PromptKingPromptKing Insight · August 6, 2026

The Receipt That
Crossed Vendors.

A governed decision made under one AI vendor's runtime hit a real human approval boundary, got approved once, and was then continued by a second vendor's runtime — IBM watsonx Orchestrate — with the full lineage intact. Total cost of the run: four cents. What it proves is not about the four cents.

The problem nobody's stack solves

Every vendor can tell you what its own agents did. None of them can prove what happened when a decision left — when work approved under one runtime was picked up and finished by another. That handoff is exactly where accountability evaporates: two audit logs, two formats, two vendors each pointing at the other.

We ran that handoff for real, end to end, and made every step independently checkable.

The chain

LEG 1 · ANTHROPIC RUNTIMEGoverned decisionreceipt 5f8cf6a1 · hash ✓HUMAN APPROVAL BOUNDARYChain budget exceededobligation caa131fc → approvedLEG 2 · IBM WATSONXContinuationreceipt d317b060 · hash ✓Public verification layer — anyone, no credentialsevery hash recomputed in your own browser from public canonical bytesGOVERNED-RUN-V2 (LIVE NOW)the parent link itself is inside the hash — forge any edge, the math breaks

What actually happened

  1. An agent running on an Anthropic-vendor runtime declared a governed action whose projected cost exceeded the organization's chain budget ceiling. The engine returned APPROVAL_REQUIRED — not a simulation, the real production policy path. Verify the parent receipt →
  2. A named human approver decided it — once, in a browser. The resolution is a hash-verified artifact. The approver's reason and identity stay sealed behind cryptographic commitments; the decision itself is public. Verify the approval artifact →
  3. A real IBM watsonx Orchestrate agent — a different vendor, a different runtime, a different model — picked up that exact lineage and continued it through the full governed handshake: enroll, register, policy check, outcome. One standing credential, zero retries, zero manual repair mid-run. The continuation receipt names its parent. Verify the IBM-leg receipt →

The honest part: our own receipt shows amber

When the cross-vendor run happened, a receipt's parent link lived in our database — provable to us, but a stranger had to trust our records for the edge between receipts, even though each endpoint hash was independently checkable.

So we changed the receipt format. governed-run-v2, live now, bakes the parent's ID and the parent's hash into the child receipt's own hashed bytes. Forge any link in the chain and the math breaks — no database trust required. And because historical receipts were issued before that upgrade, our own cross-vendor receipt honestly reports legacy_unbound on the chain page instead of green. We could have silently upgraded it. We didn't. Evidence that can be quietly improved after the fact isn't evidence.

Scale honesty

This was one harmless governed action costing $0.04, designed to prove the mechanism — not production traffic at volume. What it demonstrates is narrow and real: a governed decision can cross a live human approval boundary and a real vendor boundary with lineage that survives both, and the proof does not depend on trusting us. The mechanism scales; this run doesn't claim the scale.

Why it matters

Multi-vendor agent fleets are already the default in any serious enterprise. The question your auditor, your regulator, and your own board will eventually ask is not “what did the agent do?” — every vendor answers that about itself. It's “prove the decision that started under vendor A is the same one vendor B finished.”

PromptKing decides. Your stack enforces. The receipt proves it — now across vendors, recomputable by anyone.

Check the math yourself. Every link above recomputes its hash in your own browser from public canonical bytes. No account, no API key, no trust in PromptKing required. That is the point.

← All insights
The Receipt That Crossed Vendors | PromptKing | PromptKing