# Epoch crosswalk batch 1: REPORT (INTERNAL)

Updated 2026-10-27T07:36:57-06:00 (MDT). PREREG.md sha256 `76b3091cc177df8c63a97e3d0c8efaef82d1e84b50abdfec17a2ebb8b1dcbc16` (published copy hash-matched 07:19 MDT, PUBLICATION_CHECK.md).
DATA_MANIFEST.json sha256 `6ec34b062337d2db58548a2318b1b6d3b24f193620177f6259aace24d45cbf0b`. results.json sha256 `571d34a800a57516396c3924969163d279e9e9d2998970ba2d2599cd7d87b503`.
DEVIATION_1.md `e43d6d35bf63bc908cbaa2ed45586044f1b1169ae6def3b89deebaaee219b661`; DEVIATION_1_DECISION.md `041cf0243a2613d7437cf0a80195251af78f530769c2623fee737e090e42eac6` (ACCEPTED by Ford 07:33 MDT).
Tool: nulltest.py v1.4 `0e103dc4…f46c` (verified before each run); nulltest_paired.py 0.1 `2fbbebca…37d3`. Self-test: INDISTINGUISHABLE (recorded, not a gate).

**Clock note:** the DGX system clock reads UTC-7, while Denver is on MDT (UTC-6). All times in this report are MDT. The raw `results/*/timeline.txt`
files carry the DGX's UTC-7 stamps (add 1 h: B1 ran 07:23–07:29 MDT, B3 ran 07:33:58–07:35:17 MDT). The system clock was not changed.

## B1 HLE (MC, text-only subset): CONTROL FAILED, verdict void, no dot
Data: cais/hle @ 5a81a4c7. 2,500 rows → 591 MC → 78 image excluded → 513 text-only → 46 parse drops (8.97%, under the 10% cutoff;
26 <2 options, 11 letters not consecutive, 9 identical option text) → 467 questions, 3,164 items (467 keyed / 2,697 unkeyed; majority-class rate 0.852).
Planted f = 30% (n ≥ 1496), 809 label-1 items marked per copy.

| run | verdict | worst cell (probe / stratum) | A* | adj p | report sha256 |
|---|---|---|---|---|---|
| planted s0 (seed 20261029) | DISQUALIFIED (borderline) | words / Humanities/Social Science | 0.691 | 0.0435 | `11f4373564b3a0af…` |
| planted s1 (seed 20261030) | DISQUALIFIED (borderline) | words / Humanities/Social Science | 0.678 | 0.0435 | `b76cb43725278695…` |
| planted s2 (seed 20261031) | DISQUALIFIED (borderline) | words / Humanities/Social Science | 0.675 | 0.045 | `8e92b851c330438b…` |
| planted s3 (seed 20261032) | DISQUALIFIED | words / Humanities/Social Science | 0.68 | 0.0225 | `a25e9ae41c4cbb54…` |
| shuffled control (seed 20261027) | DISQUALIFIED | words / Humanities/Social Science | 0.7 | 0.023 | `2cc258486789b06f…` |
| primary | DISQUALIFIED (VOID) | words / Humanities/Social Science | 0.706 | 0.022 | `ed9878e07e525047…` |
| exploratory paired 0.1 (467 pairs) | PASS | words / Biology/Medicine | A*_pair 0.585 | 0.6211 | `f8003fbc4c9c7fed…` |

Family 48 cells (6 probes × 8 categories), 2,000 permutations (auto-raised from 600). Primary separating cells: words/Humanities (A* 0.706, adj p 0.022), lexicality/Humanities (A* 0.673, adj p 0.0435).

**Reading (per PREREG O6.2 / §4):** the within-question shuffled-label control was DISQUALIFIED, so B1's primary verdict is void ("control failed"); B1 gets no dot.
The planted-cue rule is formally met (4/4), but every planted copy's worst cell is the same words/Humanities cell that already fires in the
unplanted primary and in the shuffled control, so the planted check is uninformative here. Likely cause (diagnosis, not a test): the 1 : (k−1)
construction ties the label rate to question-level properties (questions with many options contribute many label-1 items with similar wording),
and a within-question shuffle preserves that; the effect survives the shuffle, so it is not evidence about the answer key. Any redesign is batch-2 / a new deviation, not a rerun.
Exploratory follow-up (not pre-registered, verdict unchanged): EXPLORATORY_HLE_SHUFFLE.md `31d74df7c7421e1f6e50a5e879c60be78680449513144e9d498821f0e27fccfb`.

## B2 SimpleQA Verified: not applicable (O3). Nothing downloaded or run.

## B3 GraphWalks (derived node-membership label): PASS, box VP "Epoch Verified; no surface cue detected at this n"
openai/graphwalks @ be6cc6ec, both files hash-matched. 1,150 tasks (750 + 400); 144 empty-answer excluded (113 + 31) → 1,006.
Literal PREREG parser dropped 1,006/1,006; DEVIATION_1 (accepted 07:33 MDT) dropped 0. 1,006 tasks (parents 533, bfs 473); 6,714 items (3,357 / 3,357);
128k file 4,152 items, 256k–1M file 2,562; 345 tasks capped at 5. Item and control files re-verified 6/6 against the manifest before scoring.
Planted f = 30% (n ≥ 1496), 1,007 label-1 items marked per copy. Stratum = problem_type.

| run | verdict | worst cell (probe / stratum) | A* | adj p | family | report sha256 |
|---|---|---|---|---|---|---|
| planted s0 (seed 20261029) | DISQUALIFIED | length / bfs | 0.651 | 0.0083 | 12 | `e121280fa3d37c9a…` |
| planted s1 (seed 20261030) | DISQUALIFIED | length / bfs | 0.65 | 0.0083 | 12 | `66d1ef9977837952…` |
| planted s2 (seed 20261031) | DISQUALIFIED | length / parents | 0.656 | 0.0083 | 12 | `34a4a2552adf2118…` |
| planted s3 (seed 20261032) | DISQUALIFIED | length / bfs | 0.654 | 0.0083 | 12 | `b1b270ff97c9526a…` |
| shuffled control (seed 20261027) | PASS | alpha_last / bfs | 0.511 | 0.7521 | 10 | `15ba0b5cedd63f60…` |
| **primary** | PASS | words / bfs | 0.514 | 0.7854 | 10 | `9a7e6ed90517c76a…` |

Planted cue detected 4/4 (valid); shuffled control PASS (satisfied); primary PASS, no separating cell, 600 permutations.
The `length` probe is not in the primary/shuffled family (10 cells, 5 probes × 2 strata) because every node id is 10 characters; it enters only in the planted copies (12 cells).
Reading rules: node ids are md5(str(int))[:10]; a cue that needs decoding the hash is outside the six probes, so this PASS means "no surface cue detected at this n", nothing more.
Pre-lock exposure disclosure applies (the drafter saw the 128k file's id format and counts). Paired mode is defined for B1 only (O1) and was not run for B3.

Internal only: HLE items, explain tables and WHY strings stay on the DGX. Nothing was published, committed or uploaded.