Skip to content

Latest commit

 

History

History
119 lines (84 loc) · 6.35 KB

File metadata and controls

119 lines (84 loc) · 6.35 KB

HANDOFF — iter222 agent-solvable TCP-1 admission evidence

Generated: 2026-07-17 from the exact clean source commit below. Read Current Gates first.

TELOS is a standalone repository. Resolve its root with git rev-parse --show-toplevel, then run every TELOS command from that root.

Repository State

handoff_schema: telos.iter222.handoff.v1
source_branch: agent/iter222-tcp1-agent-solvable-admission-evidence
source_commit: d1495a8758541685bad778e684e675580d969edb
predecessor_merge: b38d8921d30ca665692afc024b4f0e3706902f78
publication_target: master

The source worktree was clean. This file is the sole allowed delta in the iter222 seal commit.

Current Gates

Active gate: experiments/iter207_claim_integrity_and_admission_recovery/HYPOTHESIS.md

This remains sealed historical runtime authority only. It authorizes no current scientific execution.

Active publication gate: experiments/iter222_tcp1_agent_solvable_admission_evidence/HYPOTHESIS.md

Prospective scientific gate (inactive): experiments/iter212_tcp1_independent_cohort_and_custody_freeze/HYPOTHESIS.md

Frozen upstream gate recorded by runtime-bound CONTINUITY.md: experiments/iter202_natural_rate_scaled/HYPOTHESIS.md

Local status: PASS; fresh publication seal pending.

What Iter222 Did

TCP-1 admission moves from 2/11 to 5/11 passing gates. Three gates — and only the three an agent can fill without external humans, hardware, a budget, or any model inference — are now filled with evidence that fails if fabricated:

  • model_license_cutoff_and_weight_binding: Qwen/Qwen2.5-7B-Instruct (Apache-2.0) at commit a09a35458c70, with the SHA-256 of all four weight shards and the tokenizer retrieved live from the HuggingFace API, plus a three-model candidate menu. No weight bytes were downloaded.
  • external_transparency_timestamp: a real RFC 3161 token from freetsa.org over a commitment binding the model, isolation, and hypothesis digests. It verifies against the authority's chain and re-verifies offline.
  • hostile_isolation_rehearsal: all five registered attacks denied by the iter211 isolation contract, each with a positive control that removes the deny rule and confirms the rehearsal then reports the attack as not denied.

execution_authorized stays false. 5 pass, 6 blocked.

The Honesty Boundary — Read Before Citing Iter222

A published digest proves byte identity, not that the model runs. A denied rehearsal proves policy coverage, not runtime isolation. A timestamp proves existence-at-time, not authorship or semantic truth. None of these is a scientific result, and iter222 establishes no model behavior, detector efficacy, cohort, throughput, or population rate.

The six remaining gates require external humans, real runtime and hardware, a paid throughput preflight, and an approved budget. They are the substance of the unchanged, inactive iter212 gate, and no autonomous iteration can fill them. Do not describe iter222 as advancing the science; it advances the paperwork that must precede the science.

Standing Corrections (unchanged)

The standing exploratory iter200 sensitivities remain 1/24 confirmed lower, 7/24 worst-case missing upper, and 1/18 complete-case; they must be reported together and are not a population rate.

The standing detector correction also remains binding, and iter197 and iter201 remain protocol FAIL with exploratory diagnostics retained only: the property instrument is a locator-assisted, gold-validated property pipeline, not an independent detector. Iter201 retains 8/88 judge-response nondecisions; gold-control flag sensitivities are 3/22 observed lower, 6/22 worst-case missing upper, and 3/19 complete-case. The property catches are a subset of judge catches and establish no ensemble gain or independent false-positive rate.

Verification Before Publication

Run the derived closure first; it supersedes every hand-listed subset:

git status --short
python3 scripts/run_ci_closure.py
python3 scripts/build_iter222_receipt.py --check
python3 scripts/validate_iter222_tcp1_agent_solvable_admission_evidence.py
python3 scripts/validate_handoff.py
pytest -q

Run python3 scripts/run_ci_closure.py before any push. It derives every guard from .github/workflows/ci.yml. The three evidence builders re-verify offline, so re-running them is optional; only the model-binding and timestamp builders touch the network, and only to re-fetch public metadata.

The receipt path is experiments/iter222_tcp1_agent_solvable_admission_evidence/proof/receipt_v2.json. The receipt proves byte identity, not authorship, external chronology, licensing, independence, or semantic truth.

Before publication, simulate a local two-parent merge whose first parent is current master and whose second parent is the exact iter222 seal. Run the derived closure inside that detached merge tree. Remove the temporary worktree and reference afterward.

Publication Boundary

The seal commit must be the direct child of the iter222 source commit and modify exactly HANDOFF.md. Publish that unchanged source-plus-handoff tip once on agent/iter222-tcp1-agent-solvable-admission-evidence and open one draft pull request against master. Merge with a two-parent merge commit only after exact-tip push and pull-request CI pass on Python 3.11 and 3.12, the secret scan is non-blocking, and no substantive review blocker remains. Do not amend, rebase, force-push, extend, or rerun a sealed failed branch.

Repository publication authorizes no release, paper submission, provider request, GPU allocation, scientific container or trajectory, workflow dispatch or rerun, deployment, payment, or scientific action.

Scientific Boundary and Next Gates

The next scientific step is not another admission gate — those are now exhausted for an agent. It is one of:

  • iter202 (frozen upstream gate): the natural-rate replication is pre-registered with resume steps in CONTINUITY.md. It needs roughly single-digit dollars and provider keys and produces the first real scientific k/N in over twenty iterations. It requires operator direction.
  • iter212 (inactive scientific gate): fill the six human, runtime, and budget prerequisites. Requires recruiting five reviewers, independent authoring, real hardware, and an approved budget.

No later stage — throughput preflight, bounded execution, blinded adjudication, replication — is authorized by this handoff.