narwhal-check

With --fleet PATH, narwhal-check runs deployment gates in this order:

reach → contract → profile → model → pace → tokenize → produce → consume → slo

Option	Default	Contract
--version		Print the installed distribution version.
--fleet PATH	required for preflight or evidence verification	Native fleet config JSON
--ring	mesh	Mesh tests every eligible ordered pair; --ring uses ring coverage for consume.
--repeats N	1	Transfer probes per pair, clamped to a minimum of 1.
--no-kv	false	Runs reach, contract, profile, model, pace, tokenize, and slo.
--evidence-out PATH	gate output only	Write process-bound full-mesh KV evidence to a fresh JSON path. Requires a fleet engine_contract; exclusive with --ring, --no-kv and --verify-evidence.
--verify-evidence PATH	run preflight	Verify saved directed KV evidence against current fleet/profile hashes and live process generations. Requires a fleet engine_contract; exclusive with --evidence-out. --ring, --no-kv and --repeats apply to new probes.
--print-example-config	false	Prints the packaged annotated config before resolving the input config, then exits. Takes precedence over --print-contract-versions.
--print-contract-versions	false	Prints the versioned JSON interface registry before resolving the input config, then exits.

KV checks use the fixed prompt "benchmark " * 64, whose token count depends on the model tokenizer. Each consume probe creates a fresh handoff and requests up to four output tokens from a distinct, role-permitted peer. Decode uses the fleet's engine.first_token_timeout_s and must return generated output followed by a valid stream termination.

When writing directed KV evidence, qualification requires each engine's process generation to match across all pairs and repeats and to retain its profile binding at the final identity check. The fleet and profile files must retain the hashes captured before qualification.

narwhal-check --fleet fleet.json --repeats 3 --evidence-out runs/kv-evidence.json
narwhal-check --fleet fleet.json --verify-evidence runs/kv-evidence.json

--verify-evidence checks the retained qualification against current engine identities and configuration. Fresh transfer probes run through preflight, where --evidence-out records pair outcomes and requires every gate to pass for exit status 0.

The slo gate prices each profile's smallest measured decode cohort, including its active-request and KV-token costs. Capacity output states the request count used for the TPOT calculation.

Measure first-token latency across input lengths with instrumented calibration probes.

In default text mode, exit status 1 indicates a failed gate or operation; exit status 2 indicates invalid arguments or a fleet config read or validation error. JSON mode maps outcomes through the command result contract.

Use gate tables to diagnose deployment failures and the contract registry and versioned artifacts for automation.

Output option	Default	Purpose
--format	"text"	Select json for versioned command results.