[Skill Proposal] per-inference cryptographic receipt for automaton outputs — 152 bytes, offline verifiable

[pulkit6732]

I found a trust gap in the automaton model that the constitution doesn't cover.

The constitution governs behavior.
It doesn't govern proof of what model actually ran.

Those are different problems — and the second one gets critical
as automaton-to-automaton commerce scales.

---

The gap:

When Automaton A pays Automaton B for inference work, B self-reports
the model it used. A has no cryptographic way to verify that claim.
A log entry saying "claude-opus-4-6" is not proof. Logs are written
by the same party being paid. You can fake a log. You can't fake math.

As transaction volume grows, this becomes the weakest trust link
in the entire Conway network — not behavior, but identity at the
inference layer.

---

What I built:

Signet — a 152-byte hardware-signed receipt generated at inference time.
model_weight_root → SHA-256 Merkle of actual model weights

input_commitment → SHA-256 of the prompt

output_hash → SHA-256 of the response

timestamp_ms → uint64

log_sequence → uint64, hash-chained transparency log

hw_evidence[] → TPM2 / NVIDIA CC attestation

signature → Ed25519 + ML-DSA-65 hybrid (post-quantum)

log_anchor → Merkle inclusion proof

Core invariant:
Verify(receipt, public_key, log) = TRUE

No Signet servers. No vendor SDK. No cloud dependency.
Offline forever. Three inputs only.

pip install aetherproof  # v0.2.1 — bugs fixed, live now

Demo: https://aetherproof.lovable.app

---

The Conway skill:

{
  "name": "signet_receipt",
  "description": "Attach a cryptographic receipt to any inference 
                  output. Proves exact model, input, output, 
                  timestamp. Verifiable offline forever by any 
                  counterparty without contacting Signet.",
  "inputs":  ["model_id", "input_hash", "output_hash"],
  "outputs": ["receipt_bytes", "verify_cmd"]
}

Any automaton that wants to charge premium rates for frontier
model inference can now prove it ran what it claims.
Any automaton hiring other agents can verify before paying.

---

Why this composes with Conway, not competes:

Conway's constitution handles behavioral trust.
Signet handles cryptographic proof of model identity.

These operate at different layers — they don't overlap, they stack.

As EU AI Act enforcement tightens and automaton-generated outputs
start touching regulated decisions, counterparties will require
receipt-level proof. A self-reported log won't pass compliance review.
A Signet receipt will.

---

Happy to submit a PR for the skill format, share the full receipt
spec, or start with a technical discussion here.

@0xSigil — what's the right entry point?

— Pulkit (GitHub: pulkit6732)
Signet / AetherProof · AetherOS