First stable version of the project: a movie app with a real spoiler
gate (post-viewing content never leaves the server until the client
asks for it) plus the evaluation harness that measures that promise
instead of taking it for granted.
Includes
- Twistify app: catalogue with 7 curated entries researched with cited
sources, theme/sort filters, spoiler curtain, editable comments with
no accounts required. - Evals harness: leakage / grounding / richness, an offline-calibrated
judge, 20/20 titles with researched ground truth. - Baseline generator (Anthropic, tool use, no retrieval) implemented —
still pending a run across the 20 titles. - CI on GitHub Actions (8 tests, no secrets).
Pending for the next version
- Run the baseline and publish real leakage/grounding/richness numbers.
- Calibrate the judge against an external benchmark (TV Tropes Movies /
IMDB Spoiler Dataset). - Real retrieval (Milestone 1: TMDB/OMDb/Wikipedia) + a claim verifier.
See docs/DESIGN.md for the full reasoning behind
each design decision.