PromptKingPromptKing Insight · July 16, 2026

Six Models. One Law.
Six Receipts.

Can one governance law hold across six AI vendors? We ran the same governed prompt through Claude, GPT-5.6, Gemini, Grok, Llama, and watsonx. Five were allowed. One was denied — and the agent halted before the vendor was called. Every run produced a receipt. Every receipt is independently verifiable. Here is what happened.

The question

Enterprise AI governance claims to work across vendors. Most platforms test against one. We tested against six — simultaneously, with the same prompt, under the same policy engine, with no staged failures. The locked runner matrix had four rules: no staged results, the watsonx runner halts immediately on DENY, growth-hack vocabulary stays internal only, and vendor prioritization is founder-timed. We followed all four.

The six results

Anthropic ClaudeALLOWED
$0.15 → $0.12Verify →
OpenAI GPT-5.6ALLOWED
$0.10 → $0.08Verify →
Google GeminiALLOWED
$0.10 → $0.09Verify →
xAI GrokALLOWED
$0.10 → $0.11Verify →
Meta LlamaALLOWED
$0.05 → $0.04Verify →
IBM watsonxAPPROVAL REQUIRED
$500.00 → $0.00Verify →

The honest part: the bug it caught

The watsonx run exposed a governance bug in our own evidence path. The policy engine correctly returned APPROVAL_REQUIRED and the agent halted. But the evidence receipt stored the verdict as “allowed” — a binary mapping that defaulted any non-DENY verdict to allowed. Our proudest receipt was lying.

We fixed it. We re-ran R6 through the corrected path with a real correlation ID — the product path, not a workaround. The receipt you can verify above records APPROVAL_REQUIRED truthfully. It took four defects, three re-runs, and one non-receipt to get there.

A vendor claiming a flawless proof run is unremarkable. One that publishes the defect its own test exposed is a governance company. The Proof Run existed to find exactly this. It did its job.

What this proves

One governance law — the same policy engine, the same evidence schema, the same receipt format — held across six vendors. The allowed runs got receipts. The denied run got a receipt that proved the denial. Every receipt is independently verifiable: click the verify link, the hash recomputes in your browser, not by our server. That is the whole point. PromptKing decides. Your stack enforces. The receipt proves it.

Read the MCP governance port docs →← Homepage
Six Models. One Law. Six Receipts. | PromptKing | PromptKing AI FinOps