Provenance

checking…

A fleet of agents that detects, diagnoses, fixes, and verifies enterprise incidents, and writes what it confirmed into a governed institutional memory. An LLM never decides what the organization does, and never decides what the organization believes.

What the fleet believes

1

Belief inspector

stored state · computed confidence

What the organization currently believes about an entity — with the evidence, the computed confidence, and every earlier version it replaced.

2

Registry panel

stored state

Who the fleet trusts right now. Standing is re-read on every request, so a downgrade takes effect on the next action, not the next deploy.

3

Counterfactual panel

measured offline

Is memory worth it? Measured across twelve live incidents — including the result that came back against the design.

First time here? Paste the token from the README's “Run the demo” section and press Wake the fleet — the fields are already set to the supply-chain incident, which holds for human approval and changes nothing. Incident #1 (inventory-api / error_rate / 0.38) executes a real rollback and needs a fault injected first from a checkout (scripts/inject_fault.py).

An incident, live

4

Live fleet view

model output

The reasoning chain, streaming: which agent, which model, what it concluded, what it recalled.

hyp = hypotheses the agent considered (model-asserted) · tok = input+output tokens · recalled = beliefs memory handed the agent before it reasoned
5

Approval card

deterministic code

Every held action, in plain language, for a non-engineer. Generated from the risk table — never from a model.

6

Gateway ledger

deterministic code

Every decision the gateway signed, with its risk arithmetic — approvals, holds, and denials alike.