CLAUDE.md · diff
git:20260317.72654b9 to git:20260702.4eba06d
39 added, 10 removed. Audit A to A.
+ ---
+ akf:
+ v: "1.0"
+ claims:
+ - c: Trust metadata for CLAUDE.md
+ t: 0.7
+ id: cf117d0c
+ src: unspecified
+ tier: 5
+ ver: false
+ ai: true
+ evidence:
+ - type: other
+ detail: "docs reviewed, protocol matches shipped akf check behavior"
+ at: "2026-07-02T16:52:28.182634+00:00"
+ id: "akf-c8cf5407d8a6"
+ agent: "claude-code"
+ at: "2026-07-02T16:52:28.183459+00:00"
+ label: public
+ inherit: true
+ ext: false
+ hash: "sha256:2cc31b8895ecb034"
+ sv: "1.1"
+ ---
# AKF — Agent Knowledge Format
- AKF is the trust metadata standard for AI-generated content. When you create or modify files, stamp them with trust metadata so every output carries provenance.
+ AKF is the trust metadata standard for AI-generated content. Stamps are notes agents leave for the next agent: what was done, what was verified, and how much to trust it. A stamp costs ~15 tokens; re-verifying costs thousands.
- ## Rules
+ ## The protocol
- ### After creating or modifying files
+ ### Before building on an existing file — check it
```bash
- akf stamp <file> --agent claude-code --evidence "<evidence>"
+ akf check <file>
```
- Evidence examples: "tests pass", "docs reviewed", "refactored from existing code", "generated from spec"
+ - `OK` — fresh stamp with verified evidence. Build on it; skip re-verification.
+ - `STALE` — modified after stamping. Re-verify before trusting.
+ - `LOW` — stamped but unverified (no test/review evidence). Verify before trusting.
+ - `UNSTAMPED` — no metadata. Treat as unverified.
- ### After completing a task with commits
+ ### After creating or modifying files — stamp what you verified
```bash
- akf stamp <file> --agent claude-code --evidence "<summary of evidence>"
+ akf stamp <file> --agent claude-code --evidence "<evidence>"
```
+ Evidence is auto-classified and drives trust: `"42/42 tests passed"` (test_pass), `"mypy: 0 errors"` (type_check), `"reviewed by @user"` (human_review), `"CI green"` (ci_pass). Only claim evidence you actually observed — a stamp that says tests passed when they didn't poisons every agent that trusts it.
- ### When working with existing AKF-stamped files
+ ### Deeper inspection when needed
```bash
- akf read <file> # Check existing trust metadata before modifying
- akf inspect <file> # See detailed trust scores
+ akf read <file> # Full trust metadata
+ akf inspect <file> # Pretty-printed trust scores
```
### Classification
- Use `--label confidential` for files in `*/finance/*`, `*/secret/*`, `*/internal/*`
- Use `--label public` for README, docs, examples
- Default is `internal`
## Available Commands
+ - `akf check <file>` — One-line trust check (OK/LOW/STALE/UNSTAMPED, exit codes 0/1/2)
- `akf stamp <file>` — Add trust metadata to any file
- `akf read <file>` — Read trust metadata from any file
- `akf inspect <file>` — Pretty-print trust scores
- `akf embed <file>` — Embed metadata into DOCX/PDF/images
- `akf extract <file>` — Extract metadata from any format
- `akf scan <dir>` — Security scan a directory
- `akf audit <file>` — Compliance audit (EU AI Act, SOX, NIST)
- `akf trust <file>` — Compute effective trust scores
## Project Structure
- `python/akf/` — Python SDK source
- `typescript/src/` — TypeScript SDK source
- `site/` — Website (Vite + React + Tailwind)
- `spec/` — Format specification and JSON schema
- `packages/` — Framework integrations (MCP, LangChain, LlamaIndex, CrewAI)
- `extensions/` — VS Code, GitHub Action, Office, Google Workspace
- `skills/` — Agent skill files