Runtime verification agent
Every state checked.
Every violation caught.
mcruntime verifies agent invariants two independent ways — a static BFS model checker over every reachable state, and a live runtime monitor watching a real trace step by step — against the identical invariant definitions. A disabled monitor and a bug-free system look nothing alike here.
mcruntime · model checking + runtime verification hybrid · [higgsfield: noir gold state-machine graph, BFS sweep and a live pulse both converging on the same flagged node] · proof-false-done
The failure mode
The monitor nobody missed — until they did
“We turned off the safety monitor to hit a deadline. Nobody noticed until the postmortem.”Eng director · multi-agent orchestration platform
A disabled or degraded runtime check is invisible right up until the violation it should have caught actually happens in production — and by then the incident review is the only place anyone looks. Static analysis alone doesn't catch this either: it proves a property holds across the modeled state space, but says nothing about whether the live system actually stayed inside it. You need both checks, on the same invariant, or you're trusting a monitor you never verified is watching.
How it works
Three steps. No faith required.
01
Model-check
BFS explores every reachable state in the ghost-augmented state space, against a golden model and five planted-bug mutants.
02
Monitor
The same invariant, checked live, one step at a time, on a real simulated agent trace — with no lookahead.
03
Seal
Static completeness and a runtime alarm within 2 steps both have to hold, independently, before the packet seals.
Proof band
One packet. Every claim accounted for.
packet rig/mcruntime · model checking + runtime verification claim "static checker finds all planted violations; runtime alarm ≤2 steps of live violation" battery golden + 5 mutant scenarios · live trace: direct_error_to_done static PASS — 6/6 scenarios match expected violation set exactly runtime PASS — violation at step 5, alarm at step 5, lag 0 steps ≤ 2-step budget combo_check PASS — 2 coexisting bugs in one table, both found by BFS themis SEALED — terminal verifier signed
Disable the runtime monitor and the violation still genuinely occurs at step 5 (confirmed by an independent oracle) — but with no monitor attached, runtime_alarm_raised is False. The alarm gate goes RED even though the static checker still passes. How ProofPackets work →
Before you install
The obvious objections
Q Static analysis already tells me the state machine is safe.
Static analysis proves a property holds across the reachable state space — it says nothing about whether your live agent actually stayed inside it. mcruntime does both: BFS finds every planted invariant violation in the modeled space, and a runtime monitor catches the same violation live, within 2 steps, on a real trace.
Q Why do you need two checkers for the same invariant?
Because a monitor that's silently disabled is indistinguishable from a system with no bugs — until you check both ways. Disable the runtime monitor and the violation still genuinely occurs (oracle confirms step 5) but goes completely undetected: the alarm gate goes RED even though the static checker still passes.
Q "Every planted violation" — is that really exhaustive?
Five distinct mutant transition tables, including one with two independent bugs coexisting in a single table. BFS over the ghost-augmented state space has to surface both, not just the first one found — and it does, exactly matching the expected violation set on every scenario.
Pricing
Start free. Pay when it pays.
Generalist ships today. Domain Expert adds vertical memory + priors. Custom is built to your stack.
Free
Priced after a short conversation
3 scenarios a month, static-only (no live runtime monitor). Public evidence.
Team
Priced after a short conversation
Unlimited scenarios, live runtime monitor, custom invariant sets.
Governance
Custom — Priced after a short conversation
On-prem/VPC, signed packets chained to your ledger, dedicated verifier, 24/7 support.
