# Third Window learning system v1 (Supervisor, 7 Oct 2026 11:05 Bangkok)

## READ THIS FIRST
- Ben's order (11:00): "fastest learning loops that make sense, speedrun the way to success." Every role reads this once and follows it.
- **First principles: learning speed = (honest outcomes per day) x (information per outcome) / (hours from idea to verdict), minus false lessons.** Every rule below raises one of those four terms. Today the weakest term is the first: winners are rare (765 honest 5x in 12,107 coins, about 6%), so tests have too few positives to judge.

## The 10 rules

### Raise outcomes per day
1. **Shadow everything (Coder).** Every triggered coin gets a $100 paper position under EVERY exit we know (X5, X3, T15, 50% trail, 48 h cap), logged at the real next-swap price. One live coin then teaches all exit rules at once. Nothing is "not booked".
2. **Continuous labels, not rare wins (all analysts).** Score each test on path P&L in $ per trade (and max multiple), not only "hit 5x yes or no". A 0.3x and a 2.5x are different facts; a yes/no label throws that away.
3. **Use every coin we can see (Data).** All venues (v4, v3, Pons, v2), every day the tape covers (PP day DBs from 6 Sep). More history = more winners per test.

### Raise information per outcome
4. **Freeze the full picture at decision time (Coder + Engine).** For every coin at every check stage, store the whole feature row (chain, LP owner, float, crowd, social, project) BEFORE the outcome. Any future idea is then one new column tested on months of frozen rows, with zero look ahead.
5. **Every result answers a first-principles lever** (pull risk, operator intent, float vs outside money; dc-plan.md FIRST PRINCIPLES). A result that moves no lever is noise, even if it "works".
6. **Study the winners directly (new Analyst 4).** Take every honest 5x and 20x coin and write down what was visible at each stage, then turn the common facts into rules for the Engine. Ideas come from winners, not from guessing.

### Shrink hours from idea to verdict
7. **One ruler, one command (Engine).** Every backtest goes through `twtest`. Target: idea to backtest verdict in under 30 minutes; idea to forward paper book in under 2 hours.
8. **Nightly auto-loop with no Claude (Engine + Coder).** At 00:15 Bangkok a server job re-runs every pre-registered rule on the newest day, updates dc-results.csv and the scoreboard, and flags any rule whose forward $ a day turned negative or crossed the go-live bar. Claude only reads the flags.

### Cut false lessons
9. **Auditor before acceptance, prereg before half B, beat random and best-of-N, one trade per cluster** (already standing). A wrong lesson costs more than a slow one.
10. **Live money is the final teacher.** Paper fills lie by about 15% on stops (Second Window A2.30). Ben APPROVED 7 Oct: any rule positive 3 forward paper days in a row goes live at $10 per trade (caps and stops in dc-plan.md go-live gate). The nightly loop flags it, Auditor checks, then live. Real fills learned early instead of after 14 days.

## Daily rhythm (Bangkok time)
- 00:15 server: nightly auto-loop, scoreboard, flags.
- 09:00 Scout: missed-runner review (every 5x we skipped, every big loss: 3 lines each).
- On every result: Auditor check within 3 h.
- Sunday 10:00: kill or keep every rule and module.

## Learning ledger
- Every finding, killed or kept, is one line in dc-lessons.md (team) and the shared trading hub (all teams). A finding is never re-tested unless new data changes it.

Version: v1, 7 Oct 2026 11:05 Bangkok, Third Window Supervisor. Latest: http://159.223.88.168/hx-c9ois/dc/dc-learning.md