# Reproduce the deterministic baseline

Download `evidence-guide-reference.zip`, `ai-systems-evaluation.json`, and `ai-systems-reproduce.mjs` from the AI Systems Evidence page. Extract the archive into a clean directory, then place the JSON and script in that directory alongside `package.json` and the `reference` directory.

Use Node.js 22.13 or later. From that directory:

```sh
npm ci
node ai-systems-reproduce.mjs > reproduction-result.json
```

Dependency installation requires package-registry access. Answer reproduction then runs locally with network fetch disabled, zero model calls, no API credential, no database, and no provider charges. The archive includes its dependency lockfile. The downloaded script imports the same answering core used by `reference/cli.ts`; this preserves withdrawn-evidence inputs that the simpler CLI does not expose.

The result compares the full answer objects (including visible summary, limitations and sources) against every frozen source-only answer for all 120 authored, consumed questions. It reports mismatches, provider attempts, and a newly measured local p50/p95. An exit status of 1 means a mismatch or prohibited provider/network call occurred. Different hardware can change timings.

The archive answers are frozen to the evaluated implementation: commit `06226e4914928d61df826972184fc3c2b9aedb27`, runtime SHA-256 `8918e7eb747e23b413000ec3e2f117345f322c6727560a4454b6093f15ac2af5`. The public page is a presentation successor. Archive source hashes identify the original files; the sanitized dataset is a transformed public export and does not have the original private packet's byte hash.

The selector responses, two independent agent score sheets, and eight adjudications are retained in the dataset for inspection. This script does not reproduce live model outputs, independently reproduce semantic judgments, or establish real-user usefulness. Both published sets are consumed; they are not fresh future acceptance questions. All questions are authored research prompts, with no real visitor questions included.