Inspect the code. Verify the data.
Follow the actual implementation from a decoded transaction to an audiovisual input and, when the GPU pipeline is running, a Meta TRIBE v2 result.
The source repositories
MetaTray — fourtyeightTech/metabrain contains this website, the chain adapters, rendering pipeline, model adapter and paper engine.
Meta — facebookresearch/tribev2 is the original released research implementation. Pons — ponsdotdev/ponsfamily is the protocol source used to audit the launch-record, curve and graduated-pool interfaces. These references identify independent upstream projects and do not imply sponsorship or partnership.
The code that runs the experiment
- MetaTray chain decoder: reads curve/pool events and preserves transaction hashes, block hashes, log indexes and raw token amounts.
- Pinned Pons launch interface, factory, curve and V4 PoolManager event: the immutable protocol sources matched by the adapter.
- MetaTray model adapter: imports
TribeModel, prepares events from the rendered video and callsmodel.predict. - MetaTray stimulus renderer: constructs the market video and narrated input.
- Meta inference implementation: the upstream pretrained-model and audiovisual prediction interface.
- Meta model architecture: the released encoding model.
Pinned upstream implementations
The model adapter requires Meta source revision af58661791a351a448a489042a28f6c37e1c14b7. The audited Pons interface references use revision cb5748a29e4d3a7af1c4e982baa9ed9194d25a8f. Those links are immutable; downloaded TRIBE weights and feature models have their own immutable revisions in a completed result manifest.
MetaTray source links point to its main branch. Check the deployed build’s Git revision against that branch before assuming both contain the same code. A source match is still not a live-chain receipt or a completed model run.
Evidence for real market data
The live input stream shows the chain, confirmed block window, successful read time and decoded swaps. Open an event to inspect its transaction hash, block hash, log index, raw amounts and price basis. If an explorer is configured, the receipt links to the transaction.
Direct RPC mode re-reads a bounded recent window and does not maintain a durable paper account or queue model jobs. Indexed mode supplies the persistent event history, paper ledger and GPU queue. A connected provider with no matching swaps is shown as waiting, not populated with invented trades.
Evidence for an actual Meta result
Require a completed prediction receipt with its source window, model and feature revisions, stimulus hash, output hash, input age and processing latency. Match those hashes to the retained worker artifacts. The fsaverage5 surface must match the returned vertex order and dimensions.
The market-input animation is an authored visualization of observed buys and sells. It is not a cortical prediction. Only a returned TRIBE result is labelled model output; stale results are labelled delayed. The research release predicts fMRI responses to sensory stimuli, not token prices or trading success.
The released Meta code and weights are CC BY-NC 4.0. The deployment still needs a rights basis for its intended use and a separately configured GPU worker.