# Data audit (P1)

All SHA-256 values computed with `shasum -a 256` match data/README.md and bundle.json.

| file | SHA-256 | lines (wc -l) | content |
|---|---|---|---|
| data/studies.json | 124d1f3e…bbe1 | 5 items (array of 5 objects; keys id, source, short, extract) | task-designer item list |
| data/README.md | c7d07d4b…0128 | 17 | provenance |
| data/sources/arxiv-2410.12944v3.txt | d5270b44…803b | 112 | Paradis et al. 2024 |
| data/sources/arxiv-2507.09089v2.txt | b6af8a2b…d2d | 132 | Becker et al. 2025 |
| data/sources/arxiv-2411.01067v2.txt | 46a1ee4e…053a | 109 | Bono & Xu 2024 |
| data/sources/arxiv-2511.13860v1.txt | 8fb2520f…a5 | 162 | Bono 2025 |
| data/sources/arxiv-2304.09339v1.txt | 1db48b1c…06e | 108 | Kreitmeir & Raschky 2023 |

These are text extracts, not tabular data: "rows/columns" = lines; no missing values in the usual sense. No duplicate study items (ids S1–S5 unique, sources unique).
Date ranges: retrieval 2026-10-30T04:19–04:20Z; publications 2023–2025. No availability-rule violation found in the inputs; the only rule-relevant item is that
S2's confidence interval appears only in Figure 1 (not reproduced) → must NOT be used.

## Per-source findings (line numbers refer to `cat -n`)
- **S1 Paradis** L54-58, L76-78: ExpCon Model 2 = -0.24 (0.14), marker '.', 95% CI [-0.51, 0.03], p=0.086.
  Checks: CI-derived SE = 0.54/3.92 = 0.1378 vs printed 0.14 (consistent within rounding); 2·Φ(−0.24/0.14) = 0.0865 ≈ 0.086 (consistent).
  Anomalies: Table III says n = 96 while VI-B reports N=47 (AI) + N=46 (no-AI) = 93; VI-B says the AI group (Mean LogToT=4.6) was faster than no-AI (4.3) — the log means appear swapped/inconsistent with the raw means (96 vs 114 min); "21% faster" is (1−exp(β))·100, i.e. a 21% reduction in time, not a speed rate. Stray "p<"/"p=" lines are rendering artefacts (README).
- **S2 Becker** L34, L62, L72, D.1–D.3: "+19%" increase in completion time (regression-based S = exp(β)−1, log time regressed on AI + log forecast). Uncertainty: none printed in the extract; the CI is displayed only in Figure 1/Figure 15 (omitted). Raw ratio 34% (not used; the extract designates the regression-based headline).
- **S3 Bono & Xu** L55: All Content | 53.84 | 44.29 | 63.10 | -29.79%***. 44.29/63.10−1 = −29.81% vs printed −29.79% (rounding of displayed minutes; small). Uncertainty: stars only; star scheme non-standard (*** = p≤0.01 per editorial note). Control arm could use other AI tools (25% used e.g. ChatGPT, L34) → contrast is Copilot vs no-Copilot.
- **S4 Bono** L90-96: Full Corpus Aware 50.5, Control 56.9, % Time Savings 11.1, p-value 0.471. Anomalies: (a) 1−50.5/56.9 = 11.25% ≠ printed 11.1; Malicious and Non-Malicious columns likewise do not reproduce from displayed times (−47.4% vs −52.9; 25.0% vs 23.6) → printed % presumably from unrounded data/another estimator; (b) units: row labels say "(s)" per email but Malicious/Non-malicious values (2.8, 1.9, 1.8, 2.4) and Full Corpus (50.5, 56.9) cannot both be seconds per email (full corpus is not a weighted mean of the parts); Full Corpus may be a per-queue total. A ratio-based θ is unit-free so this does not affect θ. The Aware arm also differs in queue ordering (agent-prioritised) — the AI contrast bundles ordering + verdicts.
- **S5 Kreitmeir & Raschky** L80-88: Output col (3) panel A Treat×Post = -0.278 (0.295), N = 18332. Outcome is an activity count (sum of GitHub events; whether logged is ambiguous, L42) in a DiD around withdrawal of ChatGPT — not completion time → not harmonisable. Notes typo "** p<0.5" (presumably 0.05). Row label rendered as "\times TreatPost". Date line "August 24, 2026" is a rendering artefact (L8).

## Handling
Printed values copied verbatim; harmonisation uses printed headline quantities only (S3: −29.79%, S4: 11.1%), with the ratio-of-displayed-means alternative reported only as sensitivity.