Seven instances that propose. You approve.

i wanted the memory system to notice things without quietly deciding they were true.

think of seven assistants working beside one shared inbox. one reads meeting notes, one spots duplicate cards, one finds contradictions, one draws connections. they can draft, label, and recommend. the inbox still belongs to me.

IGRIS calls this the cognition layer. the name is larger than the mechanism. it is a host for seven narrow readers, each built around the same sequence: gather a bounded view of the brain, ask a model for a judgment, validate the answer, then put the result somewhere an operator can inspect.

that last step is where the original field note was too neat. “they propose. you approve.” is the default posture, not an unbreakable law. the implementation has opt-in paths that can apply some results directly. it also has deterministic maintenance that changes housekeeping state when the janitor runs. the useful claim survives, but it needs a smaller perimeter.

01 / THE BOUNDARYThe proposal is not the truth

inference can be autonomous without making acceptance autonomous.

perception reads a session transcript and inserts candidate learnings into a pending-review queue. the other six mostly use the suggestions table: a finding arrives with evidence, confidence, and, where possible, a structured action that can be applied later. an unknown action does not get improvised into execution. it falls back to review.

this means the system does write while it observes. it writes candidates, run events, audit rows, and queue state. saying “nothing writes without approval” would be false. the narrower default is better: inferred learnings do not enter approved memory, and suggested structural actions do not execute, until the review path is used.

that is the boundary i care about. drafts are cheap. truth and structure are not.

02 / THE ROSTERSeven jobs, all implemented

the old note treated seven as a plan and then tried to keep an “active” count beside it. that count decayed. the current registry contains seven implemented instances:

• perception reads session transcripts and proposes learnings worth keeping.
• subconscious reads a digest of briefs, goals, learnings, and activity, then surfaces a missing follow-up, stalled item, or pattern.
• synapse examines candidate pairs of learnings and proposes typed edges between them.
• janitor finds near-duplicates and proposes how they should be merged or kept.
• arbiter finds same-topic opposition and proposes how a contradiction should be resolved.
• curator finds stale knowledge and proposes whether to keep it, lower its confidence, or prune it.
• cartographer finds clusters in the learning graph and proposes a meta-learning that summarizes each cluster.

these are not seven names in a roadmap. they are seven extractor instances exported into one registry. the host does not branch on their identities. each instance supplies the parts that differ: how it builds context, how it prompts, how it validates, where it persists, and how an operator can observe it.

adding another reader is mostly an instance file and one registry export. the reporting roster is then projected from the registry instead of maintained as a second hand-written list. this matters because the previous hand-written health checks knew about two instances and silently omitted five. observability had become a stale inventory of the thing it was meant to observe.

03 / FIVE STATESImplemented is not enabled

one word caused most of the confusion: active.

an instance can be implemented but disabled. enabled but waiting for its driver. scheduled but skipped by a cost gate. run successfully and produce no valid candidates. produce a proposal that nobody has reviewed. those are different states, and compressing them into “active” makes the status sound more certain than the system is.

the shipped configuration keeps the layer off. a fresh install writes enabled: false for perception, subconscious, synapse, and janitor. arbiter and curator inherit the janitor gate. cartographer inherits it too, then adds a second cluster switch that also ships off.

there is one awkward exception worth keeping visible. perception’s code-level resolver treats a genuinely absent key as enabled, for older or hand-edited configurations. the installer does not leave that key absent. it writes the explicit false value. shipped posture and resolver fallback are different facts.

so i no longer publish a frozen “how many are active” number. the code can establish what exists. configuration can establish what is allowed to run. only the live health surface can establish what is happening on a particular machine.

04 / THE DRIVERSSeven readers do not need seven clocks

perception is driven by work. a session-end hook sends it transcript material, and a manual extraction tool exists for deliberate runs.

subconscious, synapse, and janitor are clock-driven. their schedules wake the whole-brain suggestion pass, the edge-inference pass, and the maintenance pass. arbiter, curator, and cartographer do not own schedules. they run inside the janitor path, in sequence, with one maintenance audit around the family. cartographer also has a cadence throttle because clustering the graph is the expensive pass.

this arrangement is less symmetrical than seven independent daemons. it is also more honest. disabling or wedging janitor affects four instances, not one. a quiet arbiter may be healthy but blocked by its upstream driver. a quiet cartographer may simply be inside its cadence window.

“scheduled” therefore means a driver exists. it does not mean the gate is on, the run fired, the input cleared its minimum size, the model returned valid output, or a proposal was produced.

05 / REVIEWApproval is an operation

there are two review shapes.

perception has a learning inbox. i can list a candidate, inspect the full row, edit it, approve it, or reject it. approval flips the candidate into approved memory while keeping its inferred provenance. rejection removes the pending candidate.

suggestions have a separate lifecycle. i can list pending items, dismiss one with a reason, or invoke its structured action. applying an action is explicitly operator-invoked; creating the suggestion does not execute it. a successful apply marks the suggestion acted. a failed apply leaves it pending.

that is more useful than a generic confirmation dialog. the review object carries the proposed change, the evidence behind it, and the action that would be taken. approval is not a mood. it is a named transition with an inspectable before and after.

it also gives rejection memory. dismissed subconscious findings feed a suppression loop, so the same unwanted observation does not have to return forever wearing slightly different words.

06 / THE ESCAPE HATCHESThe default can be overridden

review-first is a default, not a prison.

perception can auto-approve extracted learnings. synapse can write inferred edges directly. the janitor family has separate switches for automatic merges, contradiction resolution, pruning, and cluster meta-learning creation. all of those direct-apply switches default to false, but the paths are implemented.

that changes how the headline should be read. “you approve” describes the stock contract. it does not describe every configuration the software permits.

there is another caveat. when janitor is enabled, its deterministic housekeeping can bump confidence on rediscovered learnings and reject old pending-review rows without asking a model or waiting for item-by-item approval. those operations are bounded and audited, but they are still writes. the safe statement is not “the machine never changes memory.” it is “model-proposed promotion and structural change are review-gated by default, unless the operator enables a specific direct-apply path.”

trust comes from naming the escape hatches, not pretending they do not exist.

07 / THE INSTRUMENTSMeasure the layer live

an open registry needs derived instruments.

igris cognition health reads the projected roster and reports one row per registered instance. it distinguishes disabled, wedged, blocked upstream, failing, healthy, and enabled-without-recent-signal states. that vocabulary points at different repairs. turning on a child instance cannot fix a wedged parent it does not own.

igris cognition yield asks the sibling question: did the instance produce anything, did a human judge it, and what survived that judgment? it keeps “where is the current inbox?” separate from “which rows did this instance ever produce?” a drained queue is not the same as a producer that never worked.

this is where the rebuilt note lands. seven instances exist. four configuration roots control them. four driver paths wake them. proposal queues expose their work. review operations accept or reject it. optional direct-apply flags widen the boundary when i choose.

no frozen active count. no claim that observation is write-free. no universal promise hiding behind a default.

seven instances propose, by default. i approve, unless i deliberately configure a narrower loop out of the way. the line is still human. now it is also inspectable.