Orca — Vincent Aliquo

Confident, Cited, and Wrong

A language model blends two kinds of answer into one confident sentence: what it guessed, and what could have been proven. Citations don’t fix that; checking that a source exists doesn’t check the claim. In a seven-question trial with one small local model, three answers had fully valid citations—and all three failed semantic review.

Side-by-side comparison of a cited-source check, which only confirms sources exist, and Orca's check, which verifies every claim — Checking that a source exists does not check the claim.

The Proposal Is Not the Answer

Orca makes the proposal and the answer two different things. A proposer, whether a model or a planner, never supplies a value of its own. It makes typed claims that point at evidence and name calculations. Orca checks every claim, accepts only what it can derive, and attaches a receipt. Rejected claims come back with a typed reason, and the final wording is filled in from the checked result—so a sentence that slips in a new number fails.

Four steps: propose typed claims, Orca checks every claim, an accepted result, and wording Orca permits, with typed rejection reasons below — The proposal is not the answer. The check is.

A claim can rest on exactly three things: it can copy an established fact word for word, run a registered and versioned calculation, or say “none” over a list that the evidence declares closed. Nothing else parses.

Three derivations: extract an established fact, compute with a registered calculation, and absent, meaning none over a closed list — Three ways a claim can be supported.

What It Does in Practice

When the evidence is incomplete, Orca says what is missing instead of guessing. When a rule an answer relied on is withdrawn, the answer reopens as historical, and its receipt stays unchanged. When an input is missing, it asks for it.

Two Orca console results for two pumps: one answered with its limits, one that cannot be determined yet and names the reading it needs — Same question, two assets: one answer, one named gap.

Orca console before and after a maintenance rule is withdrawn, with the result reopened as historical — When the evidence behind an answer is withdrawn, the answer says so.

Six Orca console outcomes: answered with its limit, missing reading named, input asked for, conflict surfaced, marked historical, left unresolved — Six answers a guessing system would get wrong.

A free-text request becomes a proposed scope, and nothing runs until a person confirms it.

Orca console model-assistance panel showing a proposed scope that has not run, and a request no registered question fits — A proposal is reviewed before anything runs. The proposals here come from a scripted fixture, not a live model.

I test it the way an adversary would. Fifteen hand-written invalid claims, of the kinds a model produces, were all rejected for the expected reason. Change a single value in an accepted result, and its receipt stops verifying.

Terminal output listing fifteen invalid claims, each rejected with a typed reason — Fifteen invalid claims in, fifteen rejected, each with its reason.

Terminal output where a versioned calculator answers one request with a receipt and declines others by naming what is missing — Numbers come from a versioned calculator. Gaps are named, not filled.

invented claims rejected: 15 / 15
repeated questions, identical receipts: 32 / 32
objectives matched by an independent oracle: 3 / 3
median warm answer: < 5 ms

Runs Where the Evidence Is

The check runs no AI model and needs no GPU. It is a small local service that binds only to the machine it runs on and answers only callers that hold its key. The whole Windows runtime is 61 MB, starts in 0.6 seconds, and has one third-party library.

Deployment figures: zero model weight files, 61 MB on disk, 0.6 s start, under 2 s to load a 310 MB assessment, loopback only, one third-party library — Runs where the evidence is.

Any proposer meets the same check. Behind it, I built a small in-house language model—about 300 million parameters, trained from scratch on one RTX 3080. Today, the default proposer is Orca’s own deterministic planner.

Where It Is Used

I started Orca in August 2026 to support HADES’s onboard autonomy work, where it is an advisory planner with no vehicle link and no keys; its candidates need explicit operator action. In October 2026, the University of Dayton announced a two-year, $500,000 National Security Agency grant to build a K-12 curriculum that teaches AI literacy and cybersecurity through game-based scenarios. Haiku is a named partner, and access-gated Orca is planned to power the project.