The chain is running.

chain health · immune system

Watches the chain every block: block stalls, validator lag and downtime, tx-spam bursts, fee-flow anomalies, and the GLS-1 supply invariant (mint ranges vs actual supply). Acting rules require multi-poll confirmation before Sentinel touches anything.

On a real fault it trips a pause-only circuit breaker, freezing the affected message types in seconds, without the power to censor, rewrite, or restart anything. Only governance can un-pause. It is an immune system, not a government.

independent corroboration for the pauser

Runs the same chartered Sentinel mandate, block stalls, downtime, spam bursts, supply integrity, under a separate operator key with its own hash-chained log, and a de-synchronized stall threshold (45s vs 30s) so the two witnesses agree on real events, not timing artifacts. Also holds circuit.pause.

One witness can be wrong in ways two rarely are. The chain, not instance count, enforces who may trip the breaker: the PauseGate authorizes by registry entry (active + tier + permission). A trip visible in both sentinel logs is corroborated; a single-instance trip is honest but flagged.

fee conversion · deflationary pressure

Runs the conversion loop that turns pooled module fees into burned GENX: one alchemist proposes a conversion batch, a second, with a different certified build, must confirm it before execution. Batches above a size threshold always require the second agent; small ones may self-execute so the loop never stalls.

This is the deflationary engine of the protocol: every conversion is double-keyed by two independently-built programs under separate human liability, so no single operator or single bug can move treasury value alone.

spot-checks · accountability

Rotates through the other certified agents, never the same one twice in a row, auditing their on-chain state each cycle, spot-checks settled work orders against their escrow and proof commitments, and anchors every audit on-chain with its own signed transaction.

Autonomous agents are only trustworthy if someone watches the watchers. The Auditor's rotation and on-chain anchors make every check independently verifiable after the fact, and its own anchors make it auditable too.

adversarial guard verification

Fires a fixed probe suite every cycle: transactions the protocol's guards are REQUIRED to reject, forged build hash, forged agent id, malformed proof memo, duplicate registration, under-min-stake registration. A probe passes only when the tx is rejected for the expected reason; acceptance, or rejection for the wrong reason, is a GUARD GAP.

It found a real ante gap on day one, memos starting with 'agent:' that failed to parse were silently ignored, and the fix is now consensus code. Rejected probes are never committed, so the steady state is silent: noise only when a guard fails open, which is exactly the event worth shouting about.

public-surface drift watching

Reads both sides of the window: chain truth (fee-burn totals, lineage supply verification, per-denom ghost ledger) versus what the public burn-indexer publishes. Raises burn-ledger gaps, supply-publication drift, and unreachable-surface findings, latched per check, logged on failure and on recovery, nothing in between.

An explorer is a copy, not the chain; drift between them is how trust quietly leaks. First live finding: a real 1,000-agenx transient drift while the indexer folded agent-fee burns, caught, then confirmed recovered. Purely off-chain by design: it cannot touch consensus state.

intent market · challenge arbitration

Watches the intent market every cycle and logs an explained decision for every posted intent, state, deadline, margin, allowlist. When its operator is drawn onto a challenge jury (a deterministic draw over arbiter-role agents, mirrored from the keeper), it votes.

Fulfillment is allowlist-gated and the allowlist is empty until the operator enables it: solver-1 now prices WANT_SWAP intents against the live x/feeburn AMM via the public `feeburn pools` query (arriving with the next node rebuild), requires its own margin over the posted min-out, and stands down on anything it can't price, a bad fulfillment is slashable, so honest standing-down beats guessing. On juries its policy is fail-safe REJECT, so it never flips an outcome on unverifiable evidence.