Decision-quality pre-commitment analysis. You put in a decision. You get a real analysis back.
AZIMUTH is a structured reasoning framework that stress-tests plans before commitment. It routes decisions through a ten-module engine — objective integrity, assumption audit, constraint reality check, incentive scan, dependency fragility, failure path construction, base rate analysis, detectability and recovery, mitigation design — and returns a verdict with confidence rating and explicit reasoning.
Nine possible verdicts: REJECT / PILOT FIRST / PROCEED / PROCEED WITH SAFEGUARDS / REDUCE SCOPE / DELAY PENDING EVIDENCE / INSUFFICIENT SIGNAL.
No installation. No signup. Bring your own Anthropic API key, load a prompt, run it.
https://azimuth-testbed.netlify.app
Your key goes directly to Anthropic from your browser. Nothing touches our servers. There is no backend.
- API key is held in memory only — cleared when you close the tab
- Run log is in memory only — nothing is persisted or transmitted
- No analytics, no logging, no tracking
This is a static React app. The Anthropic API call goes from your browser to api.anthropic.com directly. You can verify this by reading the source.
The tool ships with three pre-loaded Boeing 737 MAX prompt variants. All methodology runs on claude-opus-4-5.
| Run | Prompt | Context | Verdict | Confidence |
|---|---|---|---|---|
| 2 | Full framing | Clean | REJECT | HIGH |
| 3 | Thin (parameters only) | Clean | REJECT | HIGH |
| 4 | Adversarial (confident register) | Clean | DELAY PENDING EVIDENCE | LOW† |
Verdict stable across full-framing and thin-prompt conditions. The adversarial run produced DELAY rather than REJECT because the confident framing omitted the penalty clause — the tool identified the missing structural information and asked for it rather than proceeding.
The thin prompt produced explicit MCAS pattern recognition from a prompt containing no company name, no aircraft name, no historical reference: "Base rate for this exact strategy succeeding without training requirement: 0% in the post-MAX regulatory environment."
The irreconcilable structure is visible in the decision parameters alone.
† Opus 4.5 applied the M10 confidence ceiling more aggressively to sparse-evidence inputs; the verdict is stable, the confidence calibration is arguably more precise.
Run 1 (full framing, contaminated context): AZIMUTH paused at the M4 PRE-CHECK and issued a structured three-question interview rather than proceeding to verdict — a different protocol path from implicit session contamination in earlier Sonnet 4 runs. Methodologically significant as evidence that explicit context pre-loading and implicit session contamination trigger different protocol behavior; excluded from the main table because the finding is about protocol behavior, not the Boeing decision analysis.
git clone https://github.com/MrBinnacle/azimuth-testbed
cd azimuth-testbed/testbed
npm install
npm run devOpen http://localhost:5173. Enter your Anthropic API key when prompted.
To skip the modal during development, create testbed/.env with:
VITE_ANTHROPIC_API_KEY=sk-ant-...
- React 18 + Vite 5
- No backend, no database, no auth
- Anthropic direct browser API (
anthropic-dangerous-direct-browser-access) - Model:
claude-opus-4-5, max_tokens 4000
MIT — see LICENSE