ai-01 reproduction

Public source retained with the related article. Generated results, dependency directories, runtime storage, secrets, and oversized binary artifacts are intentionally excluded.

README

# AI-01 evidence-loop fixture

This fixture supports “AI Should Shorten the Evidence Loop.” It uses one
synthetic research question about large JSON identifiers and two synthetic
candidate artifacts. No named AI model produced either candidate, and the
fixture is not a model benchmark.

Run:

```sh
./run.sh
```

The digest-pinned run:

- executes seventeen safe, unsafe, reviver, source-context, BigInt,
  decimal-string, and validation cases;
- retains four declared assumptions and three unresolved production questions;
- verifies or bounds ten source, observation, inference, assumption, and
  unresolved claims;
- critiques four plausible claims, rejecting three and revising one;
- retains a score-free human-owned decision for the synthetic API contract;
- rejects thirty exercised semantic mutations;
- fetches and retains six primary or authoritative source bodies; and
- verifies runtime, dependency, source, candidate, experiment, and decision
  provenance.

The final synthetic decision selects canonical decimal strings because of its
declared clients and opaque-identifier contract. It does not claim that every
API should use strings or that a lossless numeric parser is inferior.

After committing the fixture, verify it from a clean archive:

```sh
./bin/run-clean-export.sh
```

The clean-export result retains the exact commit, experiment and research
hashes, source evidence, mutation count, command, output, and exit code under
`results/clean-export/`.

Retained files