# When Do Diverse Models Fail Together? — Stage 1 evidence bundle

Version: `0.1.0-stage1`  
Release status: public Stage 1 evidence bundle

This directory accompanies the Future Shock research briefing **When Do Diverse Models Fail Together? Difficulty, Shared Evidence, and Common Shocks Across Language-Model Fleets**.

## What Stage 1 supports

This bundle supports **report-level inspection and receipt reconstruction**. It lets a reader inspect the frozen study reports, common-shock contract, primary analysis summaries, support classifications, missingness disclosures, and the exact values used by the public briefing and PDF.

It does **not** reproduce the historical model calls or regenerate every analysis from raw event-level evidence. The supplied manuscript package references a larger local evidence estate containing code, notebooks, raw accepted payloads, matrices, and bootstrap outputs that is not included here. A fresh run against current APIs would be a temporal replication, not an exact replay of the historical endpoints.

## Included files

- `study-summary.json` — machine-readable headline study and release posture
- `common-shock-results.csv` — four primary shared-versus-independent contrasts
- `support-and-diagnostics.json` — support, neutral checks, canaries, degeneracies, and correction disclosures
- `base-study-report.md` — frozen retrospective/prospective base-study report
- `common-shock-report.md` — frozen common-shock result report
- `common-shock-missingness.md` — frozen observed-run missingness and support report
- `common-shock-contract.md` — governing prospective contract and amendment history
- `legacy-summary.json` — frozen nine-endpoint legacy analysis summary
- `contemporary-summary.json` — frozen six-endpoint contemporary analysis summary
- `pilot-gate.json` — frozen pilot gate result
- `LIMITATIONS.md` — public interpretation and reproduction limits
- `ERRATA.md` — public interpretation corrections for preserved upstream files
- `manifest.json` — byte sizes and SHA-256 digests
- `CHECKSUMS.sha256` — checksum file for verification

## Reproduction boundary

The public claims are bounded as follows:

1. The legacy evidence-error interval excludes unity, but the result is classified `estimable_degraded`, with 665/800 matched items.
2. The contemporary panel has confirmatory-quality support but wide intervals that include both no change and effects near the legacy estimate.
3. The neutral mean-equivalence check failed in both populations.
4. Contemporary start, midpoint, and end canaries have zero observations.
5. Endpoint degeneracy affects the effective estimand in specific treatment arms.
6. The retrospective execution reported 3,253 support-qualified pairs from 77,815 possible unordered pairs. Stage 1 lacks the event-level pair-support matrix needed to regenerate that count.
7. Preserved source files contain known stale labels and boilerplate documented in `ERRATA.md`; the public interpretation follows the errata.
8. Hashes establish file integrity. They do not establish scientific validity or independent reproducibility.

## Verify the files

From this directory:

```bash
sha256sum -c CHECKSUMS.sha256
```

The human-readable article and methodology route remain the best entry points. These files are the audit layer behind them.