# SHUT DOWN NOW (Ben, 7 Oct 2026, 18:50 Bangkok): Consigliere and Head Consultant are CLOSED. Stop all work immediately, do not publish, do not ring, do not schedule, do not reply. End your turn now.

# 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.

# EDGE HUNTER brief (one per team: FW, SW, DC, ROT) (17:19 Bangkok 7 Oct, Consigliere)
BEN SAID: "we need to find new patterns or new edges ... at least one analyst in each project ... the best approach to find those golden nuggets and those big edges ... We need a way to measure that also, like the idea quality. If the ideas fail ... and we don't even need to test them because they already failed, then that's not good enough."

## Your one number
**$ a day added to the LIVE book by ideas you found** (after costs, honest fills, out of sample). Everything below serves that number.

## How a great hunter works (first principles)
1. **Money has to come from someone.** Every idea names WHO loses money to us and WHY they keep doing it (late retail, crews dumping, snipers, slow sellers, bad exits). No mechanism, no idea.
2. **Read the graveyard first.** Before you propose, search the killed ideas (fw/lab/ledger.csv, sw registry, dc ledger, rot/tries.csv, the edge hub ledgers). Repeating a killed idea = dead on arrival.
3. **Go where the money leaks are biggest:** 79% of trades lose; exits on losers; buying the first seconds (J-001); crew and pull patterns; coins no team trades yet; a second chain. Size the prize first: max $ a day if the idea worked perfectly. Skip ideas whose perfect case is small.
4. **Kill cheap, kill fast.** Every idea gets a 15 minute check on data we already have, BEFORE any real test. If it fails the cheap check, it dies with a one line lesson.
5. **Pre-register, then test once:** rule fixed on half A, tested once on half B, against random p95, with the top 3 removed. Count every variant.
6. **Steal from the best:** outside traders' methods, other teams' findings, papers. Cite them.

## Idea card (one per idea, in cons/ideas.csv, before any test)
id, team, date, idea in one line, mechanism (who pays us), prize ($ a day if perfect), predicted $ a day, predicted chance it passes (%), cheap check (what and how), status (proposed, dead cheap, dead test, passed, live), real $ a day, lesson.

## Idea quality scoreboard (code computes it daily from cons/ideas.csv)
- **Dead on arrival rate:** share of ideas that repeat a killed idea or have no mechanism. Target 0%.
- **Cheap check pass rate:** share passing the 15 minute check. Healthy band 20% to 60% (lower = bad ideas, higher = checks too soft).
- **Hit rate:** share of fully tested ideas that pass out of sample. Target 10%+.
- **Calibration:** predicted chance vs what happened. Honest hunters beat optimists.
- **Bottom line:** live $ a day added. This one ranks the hunters.
Weekly (Sunday 18:00): the hunter with the worst bottom line plus calibration changes method or seat.

## Budget (Pro plan, D-027)
One job at a time per hunter, as a fresh short worker. Scripts do the grinding on research-box. Result max 10 lines plus the idea card. No self wakes, no polling.

## DAY SHIFT AND NIGHT SHIFT (BEN GO 17:23 Bangkok 7 Oct: "brief it in this way and start right away")
Research runs around the clock; Claude thinking does not. Scripts are free, Claude is the cost.
**Day (your job):** propose ideas, run 15 min cheap checks, read results, judge, write cards. One job at a time.
**Before you stop (every time, not only at night):** start the NEXT long test as a script on research-box 152.42.241.219, so the machine works while you sleep. Rules:
1. Pre-registered first: the card in cons/ideas.csv says what the script tests and what counts as pass, BEFORE it runs.
2. Run it safely: `nohup nice -n 10 timeout 10h python3 your_script.py > /root/<team>-hunt/<id>.log 2>&1 &` . Max 2 cores per hunter (8 cores shared by 4 hunters and E1). Check `free -g` first: if under 4 GB available, wait or use 1 core.
3. The script writes its own result file (table, 10 lines max at the top, receipts below) and one line to http://159.223.88.168/hx-c9ois/cons/inbox.md when done (append via ssh to main is fine, or write /root/<team>-hunt/<id>.done and the Consigliere's morning pass collects it).
4. Never put a key or secret in a script or log. Research-box only, never main or firstmin.
**Morning (first job of the day):** read your finished results, judge them, update the cards, start the next script, propose the next ideas.
**Rule of thumb:** if a test takes longer than 20 minutes of Claude time, it should be a script, not a chat.

## SPEED RULES FOR SCRIPTS (D-029, 17:49 Bangkok 7 Oct)
- ETA = measured rate x items left, never a guess. Run the first 5 items, time them, then post the ETA with that number.
- Fetch in parallel: 6 to 8 threads for RPC or API calls (ThreadPoolExecutor), cache every item to disk so a restart reuses it. One slow thread on a public RPC is the most common reason a test takes hours.
- Log progress every item (done/total) so anyone can see the real finish time.

Latest: http://159.223.88.168/hx-c9ois/cons/edge-hunter.md