Skip to content

state-feed.md

Latest commit

 

History

History
59 lines (50 loc) · 3.42 KB

state-feed.md

File metadata and controls

59 lines (50 loc) · 3.42 KB
Raw

Broca state feed

render --state-output builds Brain's shared version-1 contract using the selected Brain checkout's board/_state.py::build_state. Scientist's collector and connect_organ.py consume this contract without Broca-specific parsing. Identity comes from explicit --repository owner/repository; repo in the feed is the repository basename required by that contract. pages_url is independently configured and must be an HTTPS directory. Declared identity is not access proof.

The artifact contains only rendered HTML and JSON. It never includes raw receipts, assistant transcripts, project source or the whole checkout. Rendering does not collect, evaluate, publish by itself, or rewrite the append-only evidence store.

Clocks

  • updated is the saved latest collection attempt, whether complete or incomplete. The collection row distinguishes the successful refreshed_at timestamp from a failed/incomplete attempt.
  • With no receipt, v1's required timestamp uses 1970-01-01T00:00:00Z as an explicit unknown sentinel, not an observation. The row states timestamp unknown and the feed cannot be green. No valid_until is emitted.
  • valid_until is bounded by the attempt timestamp plus collection_max_age_days (default 30), and by the earliest expiry of any currently passing benchmark. Evaluation expiry uses the same UTC calendar-day rule as the dashboard: midnight after evidence_max_age_days whole days. A newer collection cannot extend a passing benchmark's validity. The collection policy remains separate from the evaluation-age policy.
  • Each assistant row retains the latest benchmark's original observation date. A new collection cannot freshen an old evaluation. Rendering never changes any of these dates. No inventory timestamp or quality measurement is inferred for the undated guides/workspaces catalogue.

Status meaning

The same conservative latest-day benchmark selection and evidence_state classification serve HTML and JSON. Maintenance/inventory records are not benchmark passes. Same-day failure wins over pass because day-only timestamps cannot establish ordering.

The aggregate reports recorded failure as red; otherwise expired collection is stale, incomplete collection or evidence needing interpretation is yellow, and all configured assistants having current recorded passes after complete collection is green. Missing collection, no assistants, or only never-evaluated assistants is grey. Items preserve the distinct reasons even when one aggregate status takes priority. Green describes recorded assistant benchmarks only; it does not certify all communication resources or scientific correctness.

Guides, HowTo tutorials and workspaces remain an informational inventory count, separate from assistant evaluation rows. Their projects own source and meanings. Ears listening, Memory knowledge, Cortex science, Heart readiness and Hands release responsibilities are unchanged. The feed contains no executable actions and does not create a new task ledger.

Malformed/digest-mismatched current receipts, invalid records, invalid catalogue shapes and invalid publishing configuration fail rendering. The existing Pages check also validates historical receipts. The job validates the generated feed before uploading the site; deployment runs only on main. Validate PR artifacts before merge, then verify the actual HTTPS endpoint after successful deployment before claiming a live Scientist connection.