AI-assisted educational prototype. A browser-only, deterministic diagnostic over 13 synthetic proposed-action records. No API key, model call, live tool, backend, analytics, third-party dataset, or private transcript. Prepared for human review, not evidence that the account owner independently wrote or validated it.
Open index.html locally, or serve this directory with python3 -m http.server 8000. Deploy as a static HTML Space only. No paid hardware or external inference needed.
Run node test.js. The tests verify the 13 fixture counts, six detected field changes, zero benign flags, the deliberate undetected content case, object-key ordering, extra/missing fields, and JSON escaping. The UI recalculates counts from fixtures.js. fixtures.json is the same data in downloadable JSON.
The exact-field comparator is not a model benchmark, an authorization engine, or a production security control. It flags any structural difference, so valid aliases and irrelevant changes may be false alarms. A matching action may still be unsafe. Labels and authorization fields were created together; the six detected violations are designed examples, not independent evidence of generalization.
Derived from the private synthetic action-contract diagnostic prepared alongside Anansi Watchdog. This app does not execute Anansi's Python evaluator, copy its implementation, or report live-model results. No Anthropic affiliation or endorsement. No third-party datasets or restricted transcripts. All addresses and paths are inert examples. No applicant qualifications, training claims, or private account information included.
Published as a public static Hugging Face Space with the owner's approval. No open-source license has been selected; public visibility does not itself grant reuse permissions.
Can Correct Data Prevent Hallucinations? evaluates clean training data, distillation, correlated teacher errors and a limited numeric-classifier experiment. AI-assisted working article, not peer-reviewed or a language-model benchmark. The experiment package is not included in this publication.