A research instrument, not a product. It trades a paper account, grades its own strategies against broker-confirmed fills, retires the ones that do not earn their place, and publishes every claim it makes with the command that produced it.
Every figure here is published from docs/site/figures.json, and each names the command that produced it. There are no screenshots: a reader can re-run a command and cannot re-derive an image.
- agent, after cost
- UNKNOWN
- vs SPY holding
- UNKNOWN
- trading days on record
- 27
- gate
- 59 classes
allocation and aligned rolling equity history unavailable
aligned portfolio comparison unavailable
against a 60-day target
2027 tests across 116 files
What it has shown
- The accounting is real: broker-confirmed fills, a FIFO lot ledger, and a Guardian whose limits cannot be weakened by any configuration value.
- The gate can fail, and is made to fail on purpose to prove it.
make verify-self-testplants the defects each class exists to catch, including a NaN limit that silently stops comparing. - Research requires a registered contract and validated real bars before evaluating candidates and recording trials. Scheduled operation requires a configured contract, usable cache and verified successful runs.
What it has not shown
- That it beats holding. UNKNOWN after cost against SPY UNKNOWN over 27 of 60 recorded dates. Allocated capital and aligned equity are unavailable.
- Any directional edge. 129 hypotheses appear in the historical signal ledger. These aggregate verdict counts do not establish validated alpha or certify statistical error control.
- Live trading. Not present.
livemode is refused by the configuration loader.