Why “settled Brier only”

The problem

Many systems print a confidence or “accuracy” number on every run. That number is easy to invent and hard to falsify. Readers treat it as skill when it is only a decorative float.

The rule

Hyperlex separates claims from scores:

1. Open analysis may emit forecasts (probabilities) and always sets provenance.brier = null.
2. An operator settles each forecast with an outcome (TRUE / FALSE / VOID).
3. Only then does score-series compute Brier (and related decompositions) from settled pairs.

Missing outcomes yield NOT_COMPUTABLE — never a fabricated score.

Why this is a strength

• Auditable: scores attach to receipts + settlements.
• Honest: Phase 5 research cannot launder itself into calibration.
• Comparable: series Brier means the same thing over time.

Operator path

python3 scripts/hyperlex.py pipeline "rizz" --route offline
python3 scripts/hyperlex.py pending
python3 scripts/hyperlex.py settle --forecast-id <id> --decision TRUE
python3 scripts/hyperlex.py score-series --mean-shift --verify-chain

See also

• Glossary
• Brier design
• Operator loop