Pillar 3 · Epistemic-State NLI
Impeachment Engine
“Finding the needle in 10,000 hours of testimony.”
The Impeachment Engine utilizes cross-encoder Natural Language Inference to autonomously cross-reference sworn deposition transcripts against the entire historical discovery record. It does not summarize; it flags epistemic contradictions in real-time, instantly surfacing hidden truths buried in the documentary record.
How It Works
Extract the temporal anchor date from the sworn claim — the moment the deponent was speaking about, not when they said it.
Generate the inverse hypothesis — the proposition that, if true, would contradict the claim.
Run a deterministic graph query bounded by custodian access and temporal scope. Only documents the deponent personally had access to before the anchor date are eligible.
Bi-encoder retrieval ranks the candidate set by semantic similarity to the inverse hypothesis.
Cross-encoder NLI scores each candidate for entailment. State-update markers (legitimate evolutions of position) are filtered out.
The surviving candidates are ranked by entailment confidence and presented with full geometric provenance.
Technical Detail
cross_encoder_nli(inverse_hypothesis, candidate) → entailment_score // Only after temporal + access gates pass
APEX CORP v. HORIZON LLC · Hayes Depo Vol 1, 142:18-22
Sworn: "I was unaware of any thermal defect..." → Found: PLAT-SLK-00942 (Slack DM, Aug 4 2025): "The Gen 4 thermal benchmarks are a disaster." — 98% NLI confidence, predates claim by 69 days.
98%
NLI confidence on Hayes contradiction
69
Days the evidence predated the sworn claim
< 4s
Time to surface from 1.2M document corpus