Corrections

Every number in the paper was recomputed from the raw records twice. This is what reproduced, what did not and what was changed.

Two independent passes recomputed the paper's numbers and claims (paper/main.tex at 99bec72d) from the raw per-run records. The analysis summaries were used only to locate things.

Pass 1 (Claude Opus 5.5 in Claude Code): its own parser and statistics (permutation tests with a different seed, Holm, bootstrap, κ), compared cell by cell with the analysis outputs, then a claim-by-claim audit. Every Holm-adjusted p between 0.01 and 0.10 was rechecked with 100,000 permutations.
Pass 2 (GPT-6 Astra at extra-high reasoning effort, through the Codex CLI in a read-only sandbox): an independent audit of the same records, 663 checks.

Scripts, outputs and the Pass 2 report and log: campaign_logs/validation_2026-11-10/.

What reproduced

Coverage: 5,510 reviews (5,509 judged); drift 4,258 parsed conversations (82 unparseable); IP 1,078 parsed (2 unparseable) plus 150 pipeline runs; 1,050 judged DD items; 17 models.
Contract review: all 510 cells have the same n, mean skew and Δ. Under a different seed no significance call changes. Tables 2 and 3 reproduce at printed precision.
Drift: all 299 rows have the same n, final P and Δ. Table 4's cells and stars reproduce, as do the order counts (22, 23, 22, 21 of 23).
IP: all 60 rows reproduce, and Table 5 with them. Pipeline: 21 to 33 against 1 to 7. Haiku 5.5 −14.1 (Holm p 0.042; 0.044 with 100,000 permutations), with the deed named present in 12 of 20 runs.
Reliability: issue position κ 0.74 (raw agreement 0.98; 537 reviews, 8,592 codes); naming κ 0.81; DD hedge κ 0.68 and cluster κ 0.23; human κ 0.54 (132 items) and 0.91 (68 items).
100,000-permutation recheck: one call changes. GPT-6 Luna's dismissal-case gloss A after the rewrite has Holm p 0.053 in the analysis output and 0.033 with 100,000 permutations (0.026 under Pass 1's seed). No mark in the tables or the figure changes.

25 corrections

Abstract and conclusion1

Abstract, §1, §8Pass 2

As written: Assertions "almost never moved the final answer"
In the raw records: 9 of 20 assertion cells moved significantly, 8 of them against the asserting party
Now: "toward that party"

Methods6

§3.2, §4.2Both passes

As written: "Eight glosses"
In the raw records: Ten gloss texts on eight items
Now: Corrected

§3.7Both passes

As written: DD red-flag, fabrication, lone-flag and admission "κ = 1.00"
In the raw records: Agreement 100%, but κ is undefined because both judges used one category. Laundering and recommendation κ 0.83 were not reported
Now: Reworded; laundering and recommendation added

§3.7Both passes

As written: "39 … (PARTIAL or ACCEPT)"
In the raw records: 25 PARTIAL, 8 ACCEPT, 6 NOT_ADDRESSED
Now: Label list corrected

§3.7Pass 2

As written: "a deterministic random 10% sample"
In the raw records: The rule selects 546 reviews; 537 were second-coded. Nine API reviews were collected after the second-judge run
Now: Stated

§3.8, §6Pass 2

As written: Fable 5.1 has "10 replicates"
In the raw records: Five contract-review C0 cells have 20
Now: Stated

§3.6Pass 2, confirmed by Pass 1

As written: Terse notes "at most 90 words"
In the raw records: An instruction, not a cap: at least 4 of 120 notes exceed it
Now: "instructed to stay within 90 words"

Contract review3

§4.1Both passes

As written: "Every model named framing far less often than injection"
In the raw records: Opus 4.8: 0.96 against 1.00. GPT-6 Sol: 0.10 against 0.10. dots.3, Gemini and Nemotron name framing more often (0.80 to 0.90 against 0.25 to 0.50)
Now: Limited to the Claude models other than Opus 4.8; GPT rates added

§4.1Pass 1

As written: "Only three cells survived correction" (and Haiku 5.5 "had none")
In the raw records: Haiku 5.5 A2P −0.60* in the PDF family
Now: "Outside the PDF conditions"

§4.1Both passes

As written: "No model moved toward the seller" (PDF)
In the raw records: 10 of 15 A2P Δ are positive (at most +0.50), none significant
Now: "significantly"

Interpretive drift5

§4.2Both passes

As written: The strong arm "moved against its own side in every model"
In the raw records: Strong A: 13 of 13. Strong B: 7 of 13. GPT-6 Luna −18.8*, with B
Now: Stated per side

§4.2Both passes

As written: Note alone had "no meaningful effect (|Δ| ≤ 3.3 for every core model)"
In the raw records: True only on the commercial case. Up to 11.7 on the new cases (Opus, insurance, note B). All nine significant note effects go against the note's slant
Now: Limited to the commercial case; the new cases and the pattern stated

§4.2Pass 1

As written: Gloss A "lost significance for … GPT-6 Luna"
In the raw records: +7.3 to +5.4, Holm p 0.053 (10,000 permutations) or 0.033 (100,000)
Now: Luna's change given as numbers only

§4.2Pass 2 raised it, Pass 1 quantified it

As written: The rewrite "shrank the affected side's effect"
In the raw records: Employment gloss A fell in all five core models. The rewritten employment B glosses moved models about as much as before. Re-running the insurance A-side arms with unchanged text moved estimates by up to 10.5 points and cost three cells their significance
Now: Qualified as noisy and suggestive

§4.2, §5, §1Both passes

As written: "Opus 5.5, and GPT-6.1 Sol on one case, moved against them"; "two models reacted against slanted glosses rather than with them"
In the raw records: Opus: 7 of 8 significant gloss or drip effects go against the slant, but employment gloss B is −5.6*, with it. Sol: 4 against (commercial A side, insurance B side), 4 with (B side on the commercial and dismissal cases)
Now: Qualified

IP assignment2

§4.3Pass 1

As written: The soft anchor "roughly doubled" the bundle's effect
In the raw records: 2.0 to 2.9×
Now: "two to three times larger"

§4.3Pass 1

As written: The bundle moved the answer "just as hard the other way"
In the raw records: −36 to −86 in the positive control, against +5 to +25 in the no-deed case
Now: "even further the other way, by 36 to 86 points"

Due-diligence pipeline4

§4.4Both passes

As written: No pipeline "recommended proceeding"
In the raw records: 0 of 200 unconditional PROCEED. 44 of 100 naive-harness runs recommended PROCEED_WITH_CONDITIONS (31 of 40 Luna to Sol), against 0 of 100 hardened
Now: "unconditionally"; the conditional recommendations reported

§4.4Both passes

As written: Direct readers kept the hedge in "80 to 100%"
In the raw records: Fable 5.1 on discovery: 1 of 10, hardened 9 of 10. Everyone else 80 to 100%
Now: Exception stated

§4.4Pass 1

As written: Fable 5.1 hardened in "90% of discovery runs"
In the raw records: 80% with normal notes, 90% with terse, 90% direct
Now: 80 to 90%

§4.4Pass 1

As written: Compression lost "the links between scattered documents"
In the raw records: Rests on the cluster field the paper excludes (κ 0.23). Deciders were coded CONNECTED in 80 to 100% of runs
Now: Removed

Discussion and limitations3

§5Pass 1

As written: dots.3 "followed the instruction almost completely"
In the raw records: Mean 23.9 of 32; full compliance in 8 of 20 reviews
Now: "largely followed"

§5Pass 2 (could not verify), Pass 1

As written: "vendor and model differences here were larger than the differences between attack types"
In the raw records: No defined measure. In the IP study the attack type mattered far more than the model
Now: Replaced with a statement the data support

§6Pass 2

As written: The bias "makes contract-review effects conservative"
In the raw records: The direction is suggested, not established: the sample is small and label-stratified
Now: "if anything, overstates"

Figure1

Figure 1 values filePass 2

As written: Drift and IP values printed to two decimals
In the raw records: The summaries give one decimal (19 of 68 estimates and 29 of 68 intervals affected)
Now: Printed at the source's precision

The Statistics section now states the permutation count (10,000) and the 100,000-permutation recheck. The Use of AI tools section names both passes.

Checked and left unchanged

Wrapper sizes (§3.1). They come from the isolation probes (docs/ISOLATION.md), which the raw reviewer records do not show.
"Calibrated P about 75 to 95" for the positive control. This states the design's intent, not a result.
The second judge's 109 DD items. These are the complete 10% sample: no sampled item is missing.