Loading
Loading
Release · August 24, 2026 · All news
This cycle was built from a literature pressure-test, and shaped as much by what the evidence refuted as by what it supported. Plans come only from stored decisions. Completion is accepted on what was run and read. Gates vouch only for rules they can enforce. And the skill you wrote two days ago now fires when you describe the situation in plain words.
Already installed? Ask your agent to “update Cognition” in plain English. Hosted behavior updates automatically with the same key — no new key needed.
01 1.6.84
cognition_plan now returns a supplied plan assembled from your stored decisions, with a reminder cadence so it gets re-injected as a long run goes on. In published evaluations across 21,120 coding-agent trajectories, a supplied plan improves resolution and periodic reminders add another five to twelve points, because compliance decays over a long task.
The same evidence found that a subpar plan hurts more than no plan, and that single finding decided the design: Cognition refuses to supply a plan it cannot ground. Every step traces to at least one stored atom. Hard policies sort first, because a constraint that rules out an approach is worth more than a procedure that assumed the approach was allowed. When the brain has nothing to say about the task, silence is the correct output.
02 1.6.84
cognition_gate in final mode now takes a claim and its evidence, and denies a done-claim carried only by confident wording. Agent-written tests used as a self-gate moved resolution by zero and by minus one point on two frontier models, and roughly three quarters of self-assessing coding trajectories assert a success the state contradicts, because judges key on closing language rather than verified state.
So the gate does not ask a model to judge. It asks one mechanical question: was something run and observed? Evidence with no observation, or marked unverified, does not count. Hedged wording backed by a real observation passes. Deliberately absent, because the evidence does not support them: no subagent fan-out for coding, no model judging its own work, and decision records treated as an audit artifact rather than a correctness mechanism.
03 1.6.84
The gate enforces hard policies through a forbid list — the phrases a violating step would contain. An invariant stored without one was being skipped by the matcher and still counted as checked. Now a vouch requires an enforceable invariant, the unenforceable ones are named in the result instead of silently skipped, and capture refuses a new invariant that names no violating phrase.
Two new surfaces close the loop. cognition_arm_invariants reviews every rule already in the brain, buckets it as armed, armable from its own wording, or needing a human, and applies only to ids you name — it plans, and never arms on its own. cognition_blocked is an attention queue over runs a gate refused: denials listed individually with the act that clears them, deferrals aggregated into one counted item, so the queue never cries wolf.
04 1.6.81 – 1.6.82
A skill written for exactly the situation you are in would fail to fire if you described it in plain words instead of the skill's own vocabulary. Retrieval now scores the task against the skill's When to fire block with light stemming and a small concept map, so production and deployed, or updates and update, meet instead of passing each other.
Additive, not a rewrite: the precision-tuned keyword scorer stays, and a confident trigger match is a fourth signal that faces the same precision bar. Ubiquitous org terms corroborate but never carry a match — at least one strong shared signal is required, so a benchmark skill stays blocked on an unrelated task while the deploy case passes. When a bridge is crossed, the result says so: matched on stems and related terms, not literal keywords. No embeddings, no hosted model, and a person can correct the map.
05 1.6.80
The candidate drafter used to gate on frequency, which encodes a wrong belief about where value lives. A ninety-minute migration done twice a year costs more than a thirty-second command run daily, and only the second one produced a candidate. Capture now scores effort and treats recurrence as one input to expected value rather than an admission ticket.
Routines are induced from action sequences with paths, commits, URLs, and numbers abstracted to variables, so two people doing the same thing on different branches produce one routine. Friction counts: retries, repeated actions, and failure-then-pivot weight the effort up, because they record the moment a routine was needed and absent. One-offs are marked observed once, with low confidence. Candidates are named from their distinctive verbs, and navigation scaffolding is filtered out, so the routine underneath actually gets read.
06 1.6.79
cognition_code_graph grew from three actions to nine: import a deterministic extractor's graph, bind a diff to the components that own it, report the paths nothing owns, traverse paths and neighbors, and export a report a reviewer can disagree with in a pull request. Impact and review now return the blocking invariants and the judgment inside the blast radius before the hit list, and the map reports its own drift on every read.
Underneath, every response is now bounded by tier rather than by a size cap. Telemetry is trimmed before evidence, evidence before directives, and gates never — and every omission is named in a manifest with the call that retrieves it. A silently truncated response reads exactly like a complete one; a named omission does not.
Also in this cycle
The hosted behavior is live now. Ask your agent to update Cognition in plain English, then get back to work — the next plan will be grounded, and the next done-claim will be checked.
