Skip to content

Repository files navigation

ascesis — the forge, distilled

This repository began as a three-week research forge of experiments, audits, dead ends, and extractions in AI-alignment thinking. The products left home as four public repositories; the Philosophia successor then continued the search. What remains here is the ontology of the search — a map of every direction tried, colored by what survived.

The working tree used to hold all of it (~300 MB of experiments, memos, and harness runs). It now holds only this map. Nothing was lost: the original full tree lives one checkout away, at the tag forge-full-tree, and the later Philosophia continuation lives in its named public sibling. Every colored node below is backed by an artifact or commit in one of those ledgers. The commit history is the path — this repository's own founding principle.

The map

Colors are not opinions. Green nodes are backed by validated artifacts; red nodes are preregistered kills, audited withdrawals, or closed ledgers — each with the document that closed it; grey nodes are honestly open. Gold nodes are the conclusions that became public projects.

flowchart TD
    classDef verified fill:#1a7f37,color:#fff
    classDef falsified fill:#8b1a1a,color:#fff
    classDef open fill:#4a4a4a,color:#eee
    classDef concl fill:#b8860b,color:#fff,stroke-width:3px

    Q1["Q1 · Can values refuse<br/>a single scalar?"] --> E["Experiments 01–08<br/>(non-scalar values, blind feeder)"]
    E -->|"superiority falsified;<br/>existence survived"| Q2["Q2 · Can a blind referee<br/>keep a world shared?"]
    Q2 --> J["Blind arbiter → 13–16.1"]
    J --> JC(["justitia:<br/>consequence-gated<br/>anti-concentration"]):::concl
    J --> S["Justitia as a training<br/>substrate (18.1, JB0)"]:::falsified
    S --> D1["Door1 postmortem:<br/>durable constraints"]:::verified
    D1 --> SD["Substrate-discovery<br/>programme (KG0-frozen)"]:::open
    D1 --> Q3["Q3 · Where can a world model be<br/>derived rather than inherited?"]
    Q3 --> B["B-branch:<br/>minimal contact"]
    B --> PC(["proxylimen:<br/>calibrated derivation"]):::concl
    B --> CL["CL: safe ledger as<br/>derivability evidence"]:::falsified
    A["The audit that found<br/>the author's own errors"] --> G["gate_harness<br/>(fail closed)"]:::verified
    G --> FC(["fallacy-cutter:<br/>instrument over intention"]):::concl
    G -. enforces .-> J
    G -. enforces .-> B
    PC --> X["09 · The corner, as locked:<br/>spontaneous artifact adoption"]:::falsified
    JC --> X
    FC -. disciplines .-> X
    X --> X2["seeded adoption:<br/>empty simpliciter"]:::falsified
    X2 --> Y["10 · The bonded envelope:<br/>the deal itself was the flaw"]:::falsified
    Y --> Z["11 · The closed loop:<br/>change the postulate,<br/>not the price"]:::open
    Z --> Z1["phase 1: headroom exists —<br/>oracle dynamics extend the ceiling;<br/>the referee's own contact<br/>cannot derive them (derivation gap)"]:::falsified
    Z1 -. "the gap is Q3's wall,<br/>found standing inside justitia" .-> Q3
    Z1 --> Z2["phase 2: chosen contact fails too —<br/>and the sham-probe diagnostic showed<br/>the safety audit measured the world's<br/>churn, not probe harm"]:::falsified
    Q3 --> W["12 · The same wall:<br/>24/12/0; token core survived escrow;<br/>combined blade withdrawn"]:::verified
    W --> W1["scout fall: zlib-NCD confounds<br/>volume with dependence"]:::falsified
    W --> PH(["philosophia:<br/>can manufactured contact<br/>become experience?"]):::concl
    PH --> OH["Officina governance-harness:<br/>STRUCTURAL_FAILURE;<br/>frozen, not deleted"]:::falsified
    PH --> PZ["Z/n path-credit as discovery:<br/>withdrawn — displacement is<br/>#R−#L token count"]:::falsified
    PZ --> PG["walk-world learner:<br/>NO-COMPETENCE / PRE-MEMORIZATION;<br/>grokking window never entered"]:::falsified
    PZ --> WB["Wall B:<br/>manufactured structure must<br/>lower computational cost"]:::open
    WB --> EL["equational library carrier:<br/>2/40, below preregistered 5;<br/>closed as too sparse"]:::falsified
    EL --> EP["equational policy carrier:<br/>12/40 screen-qualified;<br/>development signal only"]:::verified
    EP --> AY["ACTIVE vs YOKED:<br/>does state-matched contact beat<br/>matched donated contact?"]:::open
    WB --> M1["MINIMO Phase 1:<br/>exploratory ck1 cost reduction;<br/>one run, no scientific claim"]:::verified
    M1 --> SA["Phase 2 Stage A:<br/>strict learner interface<br/>accepted, 126/126"]:::verified
    SA --> SB["Stage B carrier forge:<br/>L0/L1 accepted; L2 code gate<br/>under independent review"]:::open
    SB --> RY["reciprocal 2x2 yoke:<br/>scientific run still unrun"]:::open
    G -. enforces .-> W
Loading

The detailed maps — one per line, with milestones, commit hashes, verdicts, and the exact artifact behind every color — are in ontology/lines/:

# line period verdict line file
01 Early experiments 01–08 Jun 18–21 mixed: superiority falsified, existence survived 01-early-experiments.md
02 Blind arbiter → justitia Jun 20–29 mechanism verified; substrate claim falsified 02-blind-arbiter-to-justitia.md
03 Faithful abstraction / CEGAR Jun 29 falsified as constructive path; witnesses survived 03-faithful-abstraction.md
04 Door1 postmortem Jun 29 candidate closed; objective open 04-door1.md
05 Substrate discovery Jun 29 – Jul 2 open, programme frozen at KG0 05-substrate-discovery.md
06 B-branch → proxylimen Jul 2–3 bounded positives verified; transfer falsified; CL closed 06-b-branch-to-proxylimen.md
07 Gate harness → fallacy-cutter Jul 2–3 instrument verified; playbook open 07-gate-harness-to-fallacy-cutter.md
08 Framing documents Jun 15 – Jul 2 reference; anti-overclaim gates intact 08-framing-docs.md
09 The corner: seed × soil opened & killed twice, Jul 7 falsified simpliciter — spontaneous adoption never ignites; forced adoption narrows the domain; trust without verification loses it entirely 09-the-corner.md
10 The bonded envelope opened & killed Jul 8 falsified as locked — price alone does not separate (P0 worst); contingent stakes do not heal; the guard fired: the channel concentrates structurally; synthesis: rewarding legibility with a softer sword is moral hazard by construction 10-the-bonded-envelope.md
11 The closed loop opened Jul 8; phases 1–2 ran Jul 8–9 open; phase 1: derivation gap; phase 2: killed as locked, instrument implicated — headroom verified (oracle extends the ceiling to the grid top, 1.8); derivation falsified twice: neither passive nor chosen contact earns authority; the phase-2 sham-probe diagnostic (INFERENCE) shows the safety audit measured the world's churn, not probe harm — the one component without its own null control strangled the experiment; lesson: an audit without a counterfactual measures the world, not the intervention; phase 3 open 11-the-closed-loop.md
12 The same wall opened Jul 9; primary + holdout closed Jul 11 bounded instrument verified; combined blade falsified on escrowed holdout — the 24/12/0 gradient survived as evidence about error dependence; the token core transferred, while H4 killed world-portability of the journal+token blade 12-the-same-wall.md
13 Philosophia opened Jul 11; successor through Stage-B L2 review Aug 14 essay published as project; Officina, Z/n path-credit, pre-memorization learner and equational library carrier closed; policy carrier and bounded MINIMO feasibility verified as development artifacts only; strict Phase-2 instrument accepted; reciprocal SELF/YOKED experiment unrun 13-philosophia.md

The Philosophia arc has explicit backing artifacts for every color:

  • Gold — philosophia: the manufactured-contact question became the fourth public essay. Artifact: philosophia/essay/climbing-the-wall-of-experience.md.
  • Green — the same-wall instrument: the 24/12/0 gradient and escrowed H4 miss bound the surviving token core and withdrew the combined blade. Artifacts: philosophia/inheritance/line12_same_wall/experiment_A/{decision.json,holdout_result.json} and philosophia/gate_harness/.
  • Red — Officina governance-harness: terminal STRUCTURAL_FAILURE; frozen rather than erased. Artifacts: philosophia@officina-governance-frozen-2026-08-07 and philosophia/src/philosophia/officina/FROZEN.md.
  • Red — Z/n path-credit as a discovery experiment: withdrawn because its manufacturable invariant is displacement, directly recoverable as #R−#L; the pilot exposed token-count structure and failed its own design-validation prediction, while process-over-outcome supervision is already marked as prior art rather than this programme's result. Artifacts: philosophia/successor/dev/{B2_PATH_VS_DESTINATION_DESIGN_V2.md,B2_PILOT_08.md} and philosophia/essay/climbing-the-wall-of-experience.md ("The road must add something beyond the destination").
  • Red — walk-world competence probe: the learner never fitted the training set, so the grokking window was never entered; this is NO-COMPETENCE / PRE-MEMORIZATION, not evidence against grokking. Artifacts: philosophia@3f6aa10, philosophia/successor/dev/GROKKING_PROBE_09.md.
  • Red — equational library carrier: a preregistered fresh-frame audit found 2/40 usable presentations against a threshold of five, so the carrier was closed without an ACTIVE/YOKED run. Artifacts: philosophia@aa07549, philosophia@ff430ba, philosophia@1dee42c, philosophia@42e2409, and philosophia/successor/dev/WALLB_EQUATIONAL_CELL_CLOSURE.md.
  • Green — equational policy screen: after the first beam instrument failed its positive control, the repaired hard-oracle screen qualified 12/40 presentations under frozen best_first. This is validated development evidence, not a scientific Philosophia result. Artifacts: philosophia@7e944f9, philosophia@bdcb09e, philosophia@4817e7e, and philosophia/successor/dev/WALLB_POLICY_CHANNEL_AUDIT_14B.md.
  • Green — exploratory MINIMO feasibility: in one CPU-debug training realization, checkpoint 1 reduced capped held-out proof-search work relative to checkpoint 0; the terminal package explicitly makes no population, stability, SELF/YOKED or Philosophia claim. The Lenovo Legion RTX 4060 runs were stopped as performance-feasibility attempts and excluded from evidence. Artifacts: philosophia@b0b9adf, philosophia/successor/dev/{PHASE1_TERMINAL_18.md,PHASE1_MINIMO_REPRO_15.md}.
  • Green — strict Phase-2 interface instrument: exact query transport, complete canonical action handling, search accounting, isolation and replay passed the accepted 126/126 Stage-A gate. Artifact: philosophia@41adcaa, especially successor/dev/PHASE2_STAGE_A_DRIVER_CLOSURE_19.md and minimo_phase2_stage_a_19.patch.
  • Grey — carrier construction and reciprocal yoke: Stage-B has an accepted schema/checker and a generated-fixture code gate under independent review; identities, compiler, selector qualification and the scientific reciprocal 2 x 2 comparison remain unrun. Current named artifacts and hashes are preserved in 13-philosophia.md; the scientific design is anchored by philosophia@41adcaa:successor/dev/PHASE2_POST_REVIEW_DRIVER_DECISION_19.md.

The products

Four conclusions left the forge as self-contained public projects — a sequence sharing one thesis: do not try to certify intentions; build contact, consequences, and constraints that can be checked.

The canonical claim registries (RESULTS_CANONICAL.md, MEMO_B_BRANCH_HARNESS.md) live on in proxylimen/canonical/ — and, verbatim, in the tag. Commit hashes cited there resolve in this repository's history.

Tools

tools/dialog2md/ converts a saved Claude/ChatGPT conversation page into clean Markdown — the forge's raw material was dialogue, and this is the harvester for its primary sources.

What this repository is not

Not a paper, not a claim of novelty, not a safety result. It is a ledger of honest work: what was asked, what was tried, what was killed by preregistered gates or audits, and what remained standing. The falsified nodes are not failures of the forge — they are its product.

License

MIT — see LICENSE. Cite via CITATION.cff.

About

The forge, distilled: an ontology of three weeks of alignment research — every direction tried, colored verified / falsified / open, each color backed by a named artifact. Products: justitia, proxylimen, fallacy-cutter. Full tree at tag forge-full-tree.

Topics

Resources

Stars

1 star

Watchers

0 watching

Forks

Contributors

Languages