EM-0034 execution plan

Source content

EM-0034 execution plan

Status: author candidate complete; fresh-clone independent dossier review pending.

Objective

Project the exact independently reviewed EM-0032 GPT-4 bar-exam packet into one

disclosure-safe, reversible Case 003 candidate dossier without admitting, featuring,

publishing, or deploying it.

Authority and identity

• immutable task: tasks/contracts/EM-0034.json;
• accepted starting base: ead2b5887da233dde4e790f0c6c0aea15e59f074;
• branch: codex/em-0034-case003-dossier;
• accepted packet ID:

em:research-packet:sha256:535d07e59563b12f66e590c31b0d53a21db1a8dfce1487129a54c5e86b9fd55b;

• accepted packet SHA-256:

07dec44bc2cf893b19ba6307e34aef60b6127b5e881e319ff4132c8e69972d1a;

• accepted independent-review SHA-256:

0606e98ed7ff98f1d0e69509db5edcd8dce774c62c48981bd5aec8f23a95ad20;

• accepted source-register SHA-256:

385bbc193be495597ef0fe2a675220ca443297e673b32947c8d8ffcd1a6366da.

Every accepted EM-0032 research, plan, note, and run-receipt byte is pinned by

the candidate verifier.

The builder reads only the accepted packet and its reviewed supporting records; it performs

no retrieval.

Candidate construction

1. Preserve 8 research source-work identities, 19 source-edition identities, and all 35

reviewed parent spans byte-for-byte.

2. Retain 298 and approximately 90th percentile as the launch-edition historical label.

3. Keep the undisclosed launch comparison distribution unresolved.

4. Preserve 297 and 298 as scoring variants within one experiment, never separate roots.

5. Reproduce the three Illinois sensitivity interpolations and seven Martínez calculations,

adding one repository audit work, one structured edition, and 10 exact calculation spans

that retain equations, inputs, input cells and values, comparison populations, uncertainty,

dependencies, and results.

6. Preserve the Martínez 45/48 version-of-record discrepancy without choosing one silently.

7. Keep one historical model-performance root and seven independent roots across five empirical lineage groups.

8. Project all 10 accepted typed dependence-edge groups into exact evidence-linked relations,

expanding every multi-source endpoint instead of retaining secondary endpoints only in prose.

9. Withhold current-model, practicing-lawyer, and general legal-competence inferences.

10. Apply encyclopedia and skeptical policies to one unchanged source graph with materially different labels and reasons.

Deterministic dossier

• dossier ID:

em:dossier:sha256:babe89ba3bda594a8d9f2db86a5a2987f284437a069b940d19b6928856d936d1;

• candidate file SHA-256:

32c4457b3823237b2f988a26d51b2f6222af8060e662993524aff1c1a5d79e5d;

• source-graph SHA-256:

2bb896ded74d63469ec4e3947d75adfd6ac9a9c5dee6fe0db7ef9251760da38e;

• 8 accepted research works, 19 accepted source editions, and 35 exact parent spans;
• one calculation-register work and edition with 10 exact derivation-and-input-cell spans;
• 9 total works, 20 total editions, 45 total spans, 21 propositions, 7 lineages,

21 assertions, and 33 evidence relations;

• one claim family and two materially distinct policy evaluations;
• 8 accepted claims, 10 accepted calculations, five empirical lineage groups,

seven independent roots, and 10 accepted typed dependence edges.

Validation gates

• deterministic rebuild and exact dossier content-address recomputation;
• byte-for-byte preservation of every accepted EM-0032 artifact and receipt;
• strict dossier-format, reference, span, public-disclosure, and policy-frontier validation;
• exact source/work/edition/span/claim/calculation/lineage/relation identity sets and counts;
• complete material-sentence assertion, parent-span, calculation, and typed-edge closure;
• explicit 297/298, 45/48, undisclosed launch-population, and lawyer-comparator boundaries;
• materially distinct encyclopedia and skeptical evaluations over one source graph;
• fail-closed independent receipt binding exact base/head/tree, reviewer non-authorship,

timestamps, command-output digests, clean pre/post state, complete coverage, limitations,

findings, and decision;

• adversarial candidate and receipt tests, Ruff, git diff --check, full make check,

deterministic rebuild, and clean tracked state.

Author validation record

• build_candidate.py --check reproduced the committed candidate and content address;
• verify_candidate.py --self-test verified exact closure and rejected ten forged or

incomplete review receipts;

• nine adversarial candidate and receipt tests passed;
• verify_candidate.py --require-review fails closed until the independent EM-0034

receipt is appended;

• Ruff, git diff --check, and full make check passed with 93 tests, clean public

disclosure audit, deterministic rebuild, and unchanged accepted source state;

• append-only claim receipt:

runs/proposals/20260828T003959Z-536ec913dfdf.json;

• append-only author validation receipt:

runs/proposals/20260828T024437Z-d153517a673c.json;

• append-only calculation/endpoint-closure validation receipt:

runs/proposals/20260828T155323Z-ee99db0ad3c3.json;

• exact author head, tree, draft PR, and CI identities are recorded after push.

Boundaries

The output remains a research candidate. No independent-review authorship, catalog

admission, feature selection, production-availability claim, publication, deployment,

DNS, provider, credential, package, release, or spend action is authorized by this plan.