← Eichor · All briefs

Private Reasoning — your model, your walls, our verified facts

Your questions never leave your network. Our verified facts come to them. Eichor Private Reasoning is a deployment pattern for teams that cannot send pipeline, portfolio, or competitive questions to any cloud AI: your own local model, running inside your own walls, reasons over Eichor's deterministic clinical-trial intelligence — wrapped in the same anti-hallucination harness we run in production ourselves.

The twin architecture

• Eichor supplies the facts — per-entity risk scores, feasibility grades, geographic risk, the Overdue-Results Index — every figure deterministic, sourced, and served over a simple keyed API.
• Your local model supplies the reasoning — any on-prem model (Llama, Mistral, your fine-tune) asks the sensitive questions. The question text, your documents, your context: none of it is ever transmitted — not to us, not to any model provider.
• Our harness supplies the trust — the evidence-quote gate: your model returns each claim with its own verbatim span from the facts, and the gate checks every span against the fact values before the answer reaches your team. A claim without evidence does not get downgraded or flagged — the answer is withheld. The gate fails closed. (This is the exact discipline our own overnight extraction pipeline runs on — we sell what we use, including the part where it refuses.)

What leaves your network — the honest table

The harness — what you receive, and what it does

The whole client is about forty lines; the value was never the plumbing. It is provided with access rather than posted publicly, and it does three things worth stating plainly, so your security review can assess the design before anyone downloads anything:

• The model answers in a structured form — each claim carries its own evidence span. “No evidence” is therefore a state the gate can detect, not a sentence it has to interpret.
• Evidence is matched against the fact values, never the JSON envelope — so punctuation, key names and braces cannot pass as proof.
• A failed check raises and withholds the answer. It is not an assert (which Python strips under -O, silently removing the protection in production) and it is not a warning appended to an answer your team has already read.

How you verify it — on your hardware, not our word

With access you also get a self-test that runs entirely inside your network. It asks your local model two questions about a real trial and reports three things: an answer where every claim carries verified evidence; a second question the data cannot support, where the answer is withheld; and a log of every request that crossed your network boundary, so you can confirm the question text was not among them. It exits non-zero if any of the three fail, so it works as a check in your pipeline, not just a demo.

Both files are fetched from an authenticated URL on this domain using your own key — we do not email code, and you should be wary of anyone who does. Source is provided in full for your review before you run it; nothing is obfuscated or compiled. If you are evaluating and want to watch it run first, we will walk through it live.

Who this is for

Sponsors, CROs, and buy-side teams whose security or legal review has said "no cloud AI for pipeline questions" — and who still want current, verified trial-risk intelligence in their AI workflows. Delivered as: API access sized for bulk slices, the harness kit, an integration session, and support. Enterprise terms; pricing by scope.

Questions: hello@eichor.com · Related: the machine channel · watch our own local model work the rails