A team needs shared context. It still needs clear boundaries.
David built COHORT to organize his own projects and tasks as a team of engineering roles with deliberately different perspectives. Research and recorded decisions helped that team develop and build ideas together.
Every control here prevents a real failure: duplicate work, approval of the wrong revision, lost rationale after a handoff, or an unknown result reported as done. More context helps a successor understand a decision, and permission, ownership and review still come from the records that own them.
This atlas links every system to its source. The dashboard code is included, most services below are documented contracts that run elsewhere, and the public demo uses sample data offline.
ONE INVENTED EXAMPLE / NO HISTORICAL WORK DATA
A task count looks current. It's already stale.
Observe the failure. An analyst notices that a sample task counter misses a newly changed item. Spyglass helps locate relevant research; the team confirms against the Cairn source before acting on a search snippet.
Choose the right lane. A narrow defect may be a SWAT. Existing approved build work may be a task. A new cross-system freshness contract belongs in an RFC. Only new cross-system designs need a proposal.
Keep the different perspectives. The architect asks who owns the count; the PM asks how adoption fits the work; builders ask how to implement it; reviewers ask which revision was checked; the analyst asks whether the operator can see uncertainty.
Decide, then implement. On an RFC, each node weighs in from its role and the work it leads, and the PM writes up where the team landed. That summary carries the most weight, and the operator has the final say on when the design is ready to build. Routed work and independent code review follow as separate steps.
Deliver and carry the rationale forward. Completion must reach the owning record and the operator's view. A successor can retrieve the saved decision, handoff and lessons learned, so the reasoning never has to be rebuilt from chat history.
Reusable conventions grow from preserved research and decision history, and each change still takes the lane that fits it.
Boxes group related boundaries for readability, and several may run inside one service. Supporting components are browsable below.
Network / API observations
Stored / retrieved context
Control / authority
Execution / orchestration
Work handoff
Dashed box: documented external runtime. Double outline: included reference UI. Hue shows the kind of connection.
Read the same map as text
Operator intent and OPA constrain coordinator actions.
Research/search retrieves context; Cairn owns the source revisions and decisions.
RFCs settle a new design with the team's input; tasks and SWATs own their separate implementation/fix paths.
Agent identity, memory and handoff support execution without completing work.
Independent review binds to the artifact; accepted closure and delivery need more than a message or a nod in discussion.
Superdash consumes APIs and event projections, displaying facts without becoming their owner.
BROWSE THE SYSTEMS AND BOUNDARIES
37 connected systems and boundaries.
The inventory follows the exported system catalog, architecture/rebuild guides and included frontend. Aliases are reconciled here: ZeroBrain is the integration umbrella, Superdash its operator UI; Spyglass and Cairn are separate; the two councils and the two search indexes are distinct.
37 entries; open a system to inspect its purpose, authority, relationships and exact evidence.
Authenticated results, durable transitions and dashboard/event projections.
Source of truth / authority
Coordinator records own live fleet state and operational actions.
Runtime boundary
Documented FastAPI/MCP contracts; server implementation and running service are external.
Failure prevented / review boundary
Completion comes from the owning record, so a message, heartbeat or frontend success indicator never counts on its own. Exact protected actions go through their backend gates.
Preserve proposals, research, rationale, revisions and decisions so later work can inspect why a design exists.
Inputs
Seeds, RFC bodies, responses, signals, syntheses and knowledge records.
Outputs
Content-addressed revisions, lifecycle/audit records and source material for search.
Source of truth / authority
Cairn's durable records; the guides distinguish its knowledge store from coordinator operational records.
Runtime boundary
Documented persistence/contracts; no Cairn backend or data service bundled.
Failure prevented / review boundary
The enforced stage is ideation, which replaced the stale rfc alias. Ratified content is guarded, and changes follow the documented revision and authority path.
Turn a proposal and the team's input, objections included, into a design the team can build from.
Inputs
Proposal body and revision, adoption plan, each node's responses and reactions, the PM synthesis and any cited approvals.
Outputs
seed -> ideation -> in_round -> ratified -> shipped; append-only lifecycle audit and a ratify binding.
Source of truth / authority
Cairn. A wave is a round of team input; building and delivery come later through tasks.
Runtime boundary
Contracts documented; original lifecycle/wave display helpers are included, and backend enforcement runs outside this snapshot.
Failure prevented / review boundary
Each node weighs in from its role and the work it usually leads. The PM's summary of each round carries the most weight, and the operator has the final say on approving the design as ready to build. There's no head count, and every response stays on the record.
Give a narrow defect a specific owner, verdict and closure path instead of forcing it into a large design process.
Inputs
Targeted observations, design input where needed, fix citations, reviewer verdicts and dispositions.
Outputs
open -> in_review -> fixed -> closed, or explicit changes/rejection/refile outcomes.
Source of truth / authority
Coordinator Swatter rows; they stay in the coordinator despite the cairn_ prefix.
Runtime boundary
Documented state machine; reference UI included, backend absent.
Failure prevented / review boundary
Design input is always considered. Author/prior-reviewer/cluster/design-input recusal is lane-specific, and unknown ancestry stays visible until verification runs.
Prevent approval of one revision from being applied to another, or an author's claim from replacing independent evidence.
Inputs
Artifact/base/head relationship, eligible reviewer, verdict, notes and required acceptance evidence.
Outputs
Recorded acknowledgements/verdicts that let the owning task or SWAT advance.
Source of truth / authority
Coordinator review records and the owning work lane.
Runtime boundary
Review contracts documented; original review table renderer included, enforcement/runtime absent.
Failure prevented / review boundary
Strict task ancestry mismatches are rejected; non-strict mismatches are audited. Depth rises for vital runtime/lifecycle changes. Authorization and completion are separate steps after review.
Accepted/rejected privileged action and audit evidence.
Source of truth / authority
Coordinator authority records. OPA grants bounded elevation, and review and evidence still apply.
Runtime boundary
Contracts documented; no keys, sessions or authorization service in the public site.
Failure prevented / review boundary
Unknown or out-of-scope authority needs a new decision. Irreversible/PM-only paths may disallow elevation; undocumented auth labels remain explicit gaps.
Avoid recurring idle prompt cost while retaining task ownership, authority and completion evidence.
Inputs
Explicit wake, PM-authored work chunk and authoritative coordinator actions.
Outputs
Executed chunk, node-owned append-only proof and PM reconciliation.
Source of truth / authority
The plan coordinates assignment; accepted coordinator actions still own task/SWAT/review closure.
Runtime boundary
Documented operating model; no wake/schedule/launcher runs in the demo.
Failure prevented / review boundary
Scheduled prompting is suspended and shown as suspended. Proof is written after the owning action succeeds. The three durable work lanes stay unchanged.
Resolve paths from the node's own identity; keep secrets out of generated source and logs. Manual mode changes the post-boot loop and leaves authentication as is.
Fleet topology / service discoveryDocumented / runtime external
Canonical static reachability reference
Make co-tenant safety and service location explicit before restart, launch or vital changes.
Inputs
Generic host roles, service identities and approved topology/profile relationships.
Outputs
Static reachability/placement references used by runtime and review.
Source of truth / authority
Fleet Topology is the static reference; coordinator state describes live fleet observations.
Runtime boundary
Contract only; no real infrastructure roster or host addresses are published.
Failure prevented / review boundary
Static placement says where a node should be, and liveness needs its own check. Vital launch review requires topology/profile-isolation evidence for the actual host.
The exported guides cover reference implementations, design intent and the questions still open. A live deployment, the full backend and every historical capability sit beyond this map.
Role and persona documents carry each perspective; typed memory, handoffs, lessons and saved decisions carry the context. Authority, reviewed revisions and accepted completion stay with the records that own them.