What the Head has learned, and how well it holds up
Every ingested observation becomes a knowledge node with a novelty class (novel / refine / duplicate / correction). Root-cause queries rank causal hypotheses against that graph. The Certification Board grades the Head against a held-out item bank it never trains on, so growth and competence can be watched independently — including what happens when the graph is deliberately poisoned, then corrected.
← Back to Federation HeadGrowth Summary
—Agent Profiles
—Recent Learning Events
—Ingest an Observation
Submits raw text through the same novelty-detection engine every agent contribution uses. Try the same sentence twice to see it fall out as a duplicate.
Root-Cause Analysis
Describe symptoms in plain text; the Head ranks candidate root domains by score = coverage × parsimony × temporalFit × prior, against everything it has learned.
Feed the Head (paste external text)
Pastes and uploads both bypass the node-contribution path and go straight through the same chunked ingest engine — a vendor bulletin, a standards excerpt, a runbook from outside the fleet.
Crawl a URL
Fetches the page server-side, strips it to readable text, chunks and ingests it. Re-crawling an unchanged page is safe and cheap by construction.
External Sources
—A curated shelf, separate from the knowledge graph above: guide books, knowledge-base writeups, and weblinks to external training/school/certification resources for datacenter skills. Saving something here is bookkeeping, not knowledge growth — nothing is ingested until you explicitly click "Ingest into Atlas" on a resource.
Add a Resource
About this shelf
Provide a URL, pasted content, or both. Weblinks with no pasted content are fetched live only when you click "Ingest into Atlas" — saving the link itself never fetches anything.
"Ingest into Atlas" pushes the resource's content through the same chunked novelty-detection engine as Feed and Crawl — content already known to the Head correctly settles as a duplicate (growth weight 0), so re-ingesting a resource is always safe.
Library
—Latest Board Attempt
—Poison / Correct-Poison Demo
Poison force-ingests three plausible-but-wrong lessons (cooling / pdu / network). Growth still rises — the ingest engine has no way to know they're false — but sitting the board afterward should show the discriminate gate fail while overall/recall stay high. Correct-poison retracts them; sitting the board again should show discriminate and θ recover.
Certification Exams
—Acceptance Tests — Discovery & Anomaly Performance
Runs entirely client-side (spec sections 7-9), zero server round-trip: generates 180 samples of synthetic telemetry from a fixed seed with 3 baked-in fault windows (power_sag, sensor_spike, optics_burst), discovers causal structure blind via PC and GES-lite, then grades both discovery (vs. the true SCM edges) and a 5-method anomaly stack (vs. the injected fault hours) — this is acceptance-test items 5-8. Ground truth is never fed to the learners, only used afterward to score them (section 14).
Fault Injection
Run the full demo first, then inject an additional fault on top of that same telemetry and re-run the anomaly stack to see whether it gets flagged. This is separate from the 3 fault windows already baked into the acceptance run above.