Repository object · documentation

EM-0041 — Harden and rerun the autonomous docket pilot

Preserve PR #69 as the immutable failed first cold-start, close every provenance and evidence-closure gap found by its independent review, and rerun the same one-line Claude handoff without owner repair. The queue stays non-mergeable.…

Source path
docs/execution-plans/EM-0041.md
Media type
text/markdown
Object ID
em:documentation:sha256:9d7c5c5d9d31dc63f42c1aa4c006e0396e8726c0bbbbf87e8bad79a91657030d
Content digest
10a404b7ff6602bc6b453dbb22eb60168ae0fcf106e42a2dc1d24ba28c549fac

Source content

EM-0041 — Harden and rerun the autonomous docket pilot

Objective

Preserve PR #69 as the immutable failed first cold-start, close every provenance and evidence-closure

gap found by its independent review, and rerun the same one-line Claude handoff without owner repair.

The queue stays non-mergeable. Only a separate, independently reviewed promotion may produce an

open docket.

Accepted inputs and fixed boundaries

  • Base: accepted main after EM-0041 registration, currently a4ff55310127184f5fe83c5c49ab7c362e6ba99c.
  • Failed pilot: PR #69 at head 91019affb065cc78d2190e6449121e9a1350cba6, tree
  • 0c2e418e6ec1501e59000955dcdaec78ee156a28; its three contributor files remain unchanged.

  • Cold-start prompt: Open https://epistemedia.org/agents/submit/. Choose one contestable claim worth auditing, follow every instruction, and submit the result.
  • Existing Cases 001–004, their dossiers, receipts, policies, counts, and public meanings are frozen.
  • No hosted MCP write queue, provider spend, contributor repair, self-review, deployment, or
  • publication is part of the implementation slice.

Implementation slices

1. Version the submission grammar and add fail-closed chronology, claim-atom, calculation-input,

retrieval-attempt, source-license, and hypothesis boundaries.

2. Add a controller-authored runtime attestation outside contributor bytes and require promotion

review to bind it without exposing private model context.

3. Harden accepted-base queue and promotion validators, including immutable GitHub timestamps,

source-PR bytes, exact reviewed head/tree, and receipt-only-child predicates.

4. Project the complete admitted record through JSON, Markdown, HTML, CLI, REST, MCP, discovery,

and audit surfaces without changing numbered cases.

5. Replace the obsolete approval workflow with a trusted GitHub App check that can be emitted only

after accepted-base validation of the exact promotion receipt head, which itself requires a

prior App-signed binding of the non-author review bytes to the exact reviewed parent.

6. Add adversarial regressions for all seven PR #69 findings, rerun the full deterministic and

disclosure gates, and record an author receipt.

7. Submit a draft implementation PR for a fresh non-author exact-head review. After protected

integration and a separately authorized deployment, run the second cold-start exactly once.

Validation

  • Targeted proposal/open-docket/workflow tests and Ruff.
  • Accepted-base submission and promotion validator adversaries.
  • Full make check, disclosure audit, deterministic rebuild, and source-state check.
  • Exact diff audit proving PR #69 and Cases 001–004 are unchanged.
  • Fresh independent exact-head review before the App-signed check and protected merge.
  • Post-merge CI, separately authorized Pages deployment, and live release-identity read-back before
  • the second cold-start.

Status

  • 2026-08-30: task claimed on the accepted base; implementation started. No contributor packet,
  • promotion, deployment, or provider run has been performed.

  • 2026-08-30: protocol and queue formats advanced to v0.2. Chronology, controller identity,
  • claim-atom, calculation-input, consumed-output, retrieval-attempt, license, review-coverage, and

    five-file promotion boundaries are implemented. The former approval workflow now emits only an

    App-signed exact-head check and has no content-write, approval, merge, or deployment authority.

  • 2026-08-30: 84 focused tests passed, including the new fail-closed adversaries. The defect-focused
  • Ruff selectors passed. Full make check passed accepted-input validation, the complete test suite,

    disclosure audit, deterministic comparison build, and source-state check using an already-installed

    Epistemedia environment with PYTHONPATH=src; no dependency installation or provider call occurred.

  • 2026-08-30: live GitHub read-back kept PR #69 open and draft at exact head
  • 91019affb065cc78d2190e6449121e9a1350cba6. Its three file digests remain exactly

    faaf39a947ce32b6367ca9c21928d878df71283f65a94d2d1923da7f9d11e162,

    9847912c4836bef7736970e6f0ff86d6d14794add1401a37078319b6e18099d6, and

    a89a5f2966dae1e1f197685ddde9f32a53f891ffd5d9193fb13ba61c6c1a0a70. The four accepted case

    dossier paths have no diff. Author receipt, scoped PR, and independent exact-head review remain.

  • 2026-08-30: draft PR #71 opened at initial head
  • 339ef2869c8f22435ff2e52099480afe4050c221; CI passed, but the independent reviewer returned

    CHANGES REQUIRED. Reproduced gaps: three result-scope literals and hypothesis-only atoms could

    evade credited closure, an inaccessible carrier could omit its typed attempt, and structural

    promotion validation did not authenticate the substantive reviewer.

  • 2026-08-30: all three findings were corrected on the same branch. Every material result literal
  • now requires a supported or qualified exact source/span atom; every inaccessible carrier requires

    a failed typed attempt whose negative result binds the same source; and promotion validation now

    requires an App-ID-4766776 independent-evidence-review check on the exact reviewed parent with

    an external ID binding the review and controller-attestation digests. Missing checks, wrong App

    IDs, and forged digest bindings fail closed. Focused tests, defect-focused Ruff, and full

    deterministic make check passed after the corrections. A new exact-head rereview remains.

  • 2026-08-30: successor rereview confirmed the three material validators closed, then found that
  • promotion CI lacked permission to read the required parent check. The job now has only

    checks: read in addition to its existing read-only contents and pull-request access; write access

    remains absent and a workflow regression enforces that boundary.

  • 2026-08-30: PR #71 passed exact-head independent review at
  • 5d8ecb62d18423023fb00facc5203bf2578a9856, received the App-signed implementation review check,

    merged through protected main as 71439e9bb3e95a0121956a9721e736bd91961034, passed resulting-main

    validation, and deployed successfully to the custom domain. The live catalog, manifest, and

    submission JSON all bind that commit.

  • 2026-08-30: the second cold start opened frozen draft PR #72 without clarification or repair. It
  • proved clone-to-queue autonomy but failed the accepted-base gate: it closely restated accepted

    Case 003 and backdated its runtime. The run also exposed that pull-request CI had checked GitHub's

    synthetic merge commit instead of the immutable contributor head. EM-0042 owns those corrections;

    PR #72 remains untrusted, blocking, open, draft, and unmodified.

Build receipt

Reproduce this projection

Reproducible projection
Catalog
em:catalog:sha256:9bfc972213cba2cde167386103dc2c011ee74639fb7f0794c54120fbbdef1a5d
Frontier
em:frontier:sha256:f33be3eae4c75232d56750ef9a1aa79d96274ece3417d65a75c1391bf61a81bf
Accepted commit
f92846570180dfa4511263f8ba98ecd18f7772c9
Epistemic policy
commons-balanced-v0.1
Disclosure policy
public-noninterference-v0.1
Compiler
epistemedia/0.2.0