The fidelity-check stage audits the generated study guide against the source corpus and flags fabricated numbers, entities, and relationships. It is the audio/written trust story, and it has never been validated by a real GPU run — its own code comment says so.
Until it has been, it should not appear in any pitch as a shipped guarantee.
Work:
- run it end to end on a real module
- confirm it catches seeded fabrications (inject known-false numbers and entities into a study guide and verify they are flagged)
- measure the false-positive rate, since an auditor that cries wolf gets ignored
- publish what it does and does not catch
Also publish alongside it: the timing and VRAM table per hardware tier, and measured cost estimates for the external-key path, so the honesty ledger in PLAN.md is backed by numbers.
The fidelity-check stage audits the generated study guide against the source corpus and flags fabricated numbers, entities, and relationships. It is the audio/written trust story, and it has never been validated by a real GPU run — its own code comment says so.
Until it has been, it should not appear in any pitch as a shipped guarantee.
Work:
Also publish alongside it: the timing and VRAM table per hardware tier, and measured cost estimates for the external-key path, so the honesty ledger in PLAN.md is backed by numbers.