EM-0032 execution plan

Source content

EM-0032 execution plan

Status: corrected author research packet complete; fresh-clone independent re-review pending.

Objective

Turn the accepted GPT-4 bar-exam readiness map into a content-addressed candidate research packet

that separates the historical composite score from every percentile comparison class, reconstructs

disclosed calculations, and leaves the missing launch derivation unresolved.

Authority and identity

• immutable task: tasks/contracts/EM-0032.json;
• accepted base: 7dca84b0c8477efc09b31215f734348d626957d4;
• branch: codex/em-0032-gpt4-bar-lineage;
• accepted preliminary packet:

research/how-we-know/library-preliminary/gpt-4-bar-exam-percentile.md;

• allowed paths: the EM-0032 research directory, this plan, the Case 003 research note, and

append-only run receipts.

Method

1. Freeze the narrow historical target and a 2026-08-27 evidence cutoff.

2. Capture credential-free public primary/authoritative carriers with exact URL, edition, bytes,

SHA-256, license treatment, and quote-minimal spans.

3. Verify the 15 preliminary core sources and 89-file mechanical artifact inventory.

4. Collapse report, preprint, VOR, repository, supplement, and OSF manifestations by underlying

experiment or analysis root.

5. Separate observed/reported score, pass threshold, first-time/repeater status, passers,

administration, and modeled percentile.

6. Recompute the disclosed Illinois sensitivities and Martínez assumptions deterministically.

7. Preserve 298/~297, Martínez 45/48, the OSF 0.27/24% comment conflict, and the

266/270 cutoff difference rather than silently choosing one value.

8. Keep the launch chart, interpolation, national total-UBE distribution, purchased inputs,

proprietary snapshots, and official essay grading visible as unresolved limitations.

9. Require a fresh-clone reviewer to re-retrieve sources, match every span, reproduce the math and

lineage, and bind a GO/HOLD/FAIL receipt to the exact author head.

10. Recompute five mutable HTML carriers through one exact-root, exclusion-aware normalizer and

bind all 78 pinned Git bodies to a deterministic negative-search manifest.

Deterministic packet

• packet ID:

em:research-packet:sha256:535d07e59563b12f66e590c31b0d53a21db1a8dfce1487129a54c5e86b9fd55b;

• artifact inventory ID:

em:artifact-inventory:sha256:17f52a5509fded7e75b08201f61122d09c7626443857c94f8727c85c0824e61c;

• Git body-search manifest ID:

em:git-blob-search:sha256:545908f30ba849c42c860185f92612f4d52a53b4f13b3c3ca5672213e23ba996;

• 19 source records, including 15 preliminary core objects;
• 35 quote-minimal parent spans and 76 typed span units;
• 89 file metadata records with zero independent-evidence credit;
• 8 scoped claim records, 10 derivations, 5 typed lineage roots, 10 evidence-bound typed lineage

edges, and 1 bounded negative search.

The author recommendation is GO pending independent review for later dossier construction. It

does not promote the packet or establish a public verdict.

Validation gates

• deterministic packet rebuild;
• exact capture identity and artifact-root counts;
• exact-root raw-to-semantic recomputation for all five mutable HTML carriers;
• pinned commit/tree readback of all 78 Git bodies, with 72 UTF-8 searches and 6 explicit binary

no-text-search records;

• independent analytic recomputation of every displayed derivation;
• sentence-to-span and source-to-lineage closure;
• explicit no-credit handling for the unresolved launch derivation and artifact mirrors;
• fail-closed independent-review gate;
• Ruff, JSON parsing, path-scope comparison, git diff --check, full make check, disclosure

audit, deterministic public rebuild, and clean tracked state.

Author validation record

• build_packet.py --check reproduced the committed packet and its content address;
• verify_packet.py independently recomputed all ten calculations and verified 19 source

captures, 15 core objects, 35 parent spans, 76 typed span units, 5 roots, 10 dependence edges,

and all 89 artifact records;

• the correction replaced the misidentified testing-column source, separated the Katz Git commit

from its tree, preserved the visible Rosemary Reshetar byline and conflicting Jim Leach JSON-LD,

bound all 21 July MBE distribution cells, and converted composite table/code excerpts into typed

exact units;

• a stable Texas A&M carrier binds all seven Martínez extents and retains the current direct-fetch

HTTP 403 limitation; no publisher challenge page is represented as a PDF;

• the HTML verifier reproduced five semantic captures and all their extents from recorded raw

bytes; the Git verifier reproduced all 78 body identities and bounded-search results;

• the receipt verifier now recomputes packet records, commands, timestamps, clean Git state, exact

base/head/tree identity, reviewer independence, and receipt-only-child binding, with adversarial

mutations required to fail;

• verify_packet.py --require-review failed closed with

independent review receipt missing, as required before review;

• Ruff passed for both research Python files;
• make check passed accepted-input validation, the complete test suite, disclosure audit with no

findings, deterministic independent rebuild, and unchanged accepted source state; and

• predecessor author receipt: runs/proposals/20260827T155153Z-29d70e6b9e44.json;
• corrected content commit: f49a07f53022a9366a5be81f3cf535242038aba9, tree

2ced50ea3f8c7bd49550bcf6195249ffc8b12cb3; and

• superseding author correction receipt:

runs/proposals/20260827T193633Z-20e2fc3af0dd.json;

• review-gap correction commit: 21504dd62dae69ab73b0a8b029d1ca7f543f36bf, tree

8a419ea269f7a53664f4d682c5c6ef33530a9c85; and

• review-gap author receipt:

runs/proposals/20260827T211411Z-db2c00b6869a.json;

• exact-locator correction commit: bc3150228f562dff7639eed170b97e2fb8322b02, tree

9f71d64a793ca05962dd4e594821300f997a5ac3; and

• locator-correction author receipt:

runs/proposals/20260827T213645Z-2080bbbc9450.json.

An independent exact-head receipt and protected merge evidence remain pending.

Boundaries

No dossier construction, catalog admission, public verdict, Pages deployment, current-model

ranking, provider call, proprietary rerun, purchased input, credential, spend, DNS, workflow,

package, or release action is authorized.