An honest null result

A crypto paper-trading agent built spec-first. It trades no real money, and the language model is kept out of the order path. A framework-free core runs the same indicators, strategies and backtester in backtests and live, behind a NestJS agent with a risk engine, a pessimistic paper broker and a kill switch. The report leads with the result: none of 34 configurations cleared the bar.

• A frozen specification

  The 721-line spec doesn't change; every departure goes into an append-only deviation log and errata.

  Trade-off Changing course costs paperwork, on purpose.

• Experiments registered before they run

  Parameter grids are written down with a trial budget and a deflated Sharpe ratio before any backtest, so luck can't pass for skill.

  Trade-off Fewer, slower experiments.

• The model stays out of the order path

  A later language-model analyst is for observability only, shipped switched off behind a typed guardrail.

  Trade-off No AI decisions to show off, and no AI money risk either.

• Report the null result first

  The report opens with what didn't work, including a fee analysis that moved execution to a cheaper venue.

  Trade-off A less flattering story than cherry-picked returns.