Evidence-gated claim admissibility for autonomous coding lifecycles. Lifecycle states like reviewed, tested, and done are claims — Proof-or-Stop admits them only when fresh, code-state-bound, mechanically verifiable evidence satisfies the gate. No qualifying evidence → repair, degrade honestly, or stop. Never advance on the agent's word.

plan ──▶ execute ──▶ review ──▶ reflect ──▶ done
  ▲                                          │
  └──────────  proof, or stop  ◀─────────────┘

Proof-or-Stop: Don't Trust the Agent, Trust the Evidence — Loop Engineering for Verifiable Evidence-Gated Lifecycle Control Jek Huang, Jeffery Hsia, Jiayi Sun, Freddie Shi, Wei Huang, Ian H. White arXiv:2607.14890 [cs.AI] · DOI

Selected results:

• Zero false DONE across the engine-contract stress suite, and all 18 receipt tamper classes rejected with zero false accepts
• 9,240-cell powered ablation — enforcing the gate cuts amplified test-passing-but-wrong deliverables from 31/1800 to 2/1800 (pre-registered primary comparison; CI excludes zero), at near-identical token spend to advisory review
• Self-application corpus — 565 stories / 1,007 findings; 93% of deep-set findings surfaced by independent review while the author's own tests were green

• Paper: arXiv:2607.14890 — the method, its formal admissibility model, and the full evaluation
• Code: the Proof-or-Stop implementation (the evidence-gated lifecycle engine, its gates, and the full experiment corpus behind the paper) is being prepared for public release and will open here soon; the tag paper-arxiv-v1 marks the code state at arXiv v1 submission

@misc{huang2026proofstop,
      title={Proof-or-Stop: Don't Trust the Agent, Trust the Evidence -- Loop Engineering for Verifiable Evidence-Gated Lifecycle Control},
      author={Jek Huang and Jeffery Hsia and Jiayi Sun and Freddie Shi and Wei Huang and Ian H. White},
      year={2026},
      eprint={2607.14890},
      archivePrefix={arXiv},
      primaryClass={cs.AI},
      url={https://arxiv.org/abs/2607.14890}
}