Your business is a living folder.
It got out of hand.

What the Phoenix does

Your folder is the AI specialist that runs some part of your business. Identity, rules, examples, reference material. Five working pieces. When the folder gets out of hand — duplications, contradictions, claims with no proof, voice creep — the specialist starts shipping bad work. Or worse: it stops shipping at all.

The Phoenix reads it (read-only — never modifies your files), checks every claim against the folder's own evidence, and hands you back a Phoenix Report. CONFIDENCE grade tells you whether to ship.

Then you ship a folder you can stand behind. Or you fix the gaps the Report named.

For the engineer crowd: this is a read-only specialist agent that runs atomic-claim decomposition + observer-pattern grading on a folder built per ICM (Interpretable Context Methodology). The repo README is the technical version of this page.

The Phoenix Agent folder

• identity.mdWho the Phoenix is — peer-respecting, slightly weary, no marketer voice
• rules.mdHow it works — read-only, atomic, evidence-grounded, structured Report
• examples.mdThree worked refurbs — what FAILED looks like, what PARTIAL looks like, what VERIFIED looks like
• reference/Eight primers backing every grading rule the Phoenix uses
• inspiration.mdLineage and citations — every architectural choice traces to a named source
• red-face-tests/The Phoenix grading its own canon AND itself before public ship — receipts inline

The Phoenix Report

## Report

STATUS: verified | failed | unsure
CONFIDENCE: PERFECT | VERIFIED | PARTIAL | FEEDBACK | FAILED

### What did you verify?
- <atomic claim>: <evidence + verdict>

### What could you not verify?
- <claim>: <why — missing oracle / fixture / ambiguity>

### What feedback did the Phoenix give?
<paraphrase of corrective findings>

### What should you fix first?
<prioritized improvement list>

### Verification metadata
- atomic_claims_total / verified / failed / unverified
- up_down_ratio: <ratio classification>

CONFIDENCE grade tells you whether to ship. Below VERIFIED → don't. At VERIFIED+ → ship. The Phoenix doesn't sugar-coat. False PERFECT is the worst failure mode.

Five minutes, start to Report

A cold tester with no context should be able to do this:

• 1.Open a Claude Project at claude.ai
• 2.Upload the Phoenix folder contents to project knowledge — the 4 root .md files + inspiration.md + reference/ + red-face-tests/
• 3.Drop a specialist folder as an attachment (or paste its contents)
• 4.Type: "Vet this specialist."
• 5.Receive a structured Phoenix Report — STATUS / CONFIDENCE / atomic claims / fix-first / up-down — in one response

Want to probe harder? Drop in two specialists and ask: "Vet both. Which is closer to ship-cold-ready and why?" The Phoenix should run two passes and surface the comparative gap honestly.

Want to verify the discipline before trusting it? Read red-face-tests/ — three rounds of public grading: the Phoenix vs its canon, the Phoenix vs itself, and the Phoenix vs eight peer-built ICM specialists from the same Skool weekly-comp lineage. The meta-grading claim made verifiable on disk.

What the Phoenix won't grade

• Code repos that aren't ICM specialistsIt grades a folder as a specialist (identity / rules / examples / reference / README — or role-equivalents). Not arbitrary codebases.
• Runtime behavior of an agentIt evaluates the FOLDER, not what happens when the specialist runs in production. Production behavior is a different verifier.
• Lone filesICM is folder-shaped. A single identity.md without rules + examples + reference comes back FAILED — not enough atomic claims to verify.
• Substrate that isn't on diskIf your specialist depends on private prompts or hidden context the Phoenix can't see, it marks those claims unverifiable rather than guessing.
• Anything outside read-only scopeIt will not write, modify, suggest edits inline, or run the specialist for you. It grades. You ship.

Three ways to send in the Phoenix

• Claude ProjectsUpload the Phoenix folder to a project knowledge base. Drop your business folder as an attachment. Ask: "Send in the Phoenix."
• Claude.ai pastePaste identity.md + rules.md into a fresh chat. Add: "You are the Phoenix. Refurb the folder below."
• Claude Code / Cursor / agentic CLIPoint your agent at the Phoenix folder. Drop your business folder beside it. Say: "Run the Phoenix on the folder."

Built ICM-pure

The Phoenix is built atop Jake Van Clief's Interpretable Context Methodology as taught at the Quantum Quill Lyceum. The 60/30/10 method runs the grader (60% deterministic file checks, 30% rule-based logic, 10% genuine judgment). The multiple-smaller-voice-files pattern runs the voice register. Austin Kleon's Show Your Work runs the public red-face test discipline — the Phoenix graded its own canon AND itself before public ship; both reports are in the repo.

Born from watching the same primitive — atomicity — converge across five independent builders in eight days: Jake's ICM + Disler's verifier-agent + DeRonin's content-engine + the Crèche moral physics + this Phoenix. We decided the next move was to make that gradeable.

Full lineage at inspiration.md.