bw-gate

/baywright:bw-gate — the calibration gate

Run the honesty core (see calibration) and record it, fail-closed.

1. Prior-predictive sanity (if not already settled in priors-and-prior-predictive): does the model generate plausible data? If not, stop — fix the priors first.
2. SBC. Across many prior-drawn truths, simulate, fit, and rank the truth among the posterior draws; the ranks must be approximately uniform. Start with a modest replication count to see the shape, then scale up for a confident verdict.
3. Parameter recovery. On simulated data with known truth, does the posterior recover it with honest uncertainty?
4. Read the results yourself — the rank histograms and recovery — and interpret them in plain language for the human.
5. Record honestly in the ledger. A non-uniform SBC rank means overconfidence or bias, not a reason to loosen the test. Apply an honest model fix (real measurement noise, a heavier tail) and re-gate. Never tune priors, seeds, or thresholds to pass. Set the calibration stage to done only when the ranks are honestly approximately uniform; otherwise failed, with the fix attempted noted in the honesty log.

Drive your live session to run this; consult the current documentation for your tool's SBC / LOO utilities (if SBC isn't built in, the loop is small: prior draw -> simulate -> fit -> rank). SBC is compute-heavy — say so, and scale replications deliberately.