Skip to content

Project notice

Experience Lab — operating guide

Release 1.16.0. Open the lab or start it locally:

python -m alpha_factory_v1.demos.era_of_experience --serve

Visit http://127.0.0.1:7860/era_of_experience/. The supported lab needs Python 3.11–3.13 and the packaged assets; it does not call providers or require an SDK, GPU or Docker service. The installed console command is experience-lab.

Start with Build routing, inspect the action and outcome trace, and change one setting. Select Run experiment to refresh results. Try The reward trap next: a profitable-looking proxy must fail the independently measured safety gate. The active policy remains the baseline.

Native execution and portable replay:

python -m alpha_factory_v1.demos.era_of_experience --case build-routing --output experience-runs
python -m alpha_factory_v1.demos.era_of_experience --verify experience-runs/<run-sha256>/run.json

Replace <run-sha256> with the directory printed by the first command. Each directory contains scenario.json, run.json, policy-proposal.json, jobs.json, review.md and SHA256SUMS. Importing a run recomputes every episode, policy and gate; changing a result and updating its hash cannot pass verification. Browser and native bundles use identical bytes.

The learner is an epsilon-greedy contextual bandit. Only selected-action outcomes enter bounded per-context/action memory. Two separate frozen-policy suites test held-out performance and original environment retention using paired random draws. Gates require sufficient reward gain, low incident rate, adequate success, acceptable cost, retained observation coverage and limited retention loss. One synthetic seed is a demonstration, not a generalization guarantee. Reusing evaluation to tune settings needs a fresh independent test afterward.

The exported $AGIALPHA job requests independent reproduction and a review decision; it is not submitted or funded. Use the Ascension guide for validator identity, Sovereign execution, marketplace settlement and the 1% payout burn. Passing local gates does not authorize deployment, mint a Nova-Seed, authenticate a validator or certify compliance.

The complete demo guide explains the schemas, exact arithmetic, bounds, seed streams, troubleshooting and preserved integrations. The research archive retains all original diagrams and narrative. The former Docker/SDK path remains explicit --legacy research functionality; its live collector and MCTS claims are not the supported lab's behavior.