Skip to main content

Lag-0 and Tier-2.5

What "lag-0" means​

Circuit discovery mines same-token-position co-activation: it asks whether an upstream feature and a downstream feature tend to fire on the same token. That is exactly right for computed structure — a feature at one layer influencing a feature at a higher layer within a single position's residual stream.

It is blind to the circuits people most want to name: attention-mediated, cross-position ones. When a feature fires on a name three sentences ago and drives a pronoun-resolution feature here, the two never share a token position — so a same-position count can't see the link. This is the lag-0 limitation, and miStudio discloses it on every discovery surface: the UI Discovery tab, every MCP tool description, and a lag0_disclosure line in every run report.

Why the limit is liftable without re-capturing​

miStudio's capture store and its portable circuit contract were built position- and attention-aware from day one:

  • every activation event stores its token position as a first-class column;
  • capture can record an attention sidecar (which key positions each query token attended to);
  • the circuit contract reserves an attention_mediated edge type and nullable position/head fields.

Because those fields already exist (populated or explicitly null), enabling cross-position discovery is a matter of turning on new mining — not migrating any stored circuit or re-running any capture that already recorded attention.

Tier-2.5: the designed fast-follow​

Tier-2.5 is the name for attention-mediated cross-position mining. Its design is fully specified (attention-weighted candidate generation, a shuffled-join null that tests the attention routing, frozen-attention attribution, and position-restricted causal validation). The implementation is a named fast-follow — deliberately deferred so the honest lag-0 tier ships first — and when it lands, it populates contract fields that already exist rather than changing the schema.

The point: the interpretability studio tells you today exactly what it can and can't discover, and it's engineered so that lifting the limit is future work, not a migration.