Brier parity

ADR-0002 gates the legacy → new cutover on 7 consecutive nightly runs where the new kfish-core Brier is within 0.002 of legacy. First night green on 2026-04-19.

First-night evidence

All three archived legacy retrodiction runs, plus the first new-path computation, agree on Brier to 4 decimal places:

Run	N	legacy Brier	new Brier	Δ	gate
retro_v2_20260412_221752	30	0.2029	0.2029	0.0000	✅
retro_v2_20260413_062123	200	0.2026	0.2026	0.0000	✅
retro_v2_20260413_150229	100	0.1799	0.1799	0.0000	✅

Extremized Brier and ECE also track closely; ECE diverges only because legacy uses netcal's binning and the new path uses an equivalent numpy reimplementation. ECE is advisory per ADR-0002 — Brier is the gate.

[PARITY OK] 3 run(s) passed the ADR-0002 Brier gate.

Murphy decomposition of the 200-market baseline

For the N=200 run (the de-facto production baseline):

Component	Value	Meaning
reliability	0.0359	calibration gap — small is good
resolution	0.0734	how much the predictor separates classes — large is good
uncertainty	0.2400	data-bound, \(\bar o (1 - \bar o)\)
Brier	0.2025	reliability − resolution + uncertainty
skill score	0.1562	1 − Brier / uncertainty (vs climatology)

Reproducing

python scripts/brier_parity.py --all
# or for a specific run:
python scripts/brier_parity.py --input data/retrodiction/retro_v2_20260413_150229.json
# gate on ECE too (stricter, not required by ADR-0002):
python scripts/brier_parity.py --all --strict

Bug found during validation

The initial script treated ground_truth as boolean via int(bool(...)). One legacy prediction had ground_truth = 9.045e-08 (effectively zero / "NO"), which bool() coerces to True, flipping that row's contribution from ~0.52 to ~0.08. Fixed with 1 if float(truth) >= 0.5 else 0. Without the fix, the script reported a phantom 0.005 gap that wasn't in the data.

The 7-night schedule

1. ✅ 2026-04-19 (today)
2. ⏳ 2026-04-20
3. ⏳ 2026-04-21
4. ⏳ 2026-04-22
5. ⏳ 2026-04-23
6. ⏳ 2026-04-24
7. ⏳ 2026-04-25 — cutover-eligible

The nightly runner (.github/workflows/nightly.yml) runs the new path against a fresh batch of resolved markets; the parity script runs at the end and annotates the status here. After the 7th consecutive green, legacy src/ is archivable.