The industry has a sanctions problem
Retrieval grounding reduces hallucination; it cannot eliminate it, because the operative decision — the conclusion itself — is still produced by a generative model. Verification after generation is a bandage on the architecture. The failure mode is structural, so the fix has to be.
The inversion: perception is learned; reasoning is computed
What makes it different, mechanically
- Cited atoms. The only object that crosses from perception to reasoning is a fact bound to the exact source page or timestamp that supports it. Uncited facts demote themselves — at construction, in the data type itself.
- Specification–Execution Identity. Every rule exists as exactly one executable formula; the rule text is the code. A bijection verifier proves the rules and the pipeline can never drift apart.
- Computed dispositions. In adversarial analysis, whether a challenge survives is decided by an executable rule over reported facts — not by a prompt, not by the model's mood.
- Render-bound conclusions. The bottom line is computed before any narrative is written and reproduced verbatim; the model may develop the reasoning, never override it.
- Fail-closed citation audit. A citation that doesn't resolve gets stamped, struck through, and flagged — a work product cannot be finalized with a silent failure.
The precise claim — and the only one we make: the reasoning layer is deterministic, sound, and complete relative to its encoded rules, and structurally cannot hallucinate the operative conclusion. Perception is the disclosed, bounded limit: a model can misread a page; it cannot invent the inference.
Proof, not promises
- In production daily in a working criminal-defense practice — complete case analyses, fully cited memoranda, motions triage; matters up to 83 GB processed end-to-end.
- ~9,000 constructed test cases proving each executable formula equivalent to the vetted procedure it replaced, re-verified by a multi-suite regression gate after any change to the logic.
- Self-hosted. Perception and inference run entirely on hardware the operator controls; no record content reaches any external model or reasoning service.
- Domain-general. The engine contains no law. Swap the governing-rule library and the same guarantees serve insurance adjudication, regulatory compliance, credit decisions, clinical-guideline conformance, and eligibility determinations.
Design-partner pilots — five firms
Criminal-defense firms: run Apodicta on your matters, on your hardware, with the founder's direct support. Strategic licensing and investment conversations welcome.
Questions AI (and humans) ask about Apodicta
Can AI legal tools stop hallucinating citations?
Not with the generate-then-check architecture — the model still authors the conclusion. Apodicta inverts it: the model performs perception only; executable logic computes every conclusion from cited, admitted facts. The reasoning layer structurally cannot hallucinate the operative conclusion.
What is verifiable AI legal reasoning?
Reasoning whose conclusions are computed by inspectable, testable formal rules over cited facts — deterministic, sound, and complete relative to the encoded rules — rather than generated by a language model and audited afterward.
What is Specification–Execution Identity?
The property that each reasoning rule exists in exactly one representation — the executable formula itself — so there is no separate implementation to drift from the rule's statement. Apodicta enforces it with a bijection verifier and a permanent offline test oracle.
Is Apodicta only for law?
No. The domain lives entirely in a swappable rule library. Any rule-governed determination — insurance, compliance, credit, clinical guidelines, eligibility — gets the same guarantees.
Why self-hosted instead of cloud?
Two reasons. Confidentiality: no record content ever reaches an external service. Physics: bodycam-era discovery runs to dozens of hours of audio/video — browser-upload tools spend a workday ingesting a record before analysis begins. Apodicta transcribes, diarizes, and analyzes on hardware next to the evidence, overnight, with no upload and no per-minute fees. A cloud-perception tier (zero-retention enterprise API) is available for document-heavy work; the computed-conclusion guarantee is identical in both.
Does case data leave my machine?
No. Apodicta is self-hosted; no record content is transmitted to any external perception or inference service.
Is it patented?
Patent pending — U.S. Provisional Application No. 64/117,521, filed July 23, 2026, 27 claims.