Bechtel adopted "zero accidents" as its safety goal, gives every employee stop-work authority, and posts incident rates near the industry floor. This concept adds the layer under it: machine-verified proof that a specific task was done safely, meaning hazard-gated authoring, proof-of-safe-completion, and a settlement receipt on a public chain. On megaproject scale, per-task verification is where that proof is worth the most.
What we found: Bechtel's safety record is genuinely strong, but nothing public answers "was this specific task done safely?": no agent front door, one generic schema block, and safety commitments that live only as page copy.
What we built, free: a machine-enforced hazard gate on a live work library, a proof-of-safe-completion demo, and an agent-ready kit (llms.txt, agent card, MCP tools; MCP is the protocol that lets an AI assistant call your tools), all reviewable below.
The ask: a $750 packet where it counts, or a 10-cent automated payment call to try the live rail today. Inbound-only, no outreach.
bechtel.com/llms.txt (the file AI assistants read first), ai.txt, and every .well-known/ agent path are true 404s (soft-404 control confirmed). Zero agent front door on a top-2 U.S. contractor./approach/ safety+ESG hub each carry exactly one JSON-LD (the invisible structured labels search engines and AIs read) block: the generic Yoast SEO graph. No Organization, no project, no ESG dataset an agent can verify./approach/, none machine-readable.Baseline credited first (the safety record is real and strong). Then four evidenced findings: the missing agent door, the illegible commitments, the self-attested subcontractor path, and one hygiene note, each with the exact URL, HTTP status, and re-check command.
Read the briefTwo acts in Bechtel's own slate: Act 1: an owner's AI asks "prove this task was done safely" and gets a PDF hunt. Act 2: the same task as a work packet that cannot be authored without passing a machine hazard gate, closed with photo+GPS+signature proof, settled to an on-chain receipt that the ESG record and the sub's safety file both cite.
Run the demoConcept files: an llms.txt that makes the safety program and project data discoverable, an agent card (the file that lets an AI assistant use you correctly), MCP tool schemas for verified safety history and ESG receipts, and a concierge persona. Drafts, ready to hand to Bechtel's own team.
construction_site_safety_walk and silica_dust_exposure lanes; proof-checked packets (photo+GPS+signature); on-chain settlement on Base: EWP contract 0x76c17C…A14B, first settled packet workId 14; agent-payable tools over x402 (a web-native way for AI agents to pay per use: work-packet $0.10 · work-proof-check $0.01); and a live end-to-end agent walkthrough: the building-agent golden path.
work-packet program answers HTTP 402 with a standard x402 payment schema; the agent signs a USDC authorization and pays $0.10. A runnable client with a dry-run default is published at vealth.net/recipes.html.Why this isn't scary: the agent cannot do anything invisible here. Every action embeds its own receipt: permanent, public, verifiable. That is more accountability than most enterprise software offers, not less, exactly the property a zero-incidents culture wants.
The bet: AI competence is scaling, and Bechtel has already bet on it, standing up an SVP of EPC transformation to rebuild project delivery with AI, automation and robotics, and partnering with NVIDIA to modularize gigawatt-scale AI-factory data centers. This is the safety-verification rung of that same ladder. Being early costs $0.10 and one afternoon. Being late means the proof-verified safety track records belong to someone else.
work-packet x402 call today (vealth.net/.well-known/x402, contract 0x76c17C…A14B); the receipt lands in seconds. Inbound-only, no outreach. [email protected].
