Week 1 — poker-coach scaffolded, first reads logged

The build started with the plan at ~/.claude/plans/lucky-tinkering-pudding.md. Two deliverables: the CLI at ~/poker-decision-support/, and the lab page at /lab/poker/ you are reading now. CLI shipped first because the lab dashboard is empty without data.

What’s working: the eight-table SQLite schema, the read-add → read-resolve → calibrate flow, audit report generation (HTML body deferred until the API key is set), the export-stats command that writes to src/data/poker-stats.json and drives the calibration dashboard above the atlas.

What’s surprising: four test reads with confidences spread across the spectrum produced a Brier score of 0.1013, which the system labels as “SUPERFORECASTER RANGE.” The label is meaningless on n=4. The point is the pipeline works end-to-end and the math checks against hand calculation: (0.7-1)² = 0.09, (0.85-1)² = 0.0225, (0.30-0)² = 0.09, (0.55-1)² = 0.2025, mean 0.1013. The first session through with real hands begins this week.

Phase 1 starts with foundations. The C1 capstone unlocks at 1,500 XP — roughly 15 chapters complete or 4-5 weeks of consistent work plus the 10k-hand sample. The win-rate over that sample is not the thing being measured. The audit trail is.