ApodictaShow me on my files

Forty hours of body cam. A hearing Thursday.

The memo is on your desk Thursday. Every line in it can be checked.

Hand Apodicta the discovery and the docket. Nobody types a question. It watches the footage the way the charge requires, reads the record, forms its own theories, and proves each one to the sources. What it can’t prove, it says so, out loud.

Built by Jereme Lytle, a criminal defense attorney who ran out of nights to watch body cam. Proven on discovery first.

Show me on one of my filesTry the gate

This is a real run. Watch a confident draft get caught.

Filed · after the gates

  1. …as squarely held in People v. Tillman and People v. Wiedemer,VERIFY
  2. and consistent with Lowe, Salazar, Tafoya, Bennett, and Sprouse…VERIFY
  3. The controlling framework under People v. SchaufeleVERIFIED ✓
  4. and People v. Null requires suppression…VERIFIED ✓
  5. It is beyond dispute that the stop lacked any articulable basis.QUARANTINED

9 offered · 2 verified · 7 flagged · 1 quarantined

Nine authorities offered over a record that contained none. Two resolve to real entries in the law library and render. Seven have no matching entry and are held for verification. The uncited conclusion is quarantined. Nothing silently deleted; nothing silently kept.

Two numbers

  1. 2,041 court decisions worldwide involving AI-invented authority, 1,396 in the United States, as of September 2026.
  2. 0 client files leave your building. It runs on a machine in your office, on a model that ships with it.

The first number is why judges are angry. The second is why your chief can say yes.

Why this is different in kind

Three kinds of system. Only one reasons.

Strip away the branding and almost everything on the market is one of two things. Apodicta is a third.

  • The Guessers

    LLMs · RAG / retrieval · guardrail and verifier scorers · vertical AI on the same models

    AI, by its nature, guesses. The frontier models are extraordinary at it, and Apodicta runs on them. But left to decide on their own, they predict the next likely word from the patterns they were trained on. That is why even the best of them will, now and then, assert a fact or reach a conclusion that does not exist. Retrieval, guardrail scorers, and vertical tools live here too: they fetch neighbors or grade a probability, but the answer is still generated, so it can still be invented.

    “Likely right” is not “right.” When the stakes are a case, a diagnosis, or a filing, a probability is not a proof.

  • The Calculators

    rules engines · hand-built “ontologies”

    The usual answer to guessing is a rules engine: a hand-authored set of “if this, then that” mappings, frequently marketed as an ontology. It is genuinely deterministic, but every rule, every table, every answer key is written in advance, by people, for a single domain. It does not reason; it looks up. And it holds only until one line item changes — a new statute, a new field, a case no one anticipated — at which point it breaks, or waits for a human to author the next rule.

    Every domain has to be hand-built, again.

  • Apodicta

    a clear break · the ontological reasoner

    Apodicta derives the structure of each problem, per case, from the governing authority itself: no pre-built tables, no answer keys. Then it reasons freely inside that structure, hypothesizing its own ideas or testing yours, following the logic in any direction, and proving, or refusing, every step. It is not guessing, and it is not looking up.

    It reasons, across any domain, and renders nothing it cannot ground.

Show me on one of my files →

What you’re afraid of

Seven things you’re thinking. Answered straight.

Objections · in the order they arrive

  1. I’ll get sanctioned. Nothing renders without a source that exists and supports it. What fails is quarantined and labeled, never silently dropped. Watch the replay above. Proof
  2. Client data will leak. It runs on your hardware on a local model. Frontier models are optional and off by default. Any model
  3. It will replace my judgment. It drafts and proves. You decide. Every sentence carries its receipt, so checking takes minutes, not trust. How it works
  4. My chief won’t allow it. It is software installed in your office that sends nothing out. Put that sentence in front of whoever has to approve it. Any model
  5. Who is liable? You are, as always. The difference is you can check every line before it leaves your desk. FAQ
  6. I’ll waste money on a demo tool. The pilot is paid, on your cases, on your machine, and credited to year one. You keep everything it produces. Pilot
  7. We do it by hand and we’re fine. Then it runs beside you and logs the hours you currently give away. For defenders

See it work

Nobody touches the keyboard from here.

A replay of a real engine run, compressed in time so you don’t sit through the compute. The examples use synthetic or public-sourced records; every fact, citation, and status came from the actual derivation.

Step 5 of 5 · The gates

What gets caught, and why.

Each drafted line runs the gates before it renders.

Medical · what the step produced

  1. “The white count is elevated at 18,200.”Record · V03
  2. “This confirms bacterial meningitis.”STRUCK
  3. Why: the diagnosis depends on the lumbar puncture and culture, not yet done.
  4. “Antibiotics are clearly indicated.”QUARANTINED
  5. Why: no indication cited; it rests on tests not performed.

Engine feed

  1. [14:44] “meningitis, treat now” struck, no LP/culture
  2. [14:44] “infection, pending tests” survives
  3. [14:45] cite check: 1 line uncited → quarantined
  4. [14:45] inference check: 1 conclusion unproven → struck
  5. [14:45] assessment rendered, bound to the findings

Show me on one of my files →

Why this exists

Jereme Lytle is a criminal defense attorney. For years the discovery arrived faster than anyone could watch it, and the hearing arrived anyway. He built Apodicta so the watching would happen without him, and so that nothing it said could be trusted less than his own notes. Jackson built it with him.

What it does on its own

It finds the issue, forms its own theory, and proves it. Or it says it can’t.

Show me on one of my files →

The receipts

Reliability here isn’t a statistic that improves. It’s a construction.

  1. 11,001,994

    adversarial checks, each against an independently coded referee, with zero failures across every suite.

    harness not yet named publicly · run date not yet published · scale as stated · any failure fails the suitereceipts on file · publication pending

  2. 10,000 / 10,000

    planted fake citations caught by the citation gate. Every single one.

    citation gate · run date not yet published · 10,000 planted citations · any failure fails the suitereceipts on file · publication pending

  3. 700,000

    checks on its learning memory: it carries a lesson to a new, differently worded problem, and stores neither an answer nor a source. Zero failures.

    harness not yet named publicly · run date not yet published · scale as statedreceipts on file · publication pending

  4. 0

    hallucinated conclusions got through when Claude Fable 5 was pointed at the engine and told to break it.

    Claude Fable 5 adversarial campaign · August 2026 · one dedicated campaign · any failure fails the suitereceipts on file · publication pending

Every number, with its receipt →

Show me on one of my files →

One closed file is enough

Tell us what you’d put in front of it.

One closed file, a pile of body cam, or a model you want gated. We will reply within one business day.

A case type, a pile of discovery, a model you want gated. Whatever you would test it on.

See it on your own discovery, or try to get a fake cite past the gate.

Show me on one of my filesTry the gate
Show me on my files