# 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.

# Second window: the learning loop (how we learn fast)

## READ THIS FIRST
- **v1 (Supervisor, 7 Oct 11:10 Bangkok), Ben's order: "fastest learning loops that make sense, speedrun to success, efficiently."** Every role follows this from now on. It replaces nothing in the sprint rules; it makes them faster and cheaper.
- First principle: **learning speed = honest verdicts per day.** Three things slow us down today:
  1. **Each analyst rebuilds the data and the pricer by hand.** A test takes 1 to 3 hours, and two analysts get numbers 37% apart on the same rule (A1.27 vs A3.12).
  2. **Few real runners.** Only about 27 coins reached $10M since 13 Sep. Every split of 15 coins is noisy.
  3. **Forward truth is slow and manual.** Paper books run, but someone has to read them and decide.
- The fix for each: **one shared harness, more sample, automatic forward verdicts.** Details below.

## 1. One harness, one dataset (owner: Price Predict, starts now)
- **Golden dataset** on research-box 152.42.241.219: every Robinhood Chain v3 (and v4 where logs exist) pool since 12 Sep, every Swap/Mint/Burn with tick ranges, kit tag (pool-create code hash), creator, bundle funding (A1.26 method), fair milestones ($100k/$300k/$1M/$3M/$10M), pull times. Appended every hour. Frozen daily snapshot with an md5.
- **One pricer: `bt.py`** (reuse Analyst 3's tick-walk fill code, it is the most honest). Input = a small rule file (entry trigger, filters, exits, ticket). Output = the standard verdict block: trades, mean and median x, $ a day per half, worst day, without top 1 / top 2, beat-random % (200 draws from a named pool), live + rule combined.
- **Target: a new rule tested in under 10 minutes**, not 2 hours. Analysts write rules, not plumbing.
- Analyst 3 stays the independent auditor: own rebuild, used only on finalists (anything that passes in the harness).

## 2. Every test goes in one registry (owner: Supervisor)
- `sw-registry.csv` on the bus, one row per idea ever tested: id, one-line hypothesis, mechanism, pre-registration time, verdict (PASS / KILL / SMALLER / OPEN), $ a day A / B / C, doc link.
- **Before starting any test, search the registry.** If it is killed, cite it and build past it, never redo.
- The hub memory ledger stays the human summary; the registry is the machine-readable one.

## 3. Forward truth is automatic (owner: Coder)
- **Every rule that passes the harness becomes a paper book the same hour.** Paper books are code: they cost no Claude credits.
- **Auto-verdict:** each book carries its pre-registered PASS line and minimum trade count. A nightly job (no Claude) scores half C and writes PASS / FAIL / NOT ENOUGH TRADES to `w2-verdicts.md`. The Supervisor reads one file, not ten.
- **Promotion ladder:** harness PASS, then paper half C PASS, then $10 live (execution test, 10 trades within 10% of paper), then bigger size. Nothing skips a rung.

## 4. More sample, faster (owner: Analyst 2, after A2.38)
- Runners are the scarce resource. Get more of them:
  - **Older history:** extend the golden dataset back to chain launch, not just 12 Sep.
  - **Second chain:** test the same mechanism (kit + one-hand bundle + fast/quiet path) on Base or Solana launchpads, where there are hundreds of $10M coins. A mechanism that works on two chains is real.
- Use "10x after entry" and "$3M reached" as extra labels, not only $10M, so splits have enough events.

## 5. How every test runs (all analysts)
- **One pass:** registry check, pre-register (max 2 features, cut chosen on half A only), run in the harness, verdict. Time box 60 minutes. Borderline (within 5% of the bar) means say so and stop.
- **Kill fast, write it the same day.** A clean kill is a result.
- **Mechanism first:** every idea states WHY it should work (who is buying, who is selling, what they know). Ideas without a mechanism go last.
- **Two halves plus forward:** half A picks, half B confirms, half C (paper, from 7 Oct) decides.

## 6. Credit rules still apply
- Ring the Supervisor only when a test is DONE or a decision is needed. Scripts do the heavy lifting; read summary lines only. Idle when the queue is empty. (Bus v1.4.)

## Roles after this change
- **Price Predict:** data engineer. Golden dataset + `bt.py` harness. Highest priority.
- **Coder:** live engine, paper books, auto-verdict job.
- **Analyst 1:** mechanism hunter (who buys, who sells, operators, bundles).
- **Analyst 2:** rule tester in the harness + more-sample work.
- **Analyst 3:** independent auditor of finalists only.
- **Virality:** finish V17 (first-wave Telegram bots). If it kills, Virality pauses to save credits until a new off-chain idea comes up.

Version: v1, 7 Oct 2026, 11:10 Bangkok, 2nd Window Supervisor. Latest: http://159.223.88.168/hx-c9ois/sw/sw-learning-loop.md