Darwin Lab

Darwin Lab

Darwin Lab is Darwin's synthetic usability-testing workspace. It gives a bounded population of independent, inexpensive AI agents one fixed goal and lets each agent operate ProjectFlow through the visible interface in a fresh Playwright browser context.

Synthetic usability testing complements human research. It does not replace it.

Evidence boundary

Lab interactions are real semantic telemetry emitted by ProjectFlow, but the actor provenance is always synthetic. Lab studies use experiment-scoped study IDs and are never included in human cohorts, measured fitness, or the 10,000-event scale replay.

Agents do not receive the hidden answer oracle, source code, private APIs, database access, selectors, or a correct navigation path. Each run has an action budget, duration limit, repeated-state limit, and isolated browser storage.

Run an experiment

1. Start Darwin and ProjectFlow locally.
2. Open Darwin Lab in the Darwin navigation.
3. Confirm the target is an allowed local, test, preview, or staging origin.
4. Create the fixed Project Apollo discovery experiment.
5. Queue the population.
6. Run npm run lab:runner from the Darwin repository.
7. Inspect the live population, run replay, task outcomes, and linked L-EV-* evidence.
8. With live reasoning configured, request the single population-level GPT-5.6 analysis and approve a bounded implementation brief.
9. Promote a recurring failure to a retained BE-### behavioural eval. The eval becomes an outcome-based acceptance contract for future Codex changes.

The default population uses eight agents. Each per-action decision uses the model configured by OPENAI_LAB_AGENT_MODEL; population analysis uses OPENAI_MODEL. Both integrations fail closed when live reasoning is unavailable.

Safety boundary

DARWIN_LAB_ALLOWED_ORIGINS is an explicit allowlist. Production targets are not enabled by default. Darwin Lab must only run against local development, dedicated test environments, disposable previews, or explicitly approved staging systems.

The runner supports only validated, user-facing browser actions. A separate oracle scores submitted answers as success, partial success, incorrect, abandoned, timed out, blocked, or system error.

Evidence and mutation flow

bounded synthetic population
  -> isolated browser actions
  -> synthetic-provenance semantic telemetry
  -> deterministic friction detectors
  -> hashed Lab evidence pack
  -> one evidence-citing GPT-5.6 analysis
  -> human-selected implementation brief
  -> retained behavioural eval (optional)

Behavioural CI

A behavioural eval is not a prescribed click path. It records the user goal, the hidden oracle boundary, pass criteria, forbidden outcomes, action budget, seed, target snapshot, and supporting L-EV-* IDs. Its Codex brief says to make the eval pass without changing the oracle, thresholds, seed, telemetry provenance, or protected paths. This turns an observed synthetic failure into an executable acceptance test that can be rerun after a candidate mutation.

The retained eval is available through GET /api/behavioural-evals and is created with POST /api/lab/experiments/:experimentId/promote-eval.

Predicted impact remains a hypothesis. Before-and-after fitness is measured only after an equivalent population reruns the same task, fixture, budgets, and persona distribution against an evolved preview.

See API Reference for the Lab endpoints and Getting Started for environment configuration.