Data audit (P1)
All SHA-256 values recomputed with shasum -a 256 and match data/README.md and bundle.json.
| file | SHA-256 | lines (wc -l) | content |
|---|---|---|---|
| data/studies.json | 124d1f3e…babbe1 | 5 items (JSON list) | id, source, short, extract — no duplicates, no missing fields |
| data/README.md | c7d07d4b…0128 | provenance only (read; not an analysis input) | |
| sources/arxiv-2410.12944v3.txt (S1, Paradis) | d5270b44…803b | 112 | abstract, IV-C, VI-B, VI-C, Table III |
| sources/arxiv-2507.09089v2.txt (S2, Becker) | b6af8a2b…d2d | 132 | abstract, 2.3, 3, 3.1, D.1–D.3 |
| sources/arxiv-2411.01067v2.txt (S3, Bono & Xu) | 46a1ee4e…053a | 109 | abstract, 3, 4, 4.2 Table 3, App. B |
| sources/arxiv-2511.13860v1.txt (S4, Bono) | 8fb2520f…5ca | 162 | abstract, 2, 3.1 Table 1, 3.3 Table 3, App. B |
| sources/arxiv-2304.09339v1.txt (S5, Kreitmeir) | 1db48b1c…e06e | 108 | abstract, 3.1, 4, Table 1 |
"Schema": unstructured text; tables rendered one row per line with | separators. Units: S1 log minutes (coefficients), S2 percent, S3 minutes, S4 seconds (as labelled), S5 counts / log counts (DiD coefficient). No date ranges matter beyond publication versions (2023–2025).
Extracted values and anomalies
- S1 Table III lines 76–78: ExpCon "-0.25. -0.24. -0.30* -0.14", SEs "(0.14) (0.14) (0.14) (0.17)"; Model 2 = -0.24 (0.14). VI-C line 58: 95% CI [-0.51, 0.03], p=0.086. Anomalies: (a) VI-B line 48 gives Mean LogToT = 4.6 for AI and 4.3 for no-AI while also saying the AI group was faster (96 vs 114 min). ln(96)=4.56 and ln(114)=4.74, so the 4.3 value contradicts the rest of the text (not used). (b) VI-B reports N=47+46=93, while Table III says n=96 (not used for θ; sample reported as printed). (c) CI-implied SE (0.54/3.92=0.1378) vs printed SE 0.14: they agree up to rounding. The printed SE(b)=0.14 is used because it reproduces p=0.086 under a normal approximation; the CI-implied SE gives 0.081. (d) stray "p<" / "p=" lines are footnote artefacts (README).
- S2 headline "+19%" (abstract line 34; Section 3 line 62). CI only in Figure 1 / Figure 15 (not reproduced; header line 8 confirms). No SE, CI or p-value printed → not poolable. Raw ratio 34% (footnote 12) is not the regression estimate; not used.
- S3 Table 3 'All Content' (Aggregate): Overall 53.84, Treatment 44.29, Control 63.10, Improvement -29.79%. Anomaly: 44.29/63.10−1 = −29.81%, not −29.79% (a rounding difference); the printed −29.79% is used. Only stars ( = 0<p≤0.01 per editorial note) → not poolable. The control duration is a model-predicted duration at Copilot-group accuracy (App. B), not a raw mean.
- S4 Table 3 Full Corpus: Aware 50.5 s, Control 56.9 s, % Time Savings 11.1, p=0.471. Anomalies: 1−50.5/56.9 = 11.25% vs printed 11.1. The Malicious / Non-Malicious columns (2.8, 1.8, 1.9, 2.4 "s") do not match the 50.5/56.9 s full-corpus values in magnitude, and their implied savings (−47.4% / 25.0%) differ from the printed −52.9 / 23.6. The unit of those columns is ambiguous (possibly minutes, rounded). Only the Full Corpus column is used. Aware arm = agent verdicts + rationale + agent-ordered queue vs Control (no AI): an AI-vs-no-AI comparison on queues drawn from the same corpus.
- S5 Table 1 panel A col (3) Output: Treat×Post −0.278 (0.295), N 18332. Outcome = activity counts around withdrawal of access, not task-completion time → not harmonisable. Anomaly: notes legend "** p<0.5" (presumably 0.05; typo in source). The date line "August 24, 2026" is a rendering artefact (extract header).
Information-availability
No input violates the rules. Values that appear only in figures (S2 CI, S3 Figure 3, S5 Figure 1) are treated as unavailable.