Cohorts API

Base URL: /api/builder/v1. Authentication and X-Account-ID follow the existing Builder contract. Restrictive API keys require inference:execute.

Method	Path	Role	Notes
GET	/cohorts/capabilities?endpoint_id={uuid}	viewer	Effective endpoint capability; use this form in UI/client preflight.
GET	/cohorts/health	operator	Content-free aggregate registry/residency/fabric health and admissible capacity.
POST	/cohorts:plan	viewer	Read-only dry run.
POST	/cohorts:run	operator	Creates a durable root; supports Idempotency-Key.
GET	/tasks/{task_id}/cohort	viewer	Redacted summary.
GET	/tasks/{task_id}/cohort/events	viewer	Cursor JSON or SSE by Accept.
GET	/tasks/{task_id}/cohort/explanation	viewer	Redacted factors.
POST	/tasks/{task_id}/cohort:cancel	operator	Idempotent terminal cancellation.

Execution policy fields are mode, fallback, trace, max_members, latency_budget_ms and cost_budget_credits. The initial executable topology caps max_members at two even though future-compatible schemas may accept a higher server-bound value.

Cohort is opt-in and endpoint-scoped. It is currently admitted only for explicitly assigned causal/autoregressive text-generation LLM endpoints. Omitting execution continues to use Classic.

This API controls the production two-member hidden-state lane. It does not expose the R4-M textual-memory intervention, install a memory policy, or turn memory on for an endpoint. R4-CEM and R4-M run through separately frozen research harnesses; they do not establish a hidden-state, KV-cache or memory-by-multi-agent efficacy claim.

All errors use the stable Builder envelope. No endpoint returns raw intermediate representations or physical topology. Operational health contains counts and component booleans only: no node, model, container, peer or tenant identifiers.