What a cognitive map is
A telos-seeded incubation home — no inside, no outside.
A new engineer starts Monday; by Friday the aim is a merge she trusts. Something has to hold the question of how she gets there and keep working at it as it learns. That tended understanding-toward-a-purpose is what we'll come to call a cognitive map — and what it does is the better way in than what it is.
A new engineer joins epd-team-a. The goal that matters first is small and concrete: a pull request she actually trusts, inside the first week. Getting her there isn't a document anyone can hand over — it's a live question, with a few more underneath it. What does she already know that carries over? What's the smallest real change that builds confidence? Where are the sharp edges that scar newcomers?
Those aren't rhetorical. In the running scenario they're written down — the standing questions of a part of the system that exists for exactly this purpose, seeded with the goal "help a new EPD engineer reach first-merge confidence in week one."
What makes it more than a folder of onboarding notes is that something tends it. onboarding-agent#1 — an agent working as a
steward — holds those questions, watches what actually happens as engineers
come through, and records what it learns. When pairing a newcomer with a
maintainer turns out to head off the worst scars, that doesn't stay an
observation; it becomes something the map now does: a small standing
rule, pair on the first PR.
A purpose, the questions it keeps asking, the understanding gathering around it, and an agent keeping the whole thing current — that working apparatus is what we call a cognitive map. The name is a handle for the thing we've just watched do its job, not a category it belongs to.
Our work on theory ends on a hard claim — the system does not store knowledge; it stores data and the traces of acts, and computes projections so that a perspective engaging them can produce knowledge. A cognitive map is where that production is given a home and a direction. Compressed to a sentence:
Temper is an event-sourced coordination substrate whose organizing purpose is to be economical with attention. A cognitive map is a telos-seeded region of that substrate where humans and agents grow a shared, situated understanding together — and everything else is a projection over it.
The rest of this set is that sentence, working.
The thing you actually use a map for is to see where understanding has gotten without reading everything inside it. A map offers a shape: from wherever you stand, you can see the regions it has formed, how much each one matters under its purpose, and roughly how populated each is — while the material those regions are made of stays something you reach for piece by piece, and can be refused.
So it isn't a container with an interior you're admitted to. It's a shape it shows openly and a set of materials you ask for individually — no membrane, no act of entering, nothing bored through. That split — shape you can see, materials you reach for — is what later makes maps legible to one another and gives visibility something to work on.
an emergent field of concept-points that has settled into one or two density regions. Each region carries a label and a salience weight (how much it matters under the map's
telos), and a member_count blur — you can see that roughly N things cluster here, but the individual members are deliberately
not drawn. The honest picture of a cognitive map: a shape with weight and
population, not a walled garden with a gate.
Drawn from kb_cogmap_regions (centroid, salience, label, member_count); the surface read cogmap_shape(), which returns exactly salience / label /
member_count and never member identities. Scenario S6 demonstrates the surface returning the blur with
identities withheld. · conceptual
Why reach for the shape and not the contents? Because handing over everything a map holds would spend the exact attention the system exists to conserve — so instead it shows where the weight is and lets you spend attention deliberately. For the new engineer's steward, that means meeting "pair on the first PR" and the sharp-edges question first, not wading through every note the map has ever absorbed. The map is economical with attention by construction — the thesis already visible in the smallest unit.
One thing this leaves unsettled: when a map's shape gets re-drawn. The regions are materialized — computed at a moment and then held — so a shape can sit slightly behind the events that have since touched the map. The system treats that as normal and reports it honestly rather than blocking on freshness (the staleness signal is real and on-read). What should wake an agent to re-materialize a shape — event volume? a cadence? salience crossing a floor? — is genuinely open. We mark it as a seam and return to it in how a map grows and the deployment forks. Naming a seam as a seam, instead of papering over it, is the discipline the system applies to superseded thinking — and the one these pages try to keep.