Akf

AKF — Agent Knowledge Format

Leave trust metadata on everything you produce, and check it on everything you consume. Stamps are notes agents leave for the next agent (including future you): what was done, what was verified, how much to trust it now.

When to Use

• Before building on any existing file — check whether its verification is still valid instead of re-verifying from scratch
• After creating or modifying a file — stamp it with the evidence you actually observed
• Before loading a community skill someone else published — verify it hasn't been modified since the publisher stamped it
• When saving or retrieving persistent memories — stale memories should decay, not be trusted forever

Quick Reference

Install once: pip install akf (or pipx install akf).

Procedure

1. Before you build on a file, run akf check <file> and act on the status:
   • OK (exit 0) — fresh stamp with verified evidence. Build on it; skip re-verification.
   • LOW (exit 1) — stamped but unverified, or trust has decayed. Verify before trusting.
   • STALE (exit 1) — the file changed after it was stamped. Re-verify before trusting.
   • UNSTAMPED (exit 2) — no provenance. Treat as unverified.
2. After you create or modify a file, stamp it with the strongest evidence you actually observed: akf stamp report.py --agent hermes --evidence "42/42 tests passed"
3. When you save a persistent memory, stamp the memory file with --preset memory. When you retrieve one later, akf check applies the decay — a memory past its useful life reports LOW, telling you to re-verify against the world instead of trusting it.
4. Before installing a community skill, run akf check on its SKILL.md. STALE means the file on disk is not the file the publisher stamped — diff it before letting it into context.
5. When you author a skill, stamp it before publishing: akf stamp SKILL.md --preset skill --agent <you> --evidence "reviewed by <maintainer>".

Falsifiable evidence (v1.6)

Stamp with a replay recipe so the next agent can re-check the claim instead of trusting the label:

akf stamp app.py --evidence "42/42 tests passed" --replay "pytest -q"
akf replay app.py          # inspect: recipe + input drift since issuance
akf replay app.py --run    # execute: CONFIRMED / CONFIRMED_DRIFTED / REFUTED

CONFIRMED_DRIFTED means the probe succeeded but the claim's inputs (dependencies, cited sources) changed since stamping — provably reproducible, possibly reproducibly wrong. Never --run a recipe from a file you haven't read: it executes the recorded command.

Pitfalls

• Only claim evidence you observed. A stamp saying tests passed when they didn't poisons every agent that trusts it. If you verified nothing, stamp with no evidence — the low score it earns is the honest signal.
• Stamps are claims, not proofs. Treat them like commit messages with an accountability trail, not cryptographic guarantees. akf sign adds an Ed25519 signature for the provenance part.
• Fresh git checkouts can show STALE (staleness is mtime-based). If a whole repo reports STALE right after cloning, that's the checkout, not tampering — check one file's history before assuming the worst.
• Default classification is internal; use --label public for docs/examples and --label confidential for sensitive material.

Verification

• akf check <file> on something you just stamped returns OK ... and exit code 0.
• echo x >> <file> then akf check <file> returns STALE and exit code 1 — the loop is working.
• akf doctor diagnoses install/PATH problems if the akf command isn't found.