BreakoutBot

PAPER

The AI is not live — the numbers below are frozen as of 29 Sep 2026

The first run (25–29 Sep) is void: at every review the AI was wrongly told that five positions were open, while this bot held none (once, one). The bug was in our code, not the model, and every veto below cited that phantom exposure. Since 1 Oct this bot runs as a control with the AI switched off, and it has matched the bot without AI trade for trade. On 13 Oct we tested three Claude models offline over two periods after their training data ends. None passed, so the AI stays off until a new design does. The full results are on the start page.

balance —

worst drawdown to date —

expectancy / position · all-in —

Risk discipline

fixed-dollar risk, hard circuit-breakers

Equity & drawdown

Month by month

closing balance vs the month before

Breakdowns

same positions, sliced four ways

Signal funnel

how many candidates survive each filter

Risk circuit-breakers

proof the guardrails actually fire

Every position

Closed	Symbol	Sleeve	Regime	Risk	Notional	Exit	PnL	Balance	Why

Backtest validation

controlled drawdown even in deep bear

Methodology & honest disclaimers

With AI — what differs. This bot is a twin of the one on the Without AI page: same code, same settings, split from the same account at the moment shown above. The designed difference: before each full entry, Claude Opus 5.5 reviews the setup (candles on three timeframes, indicators computed in code, the bot's own stop and targets, the positions it holds) and may veto it. Each bot fetches its own prices, so the two can also drift apart without the AI; that drift is counted separately. The AI can only remove entries, never add one, and any failure — timeout, outage, budget — lets the entry through. History before the split is shared by both pages.
An experiment, not a result. One account path is not evidence. Whether the reviewer helps is decided by a test written down before it started (checked at 100 and 300 closed positions), not by which page is ahead today.
Exit change on 12 Oct 2026, 21:21 UTC — both bots. Once a trade reaches its first target, its trailing stop no longer drops below the entry price. Before, a trade could touch the first target and still close at a loss. In backtests the change helped slightly in both test windows (about +0.02R per trade, well inside the noise), so we adopted it as a small, low-risk fix, not a proven gain. Results before and after this moment come from slightly different exits.
This is system v2, tracked since 2026-07-29. Everything above covers the current bot only (regime-aware breakout + mean-reversion, wider trailing exit, curated 5-coin universe). The previous system v1 ran 2026-06-15 → 2026-08-22 and was retired at $896.38, −10.4% from $1,000; its raw log is archived and is not blended into any number here. Retiring a losing system and restarting the count flatters whatever survives — so v1's closing result is stated rather than quietly dropped.
Simulation — no orders, no real money. This bot places no exchange orders at all. Prices are real Binance futures data; entries, fills and balances are simulated locally. An earlier, now-retired system did route real Binance Futures Testnet orders — this one does not, and the page previously mislabelled itself “TESTNET” because of it. Transparency record, not a live audited fund.
Scoped to the current epoch. The record starts at the bot's last fresh start and nothing before it is counted. Until 2026-08-24 the exporter silently parsed the whole log, so a retired June run and a six-week outage were blended into the published curve; that is fixed, and the figures moved (44 positions rather than 46, win rate 40.9% rather than 41.3%).
No return promises. Nothing here predicts or guarantees future profit. Past results — including the backtest — do not imply future results.
Not financial advice. This is a tool/analytics record, not a recommendation to buy, sell, or hold anything.
What the edge actually is: conventional technical analysis + a 4h regime filter (longs only in BULL) + mean-reversion in range + fixed-dollar risk sizing and hard drawdown circuit-breakers. No secret indicators, no exotic alpha claims.
Counted per position, not per fill. When a trade scales out (half at the first target, the rest on the trailing stop) that is one position here, not two wins. Counting fills would inflate the win rate by roughly 14 points — this page reports the lower, honest number, and shows the win rate the system would need just to break even.
Verify it yourself: every number on this page is reconstructable from the bot's raw append-only log.
Benchmark methodology: "own-universe HODL" is an equal-weight buy & hold of the bot's own coin universe in USD terms since the current fresh-start epoch (Binance futures 1h closes). It exists to keep comparisons honest — a benchmark in a depreciating fiat currency (e.g. a ₺-denominated chart) can make "the market" look up when in USD it is not. It is chain-linked across curation changes: each period is priced with the universe that was actually live then, and the periods are compounded. Applying today's list to the whole epoch would restate the past every time a coin is dropped — the segments below are the ones being compounded.