Skip to content
HONESTASDecision-Evidence Operating System

For the General Counsel

Was this loss inside the wording? Part of that question is decidable — and nothing tells you which part.

The rest is judgement, and you are asked to stand behind coverage positions produced at machine speed. The uncomfortable part is not that a model answered; it is that the answer arrives with the same flat confidence whether it turned on an arithmetic sub-limit or on what a court will one day decide “reasonable precautions” means. Those two answers deserve different levels of trust, and nobody is separating them for you.

A certificate that says which lane it used, and where that lane stops

Lane A — formal proofdesigned, not built(designed, not built)

The wording, the endorsements and a binder’s delegated authority compiled into a formal encoding, so decidable questions — is this inside the pen, does this sub-limit bind, does this notification deadline fall inside the period — return an answer with a justification tree a machine can re-check.

Lane B — the cross-model gateshipped(shipped)

Everywhere lane A does not reach — which today is everywhere — heterogeneous models re-derive the answer and a computed predicate has to pass before anything is released. It is a weaker guarantee than a proof, and it is the one that is actually running.

Decision certificateSealedLane A unavailable for this question
question
is the loss inside the insuring clause as endorsed?
lane used
B — cross-model agreement
lane A status
not attempted: the wording is not formally encoded
what is proved
nothing is proved; two independent derivations agreed
what is not proved
the meaning of “gradual deterioration” in this forum
open-textured terms
listed, not silently resolved — each is a human input
scope
tenant · owner · business unit
seal
sha256:1d80be3c…77af (illustrative)

[O] Illustrative composition, built from the components the product ships. No customer data appears anywhere on this site; the seal above is a placeholder, not a real digest. The document intake this lane depends on is observed at NX/services/nexus-fileprocess/api/src/orchestration/SandboxFirstOrchestrator.ts and NX/services/nexus-graphrag/src/processors/ocr/ocr-cascade.ts. The formal lane itself is not observed anywhere, because it does not exist yet.

The mechanism, in three lines

Some questions about a policy wording are decidable. Where they are, the answer carries a machine-checkable proof; where they are not, the cross-model gate covers the gap — and the certificate says which lane it used.

First

The decidable part is separated out

Sub-limits, deductibles, periods, notification deadlines and delegated-authority conditions are arithmetic and set membership. They can be encoded, and an encoding can be re-checked by something other than the thing that produced the answer.

Then

The undecidable part is named, not absorbed

Open-textured terms are represented as inputs a human supplies, so the proof is honestly conditional: if the loss is a flood in the legal sense, then cover follows. Anything stronger over-claims, and we would rather print the conditional.

Finally

The certificate declares its lane

Every certificate says which lane answered, what was proved, and what was left to judgement. You can read the guarantee you are actually being offered instead of inferring it from a logo.

designed, not built(designed, not built)This is one of the four pillars that leads with the unbuilt glyph. The lane that backs it up is real; the lane the page is named for is not.

5open problems that bound what any proof over an insurance wording may claim[D]
1bn/dayformal queries answered in a production compliance product outside insurance[V]
2lanes, and the certificate always names the one it used[O]
0wordings formally encoded today — the compiler is not built[O]

The honest limit