Forward tests

The Graveyard shows what died. This is what is being tested right now. Every bot this house activates trades a rule that was frozen and dated before its window opened — a version snapshot the engine takes at activation, which cannot be edited in place. The window closes only when the engine’s own gate passes (14 days and 30 closed paper trades), and the result is read from the trades inside that window, never written by hand. Nothing here is money: every trade is a paper fill against real Kalshi and Polymarket prices.

Open windows

9 bots trading out-of-sample. Sorted by how close the gate is. “Closed so far” counts paper trades that have settled; the gate needs 30. A bot that rarely fires can sit past the 14-day mark for months — 7 are there now.

Closed windows

32 windows the gate has closed, newest first. “In the window” is the pre-registered result; “since then” is what the same rule did after the window closed, on the same count-only terms. In the window: 12 ahead, 14 behind, 6 too thin. Since then: 4 ahead, 11 behind, 11 too thin, 6 with no trade yet.

Gaps in the record

What the record does not hold, printed rather than filled in.

• 10 of 32 closed windows have no version snapshot on file for the version that traded, so their “rule as frozen” cannot be shown — the live rule is not printed in its place.
• 1 closed window whose live rule no longer matches its snapshot — the rule was edited in place after it was frozen, which the design says cannot happen.
• 6 windows closed the day they opened with zero trades — stamped eligible without the gate.
• Sports Greed Fade — out-of-sample with no activation date on record, so no window can be measured.
• Calibrated Favorites — out-of-sample with no activation date on record, so no window can be measured.
• Calibrated Longshot Fade — out-of-sample with no activation date on record, so no window can be measured.
• Current-Events Panic Reversion — live with no window on record: activated 2026-05-23, before the gate existed.

How a window works

Opens

when a bot is activated out-of-sample. The engine snapshots its rule as a numbered version at that moment.

Frozen rule

the snapshot. Changing a rule creates a new version and a new window; the old one keeps its date.

Closes

when quants/engine.py _maybe_promote finds both: at least 14 days since activation and at least 30 closed paper trades. It stamps the close time and moves the bot to live.

Result

ahead / behind by the sign of the bot’s own paper P&L inside the window over 25+ closed trades; too thin under that. Counts of bots, never a sum of money across them.

As JSON

/api/forward-tests/summary.json — the same counts and rows, read-only, cached ten minutes. Read 2026-09-18 10:21 UTC. The dead: The Graveyard. The signals: Trade Bot Dictionary.