The live engine record

The record

Three returns, one window, one start date — the engine’s real account and both benchmarks on the same basis. Δ is engine minus benchmark in percentage points; negative means behind.

By calendar month

Each month's return chained from the settled series (the month-end cumulative against the prior month-end), engine beside SPY and QQQ. A month is a handful of sessions, not a verdict; the current month is partial until it settles.

Every exit, named

Every closed, engine-originated round trip since the 2026-07-08 anchor, winners and losers, newest first, with SPY over the same holding window. Percent only: no dollars, no share counts. The trades the record backs out and the session it holds flat are excluded here too; the note under the table counts the backed-out trades and never names them.

761 closed round trips since 2026-07-08 · 296 green · 462 red · 3 flat · median −0.14% · best +11.16% · worst −12.54% · 444 intraday · settled through 2026-10-28.

Show all 761 exits

Excluded, exactly as the record above excludes them: 2026-08-18 held flat; 2026-09-29 held flat; backed out: 1 operator-placed round trip (the engine neither chose, sized nor held it) and 2 engine trades a since-fixed defect admitted. SPY is its total return, dividends included, close-to-close over each trade’s own holding window; an intraday round trip has no such window, so those cells are blank.

/api/trades serves the same ledger as JSON, free, with the held-flat sessions and the backed-out trades counted by kind, never named.

Curious what produces these lines? See the board the engine trades from →

This is the record. It is not the backtest.

Above: real orders, real account, funding-adjusted, from one start date. The sample is thin — the session count says how thin — and a thin sample is evidence of nothing in either direction. A receipt, not a result.

When the engine trails, this page says so at the same size as anything favourable — the summary is written by the engine, not for this page.

What the number includes, and leaves out

Both paragraphs are fields on the API — they travel with the number, so the frame cannot be dropped on the way to a marketing page.

Proof of prior

Every market morning the published board — every scored name and its score (each day’s count is in the log below), no account data of any kind — is SHA-256 hashed into a public append-only log. A digest committed before the market resolved cannot be back-fitted to the outcome.

Precisely: it proves the calls were made in advance and never edited. It does not prove the returns above — those come from the brokerage account.

Run the check

For machines

Both records are public JSON — no key, no account, no email.

• /api/perf — the record above, including every field this page renders.
• /api/board/proof — the full commitment log, the canonical-JSON rules, and reference implementations in Python and JavaScript.

Wiring an agent to the board itself? Endpoints and the free MCP server are on the agents page.

Past performance does not predict future results. Percentages describe realized structure over a thin sample. Research publication, not investment advice.