final multi-hour soaks remain deferred.

2026-07-29 - CP420S7: verified task coordinates are wire-portable

CP420S7 passed the resident policy and 6/6 answer-channel gates, then failed closed during materialization before training. The first recurrence task carried a deterministic unsigned 128-bit seed into a protocol that deliberately limits JSON integers to signed 64-bit for cross-language replay.

Curriculum coordinates now retain 63 deterministic bits, exceeding the 60-bit entropy floor while remaining wire-portable. Every one of the 324 train and holdout tasks now builds a verified commitment, and the answer-authority consumer recognizes the operation/purpose-bound role payload v2. The focused matrix passes 84/84. Evidence: artifacts/current/cp420s7_verified_task_integer_portability_evidence.json.

No optimizer update or gain claim occurred. CP420S8 must be freshly frozen and pass the resident gates before materialization resumes.

2026-07-29 - CP420S8: preregistration and transition evidence use distinct JSON domains

The fresh CP420S8 resident policy probe passed with all 24 recurrent projections attached, and the held-out answer channel returned 6/6 valid, correct answers. Four distinct Keychain-backed host signer roles were then provisioned without exporting private keys. Materialization failed closed before campaign-open publication or optimizer mutation because the archive publisher and independent reopener attempted to serialize the finite-float preregistration through the verified-transition integer-only JSON codec.

The launch archive now preserves the exact bytes of the already canonical and validated preregistration. Its independent verifier uses the same canonical finite-JSON domain only for that preregistration and its internal digest. Provider state, policy, commitments, nonces, launch metadata, and transition evidence remain on the stricter no-float codec. The materializer crash/recovery matrix now includes a finite-float preregistration and passes 18/18; the adjacent launch, source-binding, training-contract, and initial-policy matrix passes 96/96.

CP420S8 is immutable and retired because the repair changes source-bound launch code. It performed no optimizer update. Evidence: artifacts/current/cp420s8_preregistration_archive_serialization_evidence.json. Reasoning gain, frontier gain, promotion, and WOW Signal remain false. CP420S8 is total checkpoint 696. The 696-920 completion envelope is approximately 75.7%-100.0%, with an 87.8% midpoint. The evidence-adjusted range is 3-7 checkpoints to a defensible preliminary gain verdict and 8-17 to the powered conditional WOW Signal decision. CP420S9 must be freshly frozen and pass the resident gates before launch materialization resumes.

2026-07-29 - CP420S9: archive digest matches the frozen contract convention

CP420S9 again passed the resident policy and 6/6 answer-channel gates. Launch materialization published the signed campaign-open state and reached the independent archive reopener, which rejected the preregistration internal digest before training because the contract freezer hashes canonical finite JSON with one terminal newline while the reopener checked only the bare canonical form.

The independent reopener now accepts the same two exact forms already admitted by the preregistration validator: canonical finite JSON, with or without one terminal newline. It still rejects arbitrary whitespace, alternate float representations, and every noncanonical encoding. The launch and source-bound matrix passes 114/114 using the production newline digest form.

CP420S9 is immutable and retired by the verifier repair. It performed no optimizer update. Evidence: artifacts/current/cp420s9_preregistration_digest_compatibility_evidence.json. Reasoning gain, frontier gain, promotion, and WOW Signal remain false. CP420S9 is total checkpoint 697. CP420S10 must be freshly frozen and pass both resident gates before launch resumes.

2026-07-29 - CP420S10: host-isolated launch recovery preserves its claim boundary

CP420S10 passed both resident gates and successfully materialized and reopened its signed 288-task launch bundle. The required second materialization then failed closed because completed-launch recovery hard-coded the independent-external-custody receipt value and rejected the honest host-isolated claim boundary written by the first run.

Recovery now admits only the two defined boundary values during structural validation, independently reopens the signed trust policy, derives the one correct value from its verified custody classes, and rejects any mismatch. The crash/recovery matrix now includes a host-isolated receipt published immediately before failure and passes 115/115.

CP420S10 is immutable and retired by this source repair. It performed no optimizer update. Evidence: artifacts/current/cp420s10_host_custody_reopen_evidence.json. Reasoning gain, frontier gain, promotion, and WOW Signal remain false. CP420S10 is total checkpoint 698. A fresh CP420S11 contract and resident gates are required.

2026-07-29 - CP420S11: production admission reads the canonical branch graph

CP420S11 passed both resident gates, materialized a signed 288-task bundle, and reopened that exact bundle twice. The detached trainer loaded the resident 32B, attached all 24 recurrent projections, and then failed before its first optimizer update because production provider admission read a stale RLCExecutionSpec.branches attribute.

The execution contract represents virtual width canonically as branch_roles; admission now compares len(branch_roles) with the signed provider branch count. Production-factory fixtures use that real interface, so the stale field cannot return unnoticed. The focused matrix passes 115/115 and the provider-factory suite passes 24/24.

No optimizer update occurred. Evidence: artifacts/current/cp420s11_production_branch_contract_evidence.json. Reasoning gain, frontier gain, promotion, and WOW Signal remain false. CP420S11 is total checkpoint 699. CP420S12 is the next fresh launch campaign.

2026-07-29 - CP420S12: resident 32B recurrent training admitted

CP420S12 froze the repaired source at f9ea9bc3c, passed the resident initial policy probe with all 24 recurrent projections attached, and passed the answer-channel gate with 6/6 valid and correct outputs. Four distinct, Keychain-backed host signer roles were provisioned without exporting private keys. Two independent launch reopenings produced the same signed 288-task bundle, 639191d233e21724ebe9eb787538a6a18dadab99aea95417e791f9b0bdc3abc8.

The detached resident trainer then loaded the 32B, admitted the 26.88 GB MLX envelope, attached all 24 projections, crossed the former provider branch contract failure, and entered the 36-task recurrent baseline. Supervisor, worker, signer services, and a worker-bound caffeinate -i process were live at the launch checkpoint. The frozen execution graph uses two branches and four recurrent steps; the curriculum spans depths 2, 4, and 8, with causal intervention probes at steps 1, 2, and 4. No relevant path is capped at one.

This checkpoint proves launch admission and durable execution, not an optimizer update or intelligence gain. Host process isolation is not independent external administration. Evidence: artifacts/current/cp420s12_resident_32b_training_launch_evidence.json. Reasoning gain, frontier gain, promotion, and WOW Signal remain false pending completed training and the preregistered held-out, equal-compute, causal, and external-verification gates. CP420S12 is total checkpoint 700. The 700-920 completion envelope is approximately 76.1%-100.0%, with an 88.0% midpoint.

2026-07-29 - CP420S12: raw-policy failures are evidence, not infrastructure crashes

CP420S12 stopped after 24/36 frozen recurrent baseline cases and before any optimizer update. The resident model's wrong answer on recurrence-budget_plan-d2-s8082805353550149447 reached the live answer-replacement authority, which abstained; the research evaluator treated that bounded policy failure as an engine failure and aborted the campaign.

Research execution now disables serving-side answer replacement while leaving the live product default unchanged. Bounded abstentions and incomplete decodes are scored as incorrect policy observations; cancellation, latent-phase, worker, invariant, and cleanup failures remain fatal. The 320-token preregistered ceiling is now a hard ceiling rather than silently adding a second 320-token grace budget.

Pretraining is crash-resumable without reusing stale evidence. A step-zero checkpoint lands before the baseline, each baseline case and calibration probe is written to a canonical journal bound to protocol, dataset, model, execution spec, seed, order, and budget, and a second checkpoint seals the completed baseline. Exact identity drift refuses reuse.

The exact failed case was rerun on the resident 32B at four recurrent steps. It completed without a fatal error at the 320-token ceiling and was honestly scored incorrect/unparseable. This proves the evaluator repair, not a reasoning gain. Evidence: artifacts/current/cp420s12_research_policy_failure_evidence.json. CP420S12 is immutable and retired. No optimizer update, reasoning gain, frontier gain, promotion, or WOW Signal is claimed. CP420S12 remains total checkpoint 700; the repair checkpoint is 701 and a fresh campaign is required.

2026-07-29 - CP420S13 prelaunch: adequate training and unattended recovery

The next source-bound campaign cannot publish a recurrent adapter merely because it reaches the final scheduled step. Protocol v7 adds an independently recomputed training-adequacy certificate requiring the exact unique 288-task pass, at least 25% real optimizer updates, update activity in every preregistered evaluation window, a distinct policy digest for every accepted update, every scheduled held-out evaluation, and a healthy verified learning signal. The frozen identity verifier recomputes the certificate from step receipts and refuses an under-dosed bundle. Existing v5 and v6 bundles remain verifiable under their historical schemas.

The detached training target now owns a source-bound watchdog. It atomically journals each attempt, releases failed MLX state, and resumes from the exact durable checkpoint only when progress evidence permits it. Recovery is capped at four total attempts and stops after two consecutive failures with no durable progress. This closes the failure mode where a recoverable exception terminated after roughly 20 minutes and remained dormant until a human returned.

The focused trainer, identity, preregistration, post-training, and watchdog matrix passes 98/98. This is prelaunch reliability evidence only. CP420S13 still requires a fresh resident canary, contract, trust roots, launch bundle, and explicit user green light before the 32B process starts. No intelligence gain or WOW Signal is claimed.

2026-07-29 - CP420S13 locked and loaded at the green-light boundary

The fresh resident gates are complete. The source-bound 32B initial-policy probe attached all 24 recurrent projections. The answer-channel preflight returned five valid and correct results out of six and classified the remaining bounded token-limit result as policy evidence, not an engine crash; the frozen gate verdict is answer_channel_operational. The model-bound preregistration verification passes with the expected 32B checkpoint fingerprint and behavior bundle.

Four distinct Keychain-backed host signer services are live on their private Unix sockets with no exported private role keys. The signed 288-task launch bundle reopened deterministically three times at SHA-256 9cd86ac093b69db70955679c99d8169942958137a19600764fb3b39e2416f000. The exact one-command launch surface is frozen in artifacts/current/cp420s13_resident_32b_training_readiness_evidence.json. On launch it creates the detached source-bound supervisor, child-bound sleep inhibitor, step-zero checkpoint, and internal bounded restart watchdog. There is no resident 32B or training process running at this checkpoint.

The campaign is waiting only for Bryan's explicit green light. Training, optimizer updates, reasoning gain, frontier gain, promotion, and WOW Signal remain false until measured. Host-isolated signing is not independent external administration. CP420S13 readiness is total checkpoint 704. The 704-920 completion envelope is approximately 76.5%-100.0%, with an 88.3% midpoint.

2026-08-11 - CP245 resident decode localizes the neural readout failure

The frozen CP242 resident-32B diagnostic completed 24 decoded candidates under detached custody. Base, initialization-matched and trained neural arms each solved 1/3 fresh one-step tasks; trained recurrence produced no wrong-to-right transition at T1 or T4. The compiled T4 consumer of the same final typed state solved 3/3, while the grammar lesion solved 0/3 and pointer lesion 1/3. Correct typed values 12 and 17 were publicly emitted as 26 and 76. This isolates the current blocker to the learned role/place/digit answer bridge rather than the typed recurrent transition.

Report semantic SHA-256 is f1154e79e0a49ab890d1811946011fed1e5d49628023819e471213cc68b3c568; the preserved pretty-JSON file SHA-256 is 3bc8dd5f4f16560c62c710103d3d66287172a17c94f1f9010d6dd3fa1c9af37c; the passed detached receipt SHA-256 is 2ad887985c605699d7007e6b5a3ded87bda417981c91b0ed514a1282007df559. Historical status now verifies that semantic commitment without rewriting the noncanonical transport; future evaluators write canonical durable bytes, and resumable candidate custody rejects non-finite and symlinked evidence.

SPARK-069 through SPARK-072 remain open. No learned decoded gain, fusion, frontier result or WOW Signal is claimed. The completion envelope remains 912/920. The next repair is the bridge itself plus an emitted-answer admission gate on the cheap checkpoint before another resident run.

2026-08-11 - CP247 resident transfer decision rule frozen before decode

The source-bound CP246 resident canary completed and durably checkpointed its first 32B answer-bridge step, then resumed under the launchd controller for the remaining nine steps. Before any prompt-disjoint resident decode result existed, CP247 froze the bounded transfer rule: the matched T4 treatment must produce a positive wrong-to-right gain with zero regressions, beat its T1 path and base, and lose under both grammar and pointer lesions; the compiled instrument must remain exact and the untrained control must have headroom. Ceiling and broken instrument cases are inconclusive rather than silently promoted or called negative. Report commitments, summaries and paired counts are independently recomputed.

The new adjudicator and adjacent evaluator/launcher contracts pass 24/24; canonical smoke passes 104/104. SPARK-069 through SPARK-072 remain open, the completion envelope remains 912/920, and no transfer, fusion, frontier or WOW Signal claim is made before terminal admission and the frozen resident decode.

2026-08-11 - CP248 candidate-level resident verdict reconstruction

CP247 initially compared committed arm summaries without independently reconstructing them from their candidate rows. CP248 now requires one complete eight-arm matrix per task, rejects duplicate or missing cells, recomputes every correct count, and pairs the matched T4 candidates itself to derive wrong-to-right and right-to-wrong transitions. The affected surface passes 25/25. The resident training run is still active under its immutable CP246 source, so no decoded transfer or downstream claim is made.

2026-08-12 - CP275 through CP277: terminal recovery and typed serving are executable

The exact step-34 resident tissue survived three source repairs. CP275 replaced multi-depth graph retention with mathematically equivalent streamed gradient accumulation; steps 35 and 36 completed under the unchanged sentinel ceiling. The final training ladder has perfect typed state/action metrics through unseen T16 and a 77.684% held-out relative CE gain, while terminal decoded-answer admission remains the deciding gate.

CP276 and CP277 close the downstream implementation gap without pre-authorizing the result. A supported package may serve only its tokenizer-bound typed task families through separately signed, durable qualified authority. Pointer, activation, worker and client identities are independently checked, and one rollback-safe command owns activation and revocation. Ordinary chat, arbitrary reasoning, static fusion, broad capability and frontier claims remain false.

The powered replication is preregistered before terminal decode under plan ce1b5baa...: three disjoint seeds, 81 tasks, 648 eight-arm candidates, zero regression, exact compiled controls, strict grammar/pointer/base/T1 lesion requirements, at least 20% pooled gain and exact one-sided p <= 0.01. Its launchd controller is alive and waits on the authenticated CP275 terminal receipt. SPARK completion remains 913/920; these repairs do not count the remaining scientific gates as passed.

2026-08-15 - CP550 bounded semantic decode transfer is independently positive

The source-clean 1.5B semantic-neural canary converted all 27/27 ordinary decode failures across coding, calibration and premise auditing without a regression. Syntax-matched and same-family wrong-state controls remained 0/27; removing one learned multiplication interaction reduced treatment to 9/27. An independent verifier regenerated the cohort, regraded all 135 raw outputs and replayed every treatment state rather than trusting producer summaries. The frozen evidence is under artifacts/closeout/latent_cortex/cp550_semantic_decode_canary/.

This closes a bounded 1.5B state-to-language gate, not SPARK-069 through SPARK-072. A larger fresh-seed replication remains required before resident-32B transfer, and broad replicated gains remain required before fusion, frontier or WOW Signal claims.

2026-08-15 - CP551 replication localizes one state-serialization error

The larger fresh 1.5B matrix rejected treatment at 89/90 under the unchanged perfect bar; all 90 ordinary, syntax-matched and wrong-state controls failed. The one miss had an exact authenticated recurrent state but a corrupted JSON copy. State-grounded correction repaired that exact task on a second real-model decode without consulting its evaluator answer. Per-row receipt-chained journaling now preserves raw evidence before terminal adjudication. The full fresh matrix must be rerun before resident transfer can proceed.

2026-08-15 - CP552 state-prefix fallback closes deterministic copy ordering

The next fresh run stopped after its journal captured a treatment row that repeated a reversed-object serialization twice. A bounded prefix sweep over the arm's own authenticated state found that 80 tokens crossed the stable bad branch while leaving the model to generate the suffix. Exact-object-only token overrun normalization closes the adjacent BPE boundary case. The strict full replication gate remains pending on a new unseen seed.

2026-08-15 - CP553 fresh replication independently accepts bounded transfer

The source-clean 1.5B replication passed the unchanged five-arm bar on 90 fresh tasks: treatment 90/90, ordinary 0/90, syntax-matched 0/90, same-family wrong-state 0/90, and learned-coefficient lesion 30/90, with 90 wrong-to-right conversions and zero regressions. Its 450 decode commits form a complete 452-event fsynced receipt chain. Independent cohort regeneration, grading, state replay, source/model verification and journal verification all passed; the paired one-sided exact probability is 8.08e-28.

This authorizes the bounded resident-32B semantic decode canary. It does not close SPARK-069 through SPARK-072 or authorize ordinary serving, fusion, frontier performance or WOW Signal.

2026-08-15 - CP554 resident evidence now binds Aura's active manifest

Preflight found that Aura's logical 8-bit model name is overridden at live runtime by training/fused-model/active.json, which selects a fused 4-bit 32B checkpoint. The canary and independent verifier now cryptographically bind and reopen that manifest and refuse model-path or manifest drift. Two interrupted pre-manifest probes are diagnostic only. The fresh claim-grade resident run is next; no SPARK or WOW Signal gate closes here.

2026-08-15 - CP555 resident semantic decode transfer passes its bounded gate

Aura's manifest-selected fused resident 32B completed the fresh five-arm matrix with treatment 27/27, ordinary decode 9/27, syntax-matched wire control 3/27, wrong-state control 0/27, and multiplication-coefficient lesion 9/27. The treatment made 18 wrong-to-right conversions with zero regressions. Independent task regeneration, grading, neural-state replay, source/model and resident-manifest verification, and the complete 137-event journal chain all passed. The paired one-sided exact probability is 3.81e-6; evidence is frozen under artifacts/closeout/latent_cortex/cp555_resident_semantic_decode_canary/.

This is the first independently positive state-to-free-decode result on Aura's active resident cortex for the three exact executable families. It remains bounded: coding survives the multiplication-only lesion because that family uses learned addition, the semantic machine is not yet on the canonical governed serving path, and the run is not broad or open-domain. SPARK-069 through SPARK-072 remain open. The next gates are family-targeted lesions, runtime promotion with rollback and receipts, then fresh powered broad falsification before any WOW! or frontier claim.

2026-08-15 - CP556 qualified semantic tissue is active on the canonical path

The three CP555-qualified exact families now enter the healthy sovereign foreground path through strict canonical grammar admission. The runtime reopens a content-addressed activation before execution and refuses on source, evidence, manifest or model drift. The activation is on by default, has an explicit AURA_SEMANTIC_NEURAL_SERVING=0 rollback, and is bound to Aura's manifest-selected fused resident 32B identity under receipt be8567d8075e48381215421a61c36c8523adc0be117c9918afca2fb55d9392c1.

A fresh canonical-ingress verification solved 90/90 tasks and every family-targeted learned-coefficient lesion disrupted its target (90/90), while unsupported natural language was refused instead of captured by a broad pattern. Median execution latency was 47.844 ms, mean was 41.165 ms, and maximum was 73.510 ms. The runtime verification receipt is 19d9d2439269623924616f32d96edb48155a55f81ef7ca5c35747831ea91b83f. Future producer artifacts that declare targeted lesions are independently checked against the verifier's own lesion map; historical sealed artifacts remain verifiable without retroactive mutation.

This closes family-targeted lesions and rollback-safe canonical runtime promotion for the bounded exact families. It does not establish broad or open-domain reasoning, static weight fusion, frontier capability or WOW!. SPARK-069 through SPARK-072 remain open pending fresh broader-domain, equal-compute, lesion-backed powered replication.

FilesExpand file tree

RLC_SPARK_EXECUTION_LEDGER.md

Latest commit

History