research.daybreak_judge_forward_eval
Daybreak actionable paper-Call forward evaluation: booked long/short Call win rate and drift per model/prompt version. The legacy accuracy column does not cover watch/no-trade decisions.
- Coverage
- Unknown
- Freshness observation
- Unknown
- Point-in-time support
- Unregistered
Fields
Schema status: Observed. Captured .
| Name | Type |
|---|---|
model_id | LowCardinality(String) |
prompt_version | LowCardinality(String) |
resolved_model | String |
model_fingerprint | String |
temperature | Float64 |
eval_date | Date |
eval_ts | DateTime64(3, 'UTC') |
eval_run_id | String |
n_total | UInt32 |
n_forward | UInt32 |
n_dropped_preregistration | UInt32 |
n_scored | UInt32 |
accuracy | Nullable(Float64) |
cohort_is_stale | UInt8 |
accuracy_muzzled | UInt8 |
determinism_probe_repeats | UInt8 |
determinism_probe_agreement | Nullable(Float64) |
determinism_probe_status | LowCardinality(String) |
drifted | UInt8 |
drift_delta | Nullable(Float64) |
drift_tolerance | Float64 |
prev_accuracy | Nullable(Float64) |
drift_reason | String |
version_ts | DateTime64(3, 'UTC') |
The field guide explains unknown observations.