AIZOYA

ReleaseProof Recovery by AIZOYA Independent evidence before a software repair advances. A repair can look convincing and pass one visible example while still violating its contract. Built during this hackathon, ReleaseProof is an independent verifier for a bounded software-repair task. It checks two arithmetic functions against a frozen contract using eight public fixtures and 32 fresh challenge cases. It records candidate and suite identities, the replay seed, inputs, outputs, verdict and timing. Candidate expressions are parsed through a strict allowlist; arbitrary Python is not executed. Failed or inconclusive candidates are blocked. Working integrations: Meta Muse Spark 1.3 generated a repair that automatically passed all 40 checks. Subquadratic returned code wrapped in prose: the original response was blocked, then the unchanged code passed 40/40 after manual formatting cleanup. Those records remain separate. MongoDB stores synthetic verification evidence, with independent readback, integrity checking and deterministic replay. The packaged local source passes 136 Python tests with optional dependencies installed and no skips, plus 14 JavaScript timing tests. These results do not imply every local module is deployed or establish general repair accuracy. The verifier is the core product; its web interface displays the evidence. Hosted acceptance is disabled. A separate local workflow revalidates evidence before changing a local acceptance record. Nothing automatically merges or deploys. The hosted verification preview currently requires Vercel sign-in. The public brief provides the evidence and downloadable source: https://releaseproof-recovery-brief.aizoya-ai.chatgpt.site/ Fetch uAgents/ACP and Gemini challenger components are offline preparation; live Agentverse/ASI:One discovery and a Gemini call are not verified. Event-built code, tests and verification functionality are separate from earlier concept materials. OpenAI Codex assisted implementation, tests and documentation under Franklyn Rosario's direction. No private customer code or data is used.