Every AI inference produces a 74-byte post-quantum attestation. Hash-chained into a tamper-evident sequence. That footprint is an expression of the H33-74 primitive, which produces and anchors the portable post-quantum evidence — this page applies it to AI decisions rather than defining a new primitive.
Not "we logged it." Cryptographically signed. Mathematically bound.
Logs tell you what someone claims happened. Attestations prove what actually happened. The gap between these two is where liability, fraud, and regulatory failure live.
Application logs can be edited, deleted, or corrupted by anyone with admin access. A breach destroys the evidence. A malicious insider rewrites the record. In litigation, unsigned logs are hearsay — not evidence.
"Our systems show..." is a trust claim, not a mathematical proof. Regulators are moving from trust-based to evidence-based compliance. If your AI vendor says the decision was compliant, you are trusting their word — not verifying it.
RSA and ECC signatures will be forgeable by quantum computers. An AI audit trail signed today with RSA-2048 may be worthless in court by 2035. Post-quantum signatures are the only way to future-proof AI decision records.
One is a claim. The other is proof.
Every AI decision triggers a five-step attestation pipeline. The output is a 74-byte proof.
Five elements are cryptographically bound into every 74-byte attestation. Missing any one of them means the proof is incomplete. H33 captures all five by default.
H33 attestations are designed to be independently verifiable (not yet available in production).
An auditor receives the 74-byte attestation, checks the three PQ signatures against public keys, verifies the SHA3-256 hash chain, and confirms the on-chain anchor. No access to your systems required.
In litigation, the attestation is mathematical evidence that a specific AI decision occurred with specific inputs at a specific time. It cannot be forged, backdated, or altered.
Send the 74-byte proof. They can inspect it (independent verification not yet available in production).
Every attestation is hash-chained to the previous one. Verifying any single attestation also validates its position in the sequence. A gap or alteration anywhere in the chain is immediately detectable.
The 32-byte on-chain hash provides an immutable public timestamp. Even if every other system is compromised, the on-chain record proves the attestation existed at that point in time.
The full attestation details — all five attested elements plus the three signatures — are stored on Cachee.ai for instant retrieval. Storage with sub-microsecond lookup.
Two API calls. One to attest. One to verify. No SDK required.
# Attest an AI decision attestation = h33.attest( inputs=inference_request, outputs=model_response, model_version="gpt-4-0125", authority=delegation_chain ) # attestation.proof — 74-byte H33-74 attestation # attestation.hash — SHA3-256 commitment (32 bytes on-chain) # attestation.chain — hash-chain position in sequence # Verify (any party, no API key needed) result = h33.verify(attestation.proof) # result.valid — True # result.signatures — {ml_dsa: valid, falcon: valid, slh_dsa: valid} # result.chain_intact — True
AI attestation is a cryptographic proof that a specific AI decision occurred with specific inputs, produced specific outputs, used a specific model version, and was authorized by a specific authority chain. Unlike logging, attestation is designed to be immutable and independently verifiable (not yet available in production).
Logs are mutable, vendor-controlled, and unsigned. Any administrator can edit or delete them. A breach can destroy the entire log history. In court or regulatory proceedings, unsigned logs are assertions — not evidence. Cryptographic attestation produces a mathematical proof that cannot be altered after the fact.
Every AI decision attestation is designed to be distilled to exactly 74 bytes: 32 bytes stored on-chain (the SHA3-256 hash commitment) and 42 bytes stored in Cachee (the compressed three-family PQ signature bundle). H33-74 is designed to distill three full post-quantum signatures into 74 bytes. Patent pending.
Not yet for production receipts. By design, the 74-byte attestation contains the hash commitment, which can be checked against the on-chain record, and the three PQ signatures are designed to be checkable against public keys; independent verification by auditors, regulators, counsel or partners is not yet available.
Each attestation includes the SHA3-256 hash of the previous attestation in the sequence. This creates a tamper-evident chain: altering or removing any attestation breaks the hash chain from that point forward. Inserting a fake attestation is impossible without knowing the hash of the attestation that precedes it. This provides sequence integrity in addition to individual attestation integrity.
Approximately 42 microseconds per AI decision. For a typical LLM inference of 50–500 milliseconds, this is less than 0.1% overhead. Batch attestation (via H33-74 Claims 124–125) amortizes the Dilithium signature across 32 decisions via batched Merkle response attestation, further reducing per-decision cost. The attestation pipeline runs asynchronously and never blocks the inference response.
Make an AI decision. Get a 74-byte proof. Inspect it. No commitment required.
See also: Attestation API