STRATEGY — the north star (read this first, every session)

This is the context anchor. Every Claude session — CLI on the VM, reviewer
without data access, web session — starts here. If a request seems to conflict
with this document, ask; if work doesn't serve the goal below, question why
it's being done.

The goal (one sentence)

Find, prove, and operate a repeatable positive-CLV betting edge on NRL
opening markets — proven by the promotion gates, never by vibes — while every
instrument that measures it stays honest.

"Proven" means, per nrl-bet-advisor/data/promotion_gates.py: ≥100 settled
paper trades, ROI > +5%, positive mean AND median CLV, no round >30% of
profit, holding across home/away — out of sample. Until a strategy clears
that, everything is a candidate/hypothesis/paper trade. No real money is
staked on anything this repo produces; promotion decisions live with the human.

What we know (hard-won, don't re-derive)

1. The opening price is soft; the closing price is efficient. Established
   three independent ways (open-only ROI, closing-line agreement, negative
   close-priced diagnostics). Every strategy is a race to the early price.
2. Team-level features can't beat the H2H market (λ ≈ 0.02–0.10). Stop
   building Elo remixes.
3. The lineup/player channel carries real information (RAPM λ=0.141,
   CLA 0.583 @ z=5.22) but its magnitude is noisy — scaling and the
   team-list-timing window are where the work is.
4. Instruments lie unless test-enforced. Two fake edges (closing-feature
   leakage; a broken join measuring zeros) got within one step of production.
   Both are now structurally impossible — keep it that way.

The operating loop (how work happens here)

• Plans and specs live in docs; the current build/test plan is
  docs/test-cli-instructions.md; the prioritized path is
  docs/edge-roadmap.md (review-owned, evidence table at the top).
• Every model experiment runs through
  scripts/report_strategy_shift.py --phase <label> — pre-registered gates,
  committed results, no exceptions. The reviewer session reads only what's
  in reports/strategy_shift/.
• The UI is the shared state of truth for humans: /models (catalogue +
  verdicts), /strategy (edge thesis + gates), /ops (pipeline health),
  /dq (data quality), /glossary (vocabulary).
• Two-session pattern: the data-connected CLI session builds and runs;
  the reviewer session audits, gates, and sets direction. Handovers via
  committed briefs, never via memory.

Priorities right now (mirror of edge-roadmap.md — that file wins on detail)

1. open_fav_v1 — the favourite-side filtered strategy discovered by CLA
   segmentation (back favourites the models rate stronger, at open; 62.8–65%
   market confirmation, z ≥ 5.5). Backtest verdict first, then live paper
   trades every round. (docs/edge-review-2026-07-12.md N1)
2. Live evidence quota, every round — open snapshot, team lists ≤30min,
   both strategies seeded (open_fav_v1 + rapm_timing_v1), per-kickoff
   closes, CLV ≤24h. A missed item is a program incident. (N2)
3. line-v2 — state-space variance + real lineup deltas into the line model.
4. Demoted/killed: phase3-rapm-v4 idle-rounds only; cpois retired; H2H
   feature engineering frozen; divergence signal killed. (N3)

Standing rules (test-enforced where possible)

• Placement-time features only; closing data is a diagnostic label.
  (tests/test_leakage_guard.py)
• Walk-forward everything; actionable pricing at OPEN.
• No ROI without its season-bootstrap CI (≥5 seasons AND ≥100 bets to claim
  significance).
• Gates are pre-registered; never edited in the commit that reports results.
• Language: candidate / hypothesis / paper trade.

Key documents