Repository instructions

This repository evaluates semantic coordination mechanisms. Treat research validity as a product requirement.

Non-negotiable experiment rules

Never describe deterministic harness output as evidence that Sema improves model performance.
Keep pattern content identical between equal-information prose and Sema conditions. Isolate content, addressing, and enforcement effects.
Choose and document one primary endpoint before a confirmatory run.
Preserve every run, including failures, timeouts, and malformed outputs.
Record prompt, dataset, model, implementation, Sema version, vocabulary root, scorer, and protocol fingerprints in result manifests.
Report wire bytes, hydration/context tokens, and total model tokens separately. A short reference is not automatically a context-token saving.
Prefer deterministic validators. Blind subjective judges to the experiment condition and never use an LLM judge as the only scorer.
Do not use live trading capital in an experiment. Historical replay and paper execution come first.

Use pnpm and Node.js 22 or newer.
Keep TypeScript strict and avoid any.
Put reusable code in packages/; experiment-specific policy belongs under experiments/.
Generated result bundles belong under results/ and remain untracked unless intentionally promoted into a dated public report.
Run pnpm check before handing off changes.