Access matrix
Read is independent of write: if a model can hold enough context
and its provider is private enough (local / no-train), it gets full agency read
without climbing a write ladder. Household private data never goes to caution free tiers
(Gemini/Mistral free train on inputs). Tools and write
still scale with measured quality. Computed by bin/model-registry.py matrix.
Scored model registry
Sorted by average capability score. Access chips are the live grants (hover for why). The optional Score badge is a ranking summary only — it does not grant tools, write, or memory.
| Loading… |
What the scores mean
Each column is a fixed probe bank in state/model-registry-probes.json, graded by bin/model-registry.py. Hover a column header in the table for a one-line tip. Bounded 0–10 scores feed the optional ranking badge; live grants are the Access matrix. Where you see earned/possible points, those include harder “advanced” items that never change access floors — they only rank models that already maxed the old 10.0 ceiling.
Score badge (optional display only)
T0–T5 labels are an optional score badge for ranking/history — they do not grant or deny capability. Hands/tools/read/write/memory are driven by the Access matrix above (models.access_matrix) plus admin floors. Full rationale: docs/TIERED_MODEL_ACCESS.md.
Providers & privacy posture
Privacy is a provider-level property, not a per-model score — every model above joins to one of these. Local (Ollama) is 100% private by definition, nothing leaves the box. Cloud lanes are labeled from each provider's actual published data-retention/training policy, cited below; anything not yet checked is marked unverified rather than assumed favorable. This is what actually gates whether a lane may see household/private context.