Release · September 11, 2026 · All news

Cognition 1.6.92: score the definition, not the number.

Analytics answers fail when they quietly pick a definition. Cognition scores the interpretation channel, declares the reading, and refuses when experts disagree or nobody has attested one. No patient data crosses that endpoint.

01 1.6.92

Score the reading, not only the number

When an agent answers an analytics question, the risk is not a wrong number but a confident one computed under a definition nobody agreed to. Cognition now scores which reading an answer depends on, declares it, and refuses to pick when two experts disagree or nobody has attested one.

Interpretation channel: governed / ambiguous / contested — so an expert can object to the reading and not only the number. Unattested definitions are reported, never applied as authority. No patient data crosses /interpret: chart text and result sets are refused, not dropped.

02 1.6.92

Denials that name what has to change

An analytics denial now names the role, the ownership check, and the row that has to change — so a blocked answer is actionable instead of opaque.

Fail closed stays the default: unauthenticated calls, bad JSON, unknown workflow, unknown registry, and malformed observations all refuse instead of defaulting permissive.

03 1.6.92

A status line that stays honest on your machine

Persistent status line: version, active workspace, last action and session call count, always on screen. Carolina blue identity; NO_COLOR honoured. Cognition spinner verbs append to Claude Code's rotation without replacing the defaults.

Local-only and guarded: no network on render, and a missing segment prints nothing rather than a placeholder. Never replaces a status line you already configured; Cognition's own updates in place.

04 1.6.92

v1 REST when you do not want an MCP client

POST /api/v1/clinical/observe and POST /api/v1/clinical/coverage under Bearer-key auth, no MCP client required. The detector contract ships to the caller, so chart-derived output never leaves their trust boundary.

Chart text sent to either endpoint is refused outright rather than silently dropped. Only a positive observation counts as coverage; negated, absent, and unobservable all resolve to unresolved.

Also in this release

Clinical oversight that stays fail-closed

• EHR-assistant oversight pack: 7 layers, 32 checks auditing the deployment, not a trained model.
• Required-action registries for appointment prep, medication reconciliation, and results follow-up.
• Waivers are approved, credentialed, and expiring; unapproved or lapsed ones restore the requirement.
• Model-authored ground truth is refused, not downgraded — a target the model wrote cannot measure that model.
• Evidence grounding on required actions: unverified citations count for nothing; contradicting sources are never out-voted.