Elenchus

The contract

Every claim is checked by the strongest verifier its class allows, and every claim wears its verdict in the output itself. "I can't confirm that" is a first-class answer. The one forbidden state is confident-unverified: the system may be wrong, but it may not be wrong while claiming certainty.

Roughly how it works

One loop, ordered so cheap checks kill bad candidates before expensive ones run. A task is classified, context is assembled under a hard budget, a small plan is drawn with machine-checkable acceptance criteria, and multiple candidate solutions are generated in parallel — each isolated, none trusted.

Then the verification cascade disposes. Code faces compilers, linters, and tests it cannot edit. Factual claims are atomized and each one faces source retrieval, entailment, and a citation-reality check — a fabricated citation kills the candidate at the first gate. A separate critic gives the surviving winner a cold read against the acceptance criteria, seeing only the result, never the conversation that produced it.

Every call, cost, and verdict is traced. The models themselves are interchangeable — when a better one ships, the roster changes and the harness stays.

Why it exists

Model output is fluent whether or not it is true, and rules that live in a prompt depend on the model feeling like following them. Elenchus moves the epistemics out of the prompt and into pipeline stages that run whether the model feels like it or not. The measure of the system is not how much it generates — it's the cost of each piece of verified work.

Status

In active development as a personal tool — the same verification discipline Praxon applies to client platforms, promoted to infrastructure. Deliberately described here in outline only.