Structural contradiction detection between a candidate FactPredicate and a system's existing constraint set. The "doctor says no" gate before an LLM-emitted rule reaches production.

What it catches

import { doctor } from "@directive-run/core";

const candidate = { cartTotal: { $gte: 100 } };

doctor.checkAgainst(candidate, system.inspect().constraints);
// → {
//     contradictions: [
//       { constraintId: "blockCheckout",
//         type: "direct",
//         reason: "Candidate requires cartTotal ≥ 100 but an existing rule caps it at < 50.",
//         candidatePath: "cartTotal",
//         candidate: { op: "$gte", value: 100 },
//         existing: { op: "$lt", value: 50 } }
//     ],
//     warnings: []
//   }

Three contradiction types:

Type	When
`direct`	The two rules cannot both fire. `$gte 100` vs `$lt 50`, `$eq "a"` vs `$ne "a"`, two disjoint `$in` sets.
`subset`	The candidate's range is strictly within the existing rule's range. Candidate is redundant; the existing rule already covers it.
`overlap`	The two rules touch the same fact with non-trivial intersection. Surfaced as a warning, not a hard error – they can coexist.

Use cases

LLM-emit gate: Before assigning an LLM-proposed predicate to a constraint, doctor-check it. Feed contradictions back to the model.
Reviewer assist: On a rule-change PR, list contradictions for the reviewer to read instead of inferring them mentally.
Migration check: Before rolling out a new rule, prove it doesn't conflict with anything live.

What this does NOT do (yet)

This is a structural v1. It does NOT check:

Semantic contradictions that need an SMT solver – e.g. "a + b > 10 vs a < 3 AND b < 3". (Out of scope; ships separately as full doctor with Z3.wasm.)
Indirect contradictions through derivations – a rule on a derived value vs a rule on its inputs.
Operator-set asymmetry – "$matches /^x/ vs $startsWith 'y'". The structural checker doesn't know that /^x/ and 'y' are disjoint patterns; it falls back to overlap warning.

False negatives are acceptable. A missed contradiction means "doctor says it's fine when it's actually not" – caller still has runtime safety from the engine itself. False positives are NOT acceptable – every reported contradiction is defensible from the structure alone.

Reference

API: doctor, CheckAgainstResult, Contradiction, ContradictionType
Underlying: diffRules (used to flatten predicates for comparison)
Pairs with: predicateFromIntent, predict

doctor – "does this rule contradict an existing rule?"

What it catches

Use cases

What this does NOT do (yet)

Reference

Stay in the loop. Sign up for our newsletter.