13.7 → 100/100: making a corporate site agent-ready without touching it

What an agent actually saw

An ordinary marketing site. To a human it looks fine; to an AI agent trying to use the business it was close to invisible — 13.7/100 (audit v1.0, 2026-07-04). /llms.txt returned the HTML homepage instead of text; there was no agent manifest, no OpenAPI, no MCP endpoint, no typed actions, no sandbox, and no receipts. Layer scores that day: discovery 23.1, comprehension 27.3, action 0, verification 0, state 25.

What was deployed

A sidecar agent surface at a separate subdomain — the main website was not rebuilt. The only origin changes: a real /llms.txt, a 301 for the manifest, and one HMAC-verified backend endpoint (~80 lines of FastAPI). The manifest, distilled docs/context, an OpenAPI document, an MCP endpoint, typed actions, a production-shaped sandbox, and signed receipts are generated and served by the gateway.

The proof: the transcript, not a claim

connected over MCP → tools: list_services, get_service_details, submit_enquiry
sandbox mutation: submit_enquiry → 200 OK, signed receipt
independent verification: VERIFIED ✔ (signed by the gateway's published key)
idempotent replay (same key/input): same receipt id

A third party verified the receipt without trusting the platform — the verifier imports nothing from it. The after: 100/100 (v1.0, 2026-07-04), re-proven 100/100 under audit v2.0 (2026-07-07) including the new Security & Trust layer (HSTS, defensive headers, a declared 429 posture, clean injection hygiene, and a signed surface provenance document that verifies independently), and again 100/100 under audit v3.0 (2026-07-11) including the new GEO checks: a machine-readable schema.org entity graph (served at /structured-data.json, covered by the signed provenance) and confirmation the site doesn't block AI-engine crawlers — the two things that decide whether AI search can identify and quote the business.