Un microscope navigateur pour le calcul des modèles.
Un instrument local dans le navigateur pour mesurer les traces des modèles et distinguer tenseurs observés, interventions, chemins et mécanismes répliqués.
- Publié
- Mis à jour
Périmètre du projet
ACTUEL / PRODUIT DE RECHERCHE
Browser runtime and research instrumentation
- Runtime
- ONNX Runtime WebGPU with WASM fallback
- Primary artifact
- Pinned SmolLM2-135M-Instruct and distinct TinyLM fallback
- Evidence ladder
- E0 observed through E4 replicated mechanism classes
Thèmes
Why put the microscope in the browser
CMET makes the model and its measurements part of the same browser experience. Generation and supported analysis run locally through ONNX Runtime, with WebGPU preferred and WASM available as an explicit fallback rather than a hidden remote service.
That makes the runtime boundary inspectable: model graphs, tokenizers, manifests and frozen evidence are same-origin static artifacts with pinned identities.
Inspect one token decision at a time
The system studies a specific next-token decision at a specific sequence position. The interface can show the generated token, probability trajectory, attention paths and writes that moved the decision, rather than presenting a large explanation detached from one measured forward pass.
The research workspace then renders frozen evidence such as logit-lens views, interventions, path tests and annotations as receipts with their own provenance.
Evidence classes cannot be upgraded by wording
CMET separates direct observations, estimates, intervention-supported results, path or interaction evidence and replicated mechanisms. A fluent sentence describing a result does not change its evidence class.
This is a useful discipline for any scientific interface: the visual story should make uncertainty legible instead of allowing labels to outrun the run that produced them.
The microscope is not a universal explanation engine
The current platform is a scoped instrument for small decoder-only models and declared tasks. Calibration, controls, cross-prompt replication and wider causal qualification remain separate obligations whenever a result moves from exploration toward a research claim.
The product is strongest when it shows the measured trace and its limits together.
Limite de l'évidence publique
The browser instrument and reference laboratory are separate from the strength of any individual scientific mechanism claim. E1 screening, E2 interventions, E3 paths and E4 replication remain distinct; blocked, uncalibrated or unreplicated gates must not be narrated as completion.
Ces pages publient un périmètre d'ingénierie et un raisonnement vérifiables ; les preuves de projet sont ajoutées lorsqu'elles sont prêtes à être partagées.
Voir tous les projets d'ingénierie