You set the test. The rat makes its move.
Write a score, design a challenge, and share a real neural attempt.
Website ↗ · First Takes ↗ · Music studio ↗ · Observe ↗ · Create a mission ↗ · Race ↗ · Buyback ledger ↗ · X ↗
RAT LAB is a community experiment desk: place eight targets, submit a real inference run, then share or remix the resulting mission. Completed, replay-verified runs become recorded opponents. Fixed policies attempt new tasks; they do not learn during these sessions.
Our origin study asked what learned behavior could do when its actions had a checkable consequence. R-01 is a virtual rodent in MuJoCo. One artificial network steers its head to move a cursor; another coordinates its body to press a lever. Eight hits on RAT LAB's virtual control surface advanced a preset BNB Chain / Flap launch plan. The operator chose the token details and spending limit.
| The release | Where to check it |
|---|---|
| Live website | rat-lab.fun |
| Token | RAT on Flap · 0xC090…B7777 |
| Transaction | View the confirmed BSC transaction |
| Brain and session | Download the public release record for the SHA-256 fingerprints and eight-hit result |
| Updates | @gouyi420 on X |
| Experience | What is running |
|---|---|
| First Takes | Four behavioral music studies, synchronized movement-to-note playback, self-contained listening rooms and a test-only NFT archive. Free listening; no production sale. |
| Music studio | Listen first: a recorded performance with synchronized keys and source motion. Compare it with the original score, then compose with A/S/D/F, export the response and remix. First verified take ↗ |
| Create & remix | Three route templates, eight editable targets, a real experiment queue and persistent mission links. Successful verified attempts include synchronized race playback; failures remain documented. |
| Observe R-01 | Our own recurring Aim Eight inference sessions, cursor playback, replay checks and downloadable records. Fixed trained policies; no claim of ongoing learning. |
| Human vs. rat | Race a verified recording in two synchronized lanes. See every target split, finish ahead or behind R-01, and export your local result. The hosted site includes recorded 3D body motion. |
| BNB buyback rehearsal | Unique verified hits generate an illustrative paper allocation. No funded wallet or real purchases. Includes a block-specific token-tax snapshot. |
The observer is a separate service with no signer. Its bounded public submission endpoint accepts bounded target designs or eight-note scores. Community mission archives persist on disk; the separate 24-session paper allocation window resets on restart. Runtime, deployment and verification details →
- Website — mission workbench, our own recorded benchmark, origin study and evidence desk.
- First Takes — behavioral scores, note provenance and independent reconstruction. Method & testnet archive →
- Music studio — listening room, score/response comparison, keyboard composition, verified playback and WAV export. Design & timing contract →
- Mission studio — route editor, queued attempts, saved results and remixing.
- Community service — validated tasks, idempotent submissions and persistent archives.
- Product direction — design decisions, beta limits, API contract and next steps.
- Live lab — observation, proof inspection and a transparent simulation ledger.
- Challenge — shared target rules, local timing and result export.
- Observer — isolated neural inference and deterministic replay verification.
- Launch plan — binds the operator's parameters and spending limit to one transaction intent.
- Local signer — submits only after the target sequence, replay, and plan checks succeed.
- Publisher — checks the receipt and replay before producing the static release bundle.
- Architecture · Validation · Third-party provenance
npm ci
npm run dev
# http://127.0.0.1:5173To preview the confirmed release state locally, save the public JSON record as public/release.json, then run npm run build:public or refresh the dev server. Without that file, the evidence desk correctly shows an unpublished state. The public-source subject view displays an original static mascot illustration. The 2D observation surface plays our own recorded cursor data and connects to the hosted observer. The duel uses a synchronized 2D cursor fallback in this public-source build; the hosted site adds the separately supplied 3D renderer.
The homepage displays our own recorded benchmark, separate from the historical launch evidence. The adapted 3D viewer, original model assets, trained weights, upstream reference assets, and upstream source are omitted from this repository because the upstream Labrat repository has no repository-wide redistribution license. See THIRD_PARTY.md for the exact boundary. A neural session requires separately obtained compatible upstream components.
Original RAT LAB code and artwork in this repository are MIT licensed. The upstream virtual rodent research is credited to Labrat and DeepMind's dm_control. Referenced projects retain their own terms. RAT LAB is independent of Labrat, DeepMind, BNB Chain, and Flap.
The token launch is complete. The code here documents the experiment and does not call for another launch.