A track record you can audit.

Mechanism: The research-only version `short-term-crypto-half-deviation-v1` records probability = market + 0.50 × (champion − market) for every contract admitted to the unchanged v6 forecast universe. It writes a normal append-only forecast row with its own version and is structurally emitted as WATCH, so it cannot create a paper or broker order.

Metric (fixed at registration): On paired, non-voided contracts predicted before settlement by both versions, compare mean Brier contribution for the challenger and v6. PASS requires the paired bootstrap confidence interval for (v6 Brier − challenger Brier) to be entirely above zero at at least 100 independent resolved contracts, with no worsening in either BTC or ETH by more than 0.002 Brier.

Failure criterion: FAIL if at 200 paired independent resolved contracts the bootstrap confidence interval does not exclude zero in the challenger's favor, or if either BTC or ETH worsens by more than 0.002 Brier at n ≥ 50 for that asset. No shrinkage factor, population, or execution rule changes after registration.

Origin disclosure: Registered after the v6 performance review showed a near-zero aggregate model-versus-market Brier delta while mixed historical paper/shadow P&L was negative. That observation does not prove that 50% is optimal, that probability scoring causes P&L, or that the challenger will improve. The fixed halfway value is a conservative shrinkage test, not a fitted response to the historical cohort.

Mechanism: Challenger `deviation-gated` runs the champion's weights unchanged plus the reserved key emit.maxDeviation = 0.05. Its probabilities are identical to the champion's — only emission differs: directional signals (LONG_YES / LONG_NO) are refused when |model − market| > 0.05, with the refusal reason recorded. Shadow predictions and shadow signals accrue through the standard hourly refresh pipeline.

Metric (fixed at registration): Sign test on the per-prediction beat rate — the fraction of the challenger's EMITTED directional set (shadow signals LONG_YES / LONG_NO on clean resolved rows: non-DEMO, predicted before resolution, not effectively resolved at prediction time) where the model's Brier contribution is strictly lower than the market's. Beat rate, not Brier delta, because the delta is bounded by construction inside a small-deviation band. PASS requires beat rate > 0.5 with |beatRate − 0.5| > 1/√n at n ≥ 100.

Failure criterion: FAIL if at n ≥ 200 emitted clean resolved rows the beat rate is not significantly above 0.5. No threshold adjustment after registration — a variant with a different cap or confidence floor is a new experiment with its own registration.