Models are retrained from the chain mirror on a fixed interval, and at start-up any model that cannot be loaded is retrained. A model trains only once it has its minimum amount of data (shown per model above); until then transparent rules answer and the response says so. Every new version is signed (ed25519) and its hash registered; the runtime refuses unsigned, altered or outdated artifacts. Unchanged data creates no new version, and recent versions are kept for rollback.

Each active validator is expected to produce 1/N of the blocks, where N counts the active, non-jailed validators of the chain registry — retired validators never count. The score compares every validator's hourly share with that expectation; the tolerance (at least ±10 %) is learned from the network's own hourly variation. Long silence or jailing caps the score. Status: healthy, degraded or offline.

KMeans over hourly behaviour profiles (share of the expected blocks, longest gap). Cluster names are absolute — steady, bursty, covering, degraded, offline — so they describe behaviour rather than a ranking. Advisory only — never affect consensus, rewards or jailing.

Fees per hour and user transactions per hour, forecast with Holt-Winters (daily season). The model is selected against the seasonal-naive baseline on held-out hours and reports MAE / MASE. It needs 24 complete hours of chain data; until then the forecast repeats the same hour of the previous day and is labelled as such.

Transactions, DEX swaps and AMM pools are scored against robust baselines (median / MAD) of real user activity — protocol transactions such as block rewards, unjail and slash are left out. Swaps are compared with each pair's learned price. Every score lists the features that raised it. Advisory only — nothing is blocked or moved.

Communities ranked by suspicion score (size, intra-density, closure). Advisory only — never affects consensus.

Template-driven, signed, fully deterministic. Same (topic, lang, context) always produces the same paragraph — auditable end-to-end via AI Receipts.

Forecasts stay within installed capacity. Each prediction is signed and auditable via AI Receipts — same envelope shape as gas / volume forecasts.

Local extractive RAG (document-rag-v1). Raw uploads are not persisted; chunk hashes anchor receipts.

auto-intelligence-v1 orchestrates TX classify, DEX swap monitor, liquidity health, bridge hints, standards match, token scan, and wallet summary into signed advisory cards.

intelligence-interpreter-v1 aggregates chain analytics, auto-intelligence, CarbonFi signals, descriptive statistics, and optional RAG into an executive dashboard briefing.