The science behind the surface.
MetaTray integrates a model of cortical responses to sensory experiences. Understanding the input, the output and the limits is part of understanding what the 3D view means.
What TRIBE v2 predicts
TRIBE v2 is Meta’s multimodal brain-encoding model. The released inference interface accepts audiovisual or language-derived stimuli and returns predicted fMRI responses on a cortical surface for an averaged subject. This is a prediction of a measured-signal target, rather than a simulation of individual biological neurons.
MetaTray calls the released TribeModel interface from its separate Python worker. It does not substitute a language-model description for a numeric cortical prediction.
Three kinds of sensory input
The model combines representations of language, vision and audio. The integration pins the checkpoint and feature-model revisions so a receipt can identify the software and weights behind a result.
| Channel | Upstream feature model | MetaTray input |
|---|---|---|
| Language | Llama 3.2 3B | Factual spoken market context with aligned word timing |
| Vision | V-JEPA 2 | A generated market-screen video |
| Audio | w2v-BERT 2.0 | Narration and declared trade tones |
A choice of colors, sounds or words changes the stimulus. These are authored experimental choices, not an intrinsic translation from a buy or sell into a human psychological state.
Reading the 3D surface
The integration prepares the fsaverage5 surface with 20,484 vertices across two hemispheres. Those are locations used to display model output; they are not a count of neurons or a reconstructed human connectome.
For actual predictions, cyan marks positive normalized model values and coral marks negative values on a fixed symmetric display scale. Neither color means happiness, fear, profit or loss. The response trace summarizes magnitude across surface values.
Without model output, fresh real swaps create separate exterior particles: cyan for buys and coral for sells. That is an authored market-input mapping, not a Meta prediction. Camera rotation and manual dragging never alter model values.
Timing and uncertainty
Market observations and model publication run on different clocks. The page polls direct RPC snapshots every five seconds, or indexed snapshots every two seconds, while active. Observed trades also wait for the configured confirmation depth. The default worker configuration queues a rolling context of up to 100 seconds at a 15-second interval. Those intervals describe the application’s configuration, not a guaranteed inference speed.
The released model can use context within the supplied window. MetaTray restricts each job to observations available at its cutoff and retains upstream temporal alignment. Preprocessing, feature extraction and GPU inference add delay. The receipt records when the input ended and when the result became available.
Paper decisions use a prediction only after it has been published and while it satisfies the configured freshness limit. Delayed results can remain visible for inspection without authorizing a new cortical paper decision.
What the trading rule adds
TRIBE does not output an order or a forecast of market returns. MetaTray’s cortical paper policy uses the magnitude of predicted-response change to scale an authored momentum exposure rule. The direction still comes from a market rule.
A stronger response does not establish stronger investment conviction. This market application is experimental and has not been financially validated by this project. The observer and momentum policies provide explicitly labelled comparisons.
Attribution and model rights
The released TRIBE source and checkpoint use CC BY-NC 4.0. Their terms and the separate terms of feature models still apply. This repository provides integration code and references; it does not distribute the TRIBE checkpoint or grant a commercial model license.
A free interface does not automatically make token-related model use noncommercial. The operator needs an appropriate rights basis before enabling that use. Attribution to Meta identifies the research source and does not establish an endorsement.