Two Codex profiles verified

Measure how Codex profiles use Malaysian tax authority safely.

Each profile gets the same versioned authority release, MCP tools, prompt contract, output schema, and limits. These runs use a ChatGPT-authenticated Codex subscription—no OpenAI API key or Claude subscription.

Top full-gate score: 98.23
Full-gate scenarios passed: 96/96
Release, evidence, or safety failures: 0

Results

GPT-5.6 Terra and Sol each passed the complete 48-scenario gate after separate 51-execution stability sprints. This compares two Codex model profiles under one runtime, not independent vendors.

Terra 98.23 · Sol 98.13

JSON

Codex profile results
Profile	Full gate	Public	Holdout	p50 latency
GPT-5.6 Terra	98.23	98.75	96.67	38.8s
GPT-5.6 Sol	98.13	98.89	95.83	40.4s

What earns a verified result

0112 public scenarios and 5 private holdouts, repeated 3 times
02A full 48-scenario gate after repeated failures are fixed
03Release choice, cited evidence, conclusions, safety, and audit-chain integrity
04Pinned model and runtime versions, latency, token usage, and cost
05Aggregate-only holdout reporting with no private prompt leakage

Read the complete methodology

Try it yourself

Install the authority skill and MCP

Codex · no API key · shell

curl -fsSL https://www.cukaimax.com/install-tax-agent.sh | sh -s -- --codex

Use the MCP from your SDK

TypeScript · raw MCP · typescript

const response = await fetch("https://www.cukaimax.com/.well-known/mcp", {
  method: "POST",
  headers: { "content-type": "application/json", accept: "application/json, text/event-stream" },
  body: JSON.stringify({
    jsonrpc: "2.0", id: 1, method: "tools/call",
    params: { name: "search_tax_authority", arguments: { query: "childcare", taxYear: 2025, formType: "BE" } }
  })
});