decisiongate
Multiple LLMs can agree on a convincing conclusion while sharing the same unsupported assumption, and agreement is not independent evidence.
Every claim is tagged EXPLICIT, INFERENCE, ASSUMPTION, CONTRADICTED, or UNKNOWN, and the engine resolves to a deterministic GO, NO_GO, or HUMAN_VERIFY with report confidence forced to zero by any unresolved critical predicate.
The problem
LLM-assisted decisions fail in a specific way: several models converge on a convincing argument, and the convergence feels like corroboration, when all of them may share one unsupported assumption. Repeating an interpretation does not turn it into a fact. DecisionGate does not ask whether an AI can construct a convincing case for a decision; it asks whether the assumptions the decision requires have survived an explicit attempt to falsify them.
Approach
DecisionGate is a small, inspectable Python engine and CLI that keeps evidence separate from the interpretations built on top of it. Each claim is classified — EXPLICIT (a source says it), INFERENCE (supported by evidence but not stated by it), ASSUMPTION (a bridge the decision needs but evidence has not established), CONTRADICTED, or UNKNOWN — retaining source, location, evidence class, provenance, and confidence. Model output is always MODEL_INFERENCE; it can never create source evidence. The engine finds the predicates a decision needs, constructs plausible inversions of favorable assumptions, and asks what evidence would distinguish the competing interpretations.
Evidence
The final gate is deterministic: NO_GO when independent evidence refutes any critical predicate, HUMAN_VERIFY when any critical predicate is unresolved or contradicted, GO only when every critical predicate is supported by independent evidence. There is deliberately no aggregate fit score that can hide missing evidence — confidence is bounded by the weakest critical predicate, and an unresolved critical predicate forces report confidence to zero. HUMAN_VERIFY is not a generic refusal; it identifies the smallest questions a human or authoritative source must answer before the gate can move. The repo ships worked cases, including a DARPA/SBIR opportunity evaluation where the engine checks whether the authoritative documents actually support the decision to pursue it.