🧮 Epistemic / Truth Gates

Tier 3 — do-calculus gates: is the claim true or supportable? No LLM in the path.

Tier 3 · #11

Identifiability Gate

Given a causal DAG and P(Y | do(X)) — is the effect identifiable from observational data? If yes, which criterion (backdoor / front-door / empty-set / instrument) + estimand. Engine: ananke OneLineID (Shpitser–Pearl).

DAG: {Z→X, X→Y, Z→Y} · query P(Y | do(X))

DAG: {X→Y, X↔Y latent confounder U} · query P(Y | do(X))

Tier 3 · #12

Fabrication Gate (two-stage)

Stage 1 (cheap): self-consistency / disagreement across sampled answers. Stage 2 (symbolic): flagged claims adjudicated against the formal causal oracle — claims the oracle refutes are DENIED, ground truth returned.

Claim: "drug X causes recovery (p<0.01)" · 3 samples agree

Claim: "feature F causes churn" · samples disagree (2 yes / 1 no)

Tier 3 · #16 / #17

Domain Gate · Flip-Rate Gate

Domain Gate: every domain JSON admitted only if structurally valid, internally consistent, and provenance-intact. Flip-Rate Gate: did new causal edges actually improve understanding? The domain's own causal model is the oracle — it tests itself.

new edge: "stress → insomnia" (source cited: PMID 12345)

new edge: "insomnia → stress" AND "stress → insomnia" (both asserted, no source)

The design point

Generate → Verify → Persist → Reuse

The generation side varies (LLM, templates, operators); the knowledge representation varies. What never varies is the gate: a mechanical check of a candidate against ground truth. These gates are why a claim can be structurally verified or refuted without asking a model to grade itself. Abstention is a first-class outcome — "no causal path" is a real answer, not a failure.