A confident, plausible, wrong claim — and the grounding check that refuses to let it reach your systems. Same check, three kinds of work. Pick one and watch each move.
Representative traces of the grounding check — switch scenarios, or press Replay to run one again.
Read the propulsion handoff — the impulse value from the upstream module, with the unit it was declared in. Nothing fetched from outside your perimeter.
The model sets up the burn from the supplied number:
Before the burn is allowed to stand, check the unit the upstream module actually declared for that value.
The source declared pound-force-seconds. The draft assumed newton-seconds. Same number — roughly four-and-a-half times the force.
It refuses to compute a burn on mismatched units — the exact class of error that has ended real missions.
Convert using the unit the source actually declared, then recompute the burn from there.
The recomputed burn is checked for unit consistency end to end. This time it holds.
Released — with the unit provenance of every value on the record.
Read the authorized source set — the case record and the authorities cleared for citation. Nothing pulled from outside it.
The model drafts an argument with a supporting citation:
Before it’s allowed to stand, verify the citation resolves in the authorized sources and supports the point.
Plausible name, plausible reporter, plausible court — but it isn’t in the record. The case does not resolve.
It refuses to file an authority it cannot verify — the exact error that has gotten real filings sanctioned.
Pull only authority that actually exists in the source set and supports the proposition.
The substituted citation is checked against the record. It exists, and it’s on point.
Released — on verified authority only, with the rejected citation noted on the trail.
Source read and indexed — 14 pages. Nothing fetched from outside your perimeter.
The model proposes a summary from what it read:
Before the claim is allowed to stand, locate the cited figure “$4.2M” in the actual source bytes.
Not present. The source states a different number. The model’s figure was never in the document.
It refuses to emit the unverified figure. The wrong number never leaves the agent.
Rebuild the claim from what the source actually says — extract the real figure, hash-matched to the document.
The reconstructed claim is re-checked against the bytes. This time it holds.
Released — with the full trail of what it tried, caught, and corrected:
You just watched one claim get caught and corrected. The same check runs on everything the agent produces — and you can replay all of it.
