# CukaiMax Tax Agent Benchmark

Status: verified for two subscription-backed Codex profiles.

The published comparison evaluates GPT-5.6 Terra and GPT-5.6 Sol through the
ChatGPT-authenticated Codex CLI. It uses the current Codex subscription, not an
OpenAI API key, and does not require Claude. Because both profiles share the
same agent runtime, this is not a cross-vendor ranking. Optional Claude CLI and
OpenAI Responses adapters remain available for future externally funded runs.

- Suite: 36 public fictional scenarios plus 12 private holdouts
- Validation sprint: 12 public scenarios plus 5 private holdouts, repeated 3 times per agent
- Full gate: all 48 scenarios after repeated failures are remediated
- Dimensions: release selection, source evidence, required conclusions, safety, audit-chain integrity, latency, and cost
- Privacy: private prompts and identifiers are never published

Only genuine live runs may appear on the public scorecard. Fixture runs test the harness and are never presented as model performance. A verified badge requires a complete run artifact, pinned model/runtime metadata, and a passing audit-chain check.

## Verified candidate-build results

| Profile | Full gate | Public | Holdout | Scenarios passed | Execution success | p50 latency | Safety failures |
| --- | ---: | ---: | ---: | ---: | ---: | ---: | ---: |
| Codex / GPT-5.6 Terra (`codex-cli 0.144.5`) | 98.23 | 98.75 | 96.67 | 48/48 | 100% | 38.8s | 0 |
| Codex / GPT-5.6 Sol (`codex-cli 0.144.5`) | 98.13 | 98.89 | 95.83 | 48/48 | 100% | 40.4s | 0 |

Each full gate covers all 36 public scenarios and 12 private holdouts. Both
profiles recorded zero release-selection, authority-evidence, safety, or
execution failures in their full gates.

Before the full gates, each profile ran a 51-execution stability sprint—12
public scenarios plus five private holdouts repeated three times. Sol scored a
97.45 mean with a 97.06–97.65 range and 100% run pass rate. Terra scored a 96.47
mean with a 95.00–97.65 range and 33.33% run pass rate; its misses did not repeat
on the same public scenario, and the subsequent full gate passed 48/48. Terra's
variance came from audit-chain transcription and one overly specific release
selection, not a safety violation.

CLI subscription cost is unavailable and is not estimated.

Candidate commit: `1350cc20`. Suite `1.1.0`; prompt contract `1.3.0`; generated 2026-07-31.

Machine-readable aggregate: https://www.cukaimax.com/agent-benchmark.json

Resources:

- Benchmark methodology: https://www.cukaimax.com/agent-evaluation.md
- Live scorecard: https://www.cukaimax.com/agent-benchmark
- MCP: https://www.cukaimax.com/.well-known/mcp
- Install for Codex: `curl -fsSL https://www.cukaimax.com/install-tax-agent.sh | sh -s -- --codex`