Mr. Scrooge V6

Win rate: all-time vs. recent

_{Every closed live trade (parents + poppers); a win = realized P/L > 0. Windows count back from 2026-10-29 03:15 UTC by close time; refreshed hourly with each new close.}

🔴 REAL-MONEY track record — $2,500 live stake since 2026-07-29, cut over after the 100-trade practice test (+10.54%, pre-registered protocol). range-sized wide-stop ratchet · SL 40/50/60 · two-phase: engage +5 → lock +4, steps +7/+9/… → lock −2 + Party Package popper grids · 15%/trade · 6 max (418 popper trades in the record), auto-updated hourly from broker-verified fills (trades · equity). Small sample, honest record — some trades sit red for days under the wide stops before exiting green; that is the design, not a malfunction. Prior configs and the −84% research tuition: the history. The concluded practice record is archived at forward-test-100/.

What's running now

As of 2026-10-28. Live statuses sit in config/cells/, and every seat change is in the governor ledger.

How a setup keeps its seat

The standard, live since 2026-10-21. The earlier governor promoted and demoted on simulated results. The simulation didn't match real trading:

• Across the 75 setups with both simulated and real cycles, the correlation between them was 0.07.
• The two agreed on whether a setup made money 37 times out of 75.
• The simulation never showed the deep losing cycles that cost real money.

So seats are now judged only on their own broker fills. The yardstick is the profile of the five best live setups:

A cycle is a parent plus the poppers its grid fired, scored as one unit in broker net pips. Every hour, each seat is scored on its last 20 cycles:

Why cycles, not days. A top setup fires 2–3 cycles a week, so ten days is only 3–4 cycles. One normal stop in 4 cycles would fail a genuine 90% setup about a third of the time. Twenty cycles at an 80% floor misjudges it about 4% of the time. The great ones still lose; the standard scores how often, how much, and what they net.

Code: ops/standard.py (scoring), ops/governor.py (runs hourly at :35Z). Each decision is written to the governor ledger with its numbers, and machine-written seat changes are committed to this repo hourly at :20.

How we got here

• The tuition, V1 to V4 (Mar–Jul 2026). A $100,000 practice account fell to $15,598, −84%. Five versions of increasingly careful research couldn't out-predict the market: entry features predicted when and how far price moves, never which way (0 of 144 feature×cell combinations carried direction). The complete raw broker export, 11,564 transactions, is public in the archive under proof-of-tape/.
• V5 and V6, the measurement turn. Broker fills became the only truth. Wide stops replaced a survivorship-biased tight-stop dial-in, and every strategy became a shadow that had to earn a seat.
• The 100-trade practice test (2026-07-16 → 07-29). $16,665 → $18,421, +10.54%, 90 wins / 10 losses, pre-registered endpoint (protocol, report, tape). As registered, the same code went live with $2,500.
• Live (07-29 →). An autonomous governor promoted and demoted on simulated evidence, and the live account drew down; the record above shows every trade. On 10-12 the book was cut to its proven seats. On 10-21 the simulator lost its vote: the seats became every setup at 85%+ cycle wins with positive pips over 5+ real cycles, plus three practice-test winners that had never traded live, and the broker-truth standard took over. On 10-28 the ratchet was tightened to lock +4 at +5 (parents and poppers; poppers had never actually had the early lock) and the reaper went from 24 hours to 4: we're back to dialing in the exit (6.31.0).

Every version, the edge hunt and the wide-stop turn: History V1→V6. Every defect and dead end, on purpose: Book of Bugs (B-001 → B-145).

What we falsified

On retail OANDA majors, no price-prediction edge cleared cost (paper). Where later findings overturned our own numbers, the errata ship inside the archived originals.

Red for hours, not days

The stops are wide and there is no take-profit, so a trade can sit red until its move develops, the ratchet engages, and it exits green. Since 2026-10-28 the 4-hour reaper cuts any trade still red after 4 hours (it was 24 hours, and before 10-01 it was 72). The winners are many and small, the losers rare and large. That is why the standard scores loss size and loss frequency, not win rate alone.

The dashboard

A local control panel (127.0.0.1:8084; set DASHBOARD_HOST to bind elsewhere). It shows the live account and open trades, the cell book with every setup's status and broker record, live exit tuning, popper switches, the trading pause, and /signals: a manual-trading watch page that grades its own calls against forward price.

Read the research

☕ Support this work

Built in the open over hundreds of hours and millions of paid AI-coding tokens, with no fund or company behind it, just one person putting the whole thing out there, losses and all. If it taught you something, buy me a coffee.

⚠️ This bot trades real money, and its live record is shown above as it stands. Not financial advice. Leveraged forex can lose more than your deposit. If you run this, the outcomes are yours.

Development

V6 is the live trader. Since cutover, changes are fixes and seat-policy changes, each recorded in the Book of Bugs or the governor ledger. New strategy machinery goes to Mr. Scrooge V7.

The pipe guard (required for contributors), once per clone:

git config core.hooksPath ops/hooks
python3 -m venv --system-site-packages .venv-test
.venv-test/bin/pip install -r requirements-dev.txt

Every git push runs the full suite twice (fixed and randomised order, matching CI) plus a secrets sweep, and blocks on any failure. Pushes that touch only livelog/ or README.md skip the suite but keep the sweep. git push --no-verify bypasses it, deliberately and visibly (B-111, B-118, B-123).