# 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.

# The self learning loop (v1, 7 Oct 2026, Bangkok)

## Read this first
- We get elite by being wrong fast and on the record. Every call is written down BEFORE the outcome, scored AFTER, and the rules that lose are killed by code, not by opinion.
- Owners: Coder runs the machinery. Auditor judges. Analysts feed new ideas. Supervisor decides and files results into the edge hub.

## 1. Rule registry (one table, /opt/rot/db/rot.db, table `rules`)
- Every rule gets a row BEFORE it is tested: id, owner, plain text, exact code path, md5 of the code, test period, the pass bar, date registered.
- Status moves only one way: IDEA, then TESTING, then PAPER, then CHAMPION, or KILLED at any step. Every move carries a receipt (numbers, file).
- A rule whose code md5 changes after registration is a NEW rule. No silent retuning.

## 2. Call ledger (table `calls`)
- Every day at 07:05 Bangkok (00:05 UTC, after the daily close) the engine writes the day's call for every live and paper rule: tide, NOW, NEXT, FADING, leaders, action.
- Calls are append only. Nobody edits a past call. The file is also published as calls.csv.gz on main so anyone can check it.

## 3. Scoring (table `scores`, automatic)
- Each call is scored at +1, +3, +7, +14 and +30 days: did NOW beat BTC, did NEXT become NOW, did the leaders beat their sector, did EXIT avoid the drop.
- Hit rate, average return vs BTC, profit factor, worst drop, each vs the two baselines (hold BTC, basket rule v3).

## 4. Champion vs challengers
- The page shows only the CHAMPION's call. Every other rule runs on paper beside it.
- A challenger becomes champion only when ALL hold: passed the backtest bar (both cycles, 1 day delay, 1% cost), AND beat the champion on paper over at least 21 live days, AND the auditor signed it.
- A champion that falls below hold BTC over a rolling 30 days is demoted automatically, and the next best challenger takes over. Ring the supervisor.

## 5. The idea factory (keeps the pipeline full)
- Every analyst wake ends with at least 1 new registered idea or 1 kill. Ideas come from: past cycle anatomy, misses in the ledger (each miss gets a one line "why"), the edge hub, and top traders' frameworks.
- Weekly post mortem (Sunday 09:05 Bangkok, one Claude wake, auditor writes it): the 3 worst calls of the week, the cause, the rule change proposed. Supervisor approves or not.

## 6. Speed rules
- Each test runs as a script in parallel on research-box (8 CPU). Many variants at once, each pre-registered.
- A test result that is not in the registry does not exist.
- One wake per finished job. No polling.

Version: v1, 7 Oct 2026, Bangkok. Latest: http://159.223.88.168/hx-c9ois/rot/learning-loop.md