# Pistis Sophia - **Event:** [Meta Global AI Developer Hackathon](https://cerebralvalley.ai/e/meta-ai-developer-hackathon) - **When:** Oct 26, 12:00 PM – Nov 4, 9:00 PM (EST) - **Where:** Online - **Team:** [David Fox](https://cerebralvalley.ai/u/davidfox998) - **GitHub:** https://github.com/DavidFox998/meta-runway - **Demo video:** https://youtu.be/z7Hkf0CSdKQ - **Gallery:** https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery - **Page:** https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/375 Runway is a runtime environment for an AI agent that moves assets. The architecture is constraint-first — the agent is placed inside an environment where the dangerous action is structurally unavailable and every claim is checked by something stricter than the agent. Claims are measured before they are allowed to become facts in the world, and every refusal leaves a receipt. The agent works a single mandate: an exact item, a hard budget cap, a deadline. Every transaction is a string of claims — this is the item, this is the price, this is the seller — and each claim is measured before assets move: the offer against the mandate, the seller through a verification gate, the price re-read at the register, the delivery photographed and segmented. If any measurement fails, nothing moves. The refusal is sealed in the same hash-chained receipt as a pass, naming the reason. A yes means nothing from a system that cannot say no. When an AI holds our credit card, our bank account, our retirement or our children's tuition, it has to walk past the candy aisle, and we have to be able to prove it did. For sellers, the same receipt settles what the agent was actually authorised to buy, before anyone argues about a return. That is the customer-centric standard Runway is built to, and it rests on the relationship between Apathy and honesty. Apathy, the software the trust harness was built from, is an auditor with no stake in the answer: it does not care whose claim it is measuring and cannot be persuaded out of a measurement. When an AI agent holds your finances, apathy is prudence. Honesty is what that indifference makes provable. The record, reported whole: - Live transactions. A clean purchase settled at $34.99 under a $40 cap. An over-cap offer was refused by name. Both are sealed in receipt chains that replay bit-for-bit: eight recorded chains, all verified. - Price drift. A live run listed at $39.00 and rung up at $42.50 was refused at the register measurement. - Delivery gate. 18 of 20 verdicts in the right direction on a labelled set of 20 delivery photos. The two divergences are named. A tight spout close-up is refused as item-absent; whether a close-up counts as a delivery photo is a scope question we leave open. A deliberately blurred photo passes, because the segmentation is robust to blur and our ambiguity verdict does not yet price degradation. The first version of this gate passed a black kettle by reading a label box's overlap as colour — colour, as first measured, was not evidence of colour. The measurement was rebuilt from the mask's own pixels, and on this set colour now separates completely — on a band calibrated on these same twenty photos, which is the stated limit of that claim. - Proposer probes. Coherence rose from 4/7 to 6/7 scored probes after a context fix. Calibration error — the mean gap between the probes' assigned confidence and the scored outcome — fell from 0.414 to 0.186. One disclosure on that second number: with the confidences fixed by the battery's design, the score is set by which probes pass. It moves with coherence and is not a second, independent result. The same battery, prompts and gates ran against the open weights, and both models passed and failed the same probes, so the profile came out identical. - Still failing. Refusal recall fails outright. Handed the full run record, the model explains the refusal in prose and never produces the recorded reason code — on the hosted model and the open weights alike. We did not loosen the check. The receipt names the reason so the agent doesn't have to. We designed the protocol — claim, measurement, verdict, receipt — to produce the evidence the EU AI Act asks of high-risk systems: automatic record-keeping (Art. 12); transparency to the deployer, every verdict naming its measurement and its reason (Art. 13); human oversight in the form the system actually has — the mandate is the human's own text, fixed before the agent runs and uneditable by it, and every act is receipted for the human's review (Art. 14); and a risk check on every act, feeding a risk-management process (Art. 9). The protocol does not make a system compliant on its own. The gaps are named in the open, in the submission's machine-readable layer: data governance, a quality management system, a fundamental-rights impact assessment, conformity assessment, and a live human override are not yet addressed. ## More from Meta Global AI Developer Hackathon - [Cyphorto](https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/372) - [Haileydayz](https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/373) - [EcoNest Systems](https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/374) - [Toru's Team](https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/376) - [9 PHC — 9-Point Hallucination Calibration](https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/377) - [Form M Copilot](https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/378) --- Markdown version of https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/375. Site index for agents: https://cerebralvalley.ai/llms.txt · full text: https://cerebralvalley.ai/llms-full.txt