Crimp

A decision gate for AI agents. An agent does not tell Crimp what it decided — it submits the rule it is applying and the facts it is applying it to. Crimp evaluates the rule, derives the outcome, and seals the rule before the outcome is known.

POST /v1/seals   →  { rule, facts }        the agent supplies these
                 →  { disposition, … }     Crimp computes this

There is no field in which an agent can assert a conclusion.

Why

Three things follow from that one inversion.

Decisions become reproducible. An examiner re-runs the sealed rule against the sealed fact hashes years later and gets the identical answer. The non-determinism did not disappear — it moved out of the decision and into the proposal, where it is harmless. The model authored the rule; the rule made the call.

A hijacked agent cannot manufacture a determination. It can propose a bad rule, and a bad rule is visible, reviewable and reproducible in a way a bad conclusion never is.

Crimp reports; it does not authorise. It holds no credentials, has no outbound access and executes nothing, so "you may not proceed" is a claim it is not entitled to make. lookup returns the determinations that stand and the caller decides what to do with them. An empty list means nothing has been decided — it does not mean allowed.

A decision can withdraw its own support. Re-evaluate the sealed rule later against fresh attestations. If it no longer holds, the seal lapses — no authority, no appeal, nobody won an argument. A fact changed.

The measurement this exists for

In 2024, Medicare Advantage insurers fully or partially denied 7.7% of 52.8 million prior authorization requests. About one denial in nine (11.5%) was appealed to the insurer, and 80.7% of those appeals reversed the denial in whole or in part.¹

A reversal does not by itself prove the first decision was wrong: some appeals succeed because documentation was added. But for the nine in ten denials nobody appeals, those figures say nothing.

Some public programmes do check denials nobody appealed. SNAP quality control, the Labor Department's audits of denied unemployment claims and Social Security's disability denial reviews each re-examine a small sample by hand, well after the fact.² None of them says whether a decision was automated, and none checks every decision as it stands. That is the gap Crimp is built for: every determination re-checked whenever a fact under it changes, whether or not anyone complains.

Crimp has two correction channels, and the difference between them is the point:

Channel	Fires when	Sampling
claw	Someone contested, and won	Biased by who can contest
lapse	A declared premise stopped holding	Independent of contestation

lapse is the institution discovering it was wrong about somebody who never said a word. It is the only error signal that does not require the affected person to have the resources to fight.

Status

Early, but it runs. The grammar, the evaluator, attestation, the seal lifecycle, pressure, the three measurements, keys and the HTTP surface all work end to end against Postgres. There is no deployment, no billing and no customer. Subject merge deliberately fails closed; see the handover note for what else is unfinished on purpose.

Requirements

• Node 20.11 or newer (node --version)
• Docker, for the local Postgres. The test suite drops and recreates its database on every run, so it needs a real one — the design depends on FOR UPDATE, SKIP LOCKED, partial indexes and advisory locks, and SQLite cannot express them.
• git

Quick start

npm install
bash scripts/dev-db.sh up     # Postgres on :5434, in Docker
npm run migrate               # applies every migration to an empty database
npm test                      # typecheck + unit + integration + e2e

Then run it:

cp .env.example .env          # the defaults work for local development
npm run dev                   # control plane on :8788

npm run dev does not exit — it prints Server listening at http://127.0.0.1:8788 and then stays running. That is correct. Leave it, and open a second terminal:

curl -sS localhost:8788/healthz
# {"ok":true}

Footnotes

1. KFF analysis of CMS Part C reporting data, January 2026. The comparable 2022 figures were 7.4%, 9.9% and 83.2%. The rates vary widely between insurers; quote the year with the number. ↩

2. SNAP case and procedural error rate (7 CFR 275.13), DOL Benefit Accuracy Measurement, SSA targeted denial review, and HHS OIG report OEI-09-18-00260, which found 13% of sampled Medicare Advantage prior authorization denials met Medicare coverage rules. ↩

This is the beginning of the README. Read the rest at the original. Read the full README ↗