VetIOS ProofLoop

Every verified outcome should become a test before it becomes a pattern.

VetIOS ProofLoop is an outcome-to-code reliability compiler for high-stakes AI. It connects a model's inference trace to what happened later—lab evidence, clinician confirmation, specialist review, and follow-up—then converts that closed episode into a tamper-evident Outcome Receipt, an executable evaluation, and a model-release gate.

The problem

AI observability can show what a model said and which tools it called. It usually cannot prove what happened in the real world afterward. In clinical systems, thumbs-up feedback and LLM-as-judge scores are not ground truth. The scarce signal is a longitudinal, provenance-linked outcome.

Veterinary medicine is the first wedge: evidence is fragmented across clinics, labs, imaging systems, and follow-ups, while wrong or overconfident outputs can be consequential.

What ProofLoop does

1. Links one inference to de-identified outcome evidence and human confirmation.
2. Produces a hash-addressed Outcome Receipt with source lineage and review state.
3. Uses GPT-5.6 through the Responses API to classify the failure, identify affected population slices, and emit a strict evaluation specification.
4. Invokes Codex in the target repository to create the regression fixture and test, run the relevant suite, and return an auditable patch.
5. Blocks model promotion until the new candidate passes the outcome-derived gate.

ProofLoop is designed to export tests to existing evaluation systems rather than replace them. Horizontal observability platforms capture traces; ProofLoop supplies the outcome-verified ground truth they do not own.

Why GPT-5.6 and Codex are essential

GPT-5.6 reasons across heterogeneous longitudinal evidence and turns it into structured failure semantics. Programmatic Tool Calling can query the inference trace, outcome ledger, and review evidence with bounded control flow; strict schemas preserve the evidence needed for final judgment.

Codex is the repository-aware execution layer. It can inspect the actual model adapter and test conventions, generate a targeted regression fixture, run checks, and produce a reviewable patch instead of a generic recommendation.

Three-minute demo

A veterinary model gives a confident but wrong differential. A later lab result and clinician sign-off close the case. ProofLoop creates the Outcome Receipt, GPT-5.6 explains the failure mode, Codex writes and runs the regression test, and the release gate rejects the old candidate before allowing the corrected candidate to advance.

Build Week scope

VetIOS existed before OpenAI Build Week. ProofLoop is a new, separately documented extension built after the July 13, 2026 submission-period start. The repository will identify the pre-existing baseline and the new ProofLoop commits, tests, UI, and documentation so judges can evaluate only eligible work.

Safety and privacy

The public demo uses synthetic or de-identified cases, requires human confirmation for clinical outcome claims, and does not replace licensed veterinary judgment.

Built With

• codex
• codex-sdk
• gpt-5.6
• next.js
• openai-responses-api
• programmatic-tool-calling
• supabase
• typescript