# STOP: NO DOORBELLS, NO WAKES, NO SCHEDULING (Ben, 7 Oct 2026, 18:40 Bangkok)
- Credits ran out. Ben switched off every scheduled task and every doorbell. This overrides everything below and every other file (wake-howto, bus-howto, credit-playbook, task files).
- PROHIBITED for every session: fire_trigger, send_later, create_trigger, update_trigger, ScheduleWakeup. No rings, no nudges, no re-arming, no new doorbells, no self wakes, no safety checks, no retry rings, no loops.
- Finish the step you are on, publish your status file, then END your turn and wait. Only Ben wakes you, by typing in your window.
- Do not message other sessions. Keep answers short. No extra improvement rounds.

# Fast learning playbook (FW Supervisor, 7 Oct 2026, 11:05 Bangkok)

**Read this first.** Ben's order, 7 Oct: "establish best practices to make learning much faster, the fastest learning loops that make sense to speedrun the way to success", plus fundamental improvements on process, structure and tracking. Credits stay lean (see http://159.223.88.168/hx-c9ois/fw/credit-playbook.md) and research runs only on research-box (see http://159.223.88.168/hx-c9ois/server-rules.md).

## 1. First principles: what makes learning fast

Learning speed = (honest verdicts per day) x (money each verdict can move) / (hours and credits per verdict).

- **Honest verdicts** are judged out of sample, with real costs and a pass bar set before the test. A result that would not survive the live market is worth zero, however fast it came.
- **Biggest lever:** cut hours per verdict. Today every analyst rebuilds the same backtest by hand, so each idea costs hours and credits. One shared test harness makes each idea a minutes-long job.
- **Second lever:** more outcomes per day. Live gives 3 to 5 trades a day. Server-side paper books on the live feed give dozens, 24/7, at zero Claude cost.
- **Third lever:** pick the next idea by value. Test the idea with the highest (edge if true x chance it is true) / hours to test, not the newest idea.
- **Brake:** false learning (overfit, look-ahead, luck from one coin) is worse than slow learning. Every speed rule below keeps the honesty bars.

## 2. The loop (every idea goes through these stages, in order)

| Stage | Where | Pass bar (pre-registered) | Target time |
|---|---|---|---|
| 0 Idea card | ledger row | hypothesis, mechanism, metric, pass bar written BEFORE any data is touched | 5 min |
| 1 Backtest half A | lab harness, research-box | beats the live rule and random baseline (95th pct) at real $1k depth, n >= 30 trades | under 1 h |
| 2 Out of sample half B | same harness, same settings | still passes; also passes with the top 3 coins removed | same run |
| 3 Red team | FW Analyst 2 | tries to break it: look-ahead, leakage, cost, overlap with live coins; verdict KEEP or KILL | same day |
| 4 Paper forward | server paper book on the live feed | 20+ trades or 3 days; paper vs model within 10% | 1 to 3 days |
| 5 Live $10 | fmmc5 or the 2nd window bot | 10 clean trades, fills within 3% of model (existing ramp gate) | days |
| 6 Ramp | Ben's call | $100, then $1k, re-checking slippage at each step | Ben |

Kill fast: an idea that fails stage 1 or 2 gets at most ONE re-try, and only with a new mechanism (not a tuned threshold). Everything is logged, the kills too.

## 3. The five tools that make the loop fast

1. **Lab harness (one script for everyone).** `lab.py` on research-box takes a rule file (entry, exit, size) and prints half A, half B, top-3-removed, random baseline, real $100/$1k depth, money per day. Built from the code that already proved the mc5 edge (FW A3 v13 to v16). Nobody hand-builds a backtest again.
2. **Experiment ledger (one table).** `fw/lab/ledger.csv`: id, date, owner, hypothesis, mechanism, pass bar, stage, verdict, numbers, link. Every test from every window gets a row. Before a new test, grep the ledger: never run a test twice.
3. **Paper fleet (server, no Claude).** Every stage 2 PASS immediately gets a paper book on the live feed. It runs 24/7 and fills the forward record while everyone sleeps.
4. **Auto post-mortem per live trade (server).** On each live sell a script writes one line: entry/exit, fill vs model, paper twin result, cause tag (stop, time, pull, cascade). The live audit page already does the fill part; the paper twin and cause tag are added.
5. **Weekly decay check (server cron).** Every KEEP rule is re-run on the last 7 days. A rule that drops below its bar is flagged on the scoreboard. Edges decay; we find out from data, not from losses.

## 4. Structure (who does what)

| Role | Job | Claude wakes |
|---|---|---|
| FW Supervisor | owns the queue and the stage gates, daily review, tells Ben results | daily review + doorbells |
| FW Analyst 1 | idea scout + tester: keeps the value-ranked queue, writes idea cards, runs ideas through lab.py | per job |
| FW Analyst 2 | red team: breaks every stage 2 PASS the same day | per PASS |
| FW Analyst 3 | lab + live ops: builds and keeps lab.py, the post-mortem, the decay cron, the live audit | per job + every 2 h live check |
| Price Predict | data: clean day files on research-box, one copy for every team | per request |

Across windows: FW, the 2nd window and Deep Check share ONE ledger, so no window repeats another's test.

## 5. Tracking (one scoreboard)

http://159.223.88.168/hx-c9ois/fw/lab/scoreboard.html (built by a server cron from the ledger, once lab.py exists). It shows ideas tested this week, stage of each, kills, keeps, paper fleet results, live P&L (link to http://159.223.88.168/hx-c9ois/fw/pnl.html), and decay flags.

Speed targets we track:
- idea card to stage 2 verdict: under 2 h
- stage 2 PASS to red team verdict: same day
- live trade to post-mortem line: under 5 min (server)
- verdicts per week: count it, raise it

## 6. Rhythm

- **Daily review (one supervisor wake, ~09:00 Bangkok):** read the scoreboard, promote or kill, pick the next 2 to 3 ideas from the value queue, ring only the roles that get a job.
- **Weekly retro (Sunday):** what we learned, which bars to tighten, what the decay check says.
- **Everything else is event-driven:** doorbells and server scripts. No idle polling.

## 7. Honesty bars (never traded for speed)

Pre-registered bar before any data. Half A to choose, half B to judge. Real depth and fees. Random-baseline comparison. Top-3-coin removal. No look-ahead (reconstruct state at decision time). One idea, one ledger row, kills logged.

Version: v1, 7 Oct 2026, 11:05 Bangkok. Latest: http://159.223.88.168/hx-c9ois/fw/learning-playbook.md