Model Calibration
Calibration Honest Limits
- comp_n is primary: All estimates with n<5 are extremely provisional. At n<30, win probability is not computed.
- Vision misreads: The vision model can misidentify holograms, serials, and certificate issuers. Vision alone never triggers suppression — it requires corroborating text signals.
- Cards out of scope: Card (trading card) pricing requires grade-specific comps (PSA 9 vs PSA 10 can be 10x). The scanner suppresses max_bid for cards and directs to 130point.
- Null hypothesis: Until the model demonstrates better-than-baseline MAE at n≥50, the system's value is coverage and deal discovery, not estimation quality. Don't over-rely on fair_estimate until beats_baseline=true.
- Implied close / win prob: Both require n≥30 historical snapshots at the relevant time window. Below that, they display "insufficient history".
No calibration runs yet. Runs nightly at 3am after outcomes are recorded.