Evidence kinds/Input-hash-only (PII-safe)

Input-hash-only (PII-safe)

structural evidence

Only SHA-256 hashes of inputs enter the L12 chain — raw PII never crosses the audit-trail boundary.

What it is

When the substrate records an L12 entry, the input field is the SHA-256 hash of the canonical-form-serialised input, not the input itself. Same for the output (outputCanonicalHash). The raw content stays in the operator's source-data store, which the operator governs under its own retention + GDPR/HIPAA policy.

This is the substrate's structural answer to 'audit trails contain personal data'. With input-hash-only, the chain is GDPR-Article-5(1)(c) data-minimal by construction. A regulator verifying the chain doesn't see PII; the operator can replay decisions by joining the inputHash to the source data they retain separately.

For HIPAA: the chain contains no PHI by design. The covered entity / business associate retains the source ePHI under HIPAA Privacy Rule + Security Rule; Promethean's audit trail records only attestable hashes.

What it guarantees

  • Raw input + output never enter the L12 chain.
  • GDPR Art. 5(1)(c) data minimisation satisfied at the audit-trail layer by construction.
  • Substrate-infrastructure-side access to PII is structurally impossible.

What it does NOT guarantee

Honest limits. Every primitive has them.

  • That the operator's source-data store is GDPR/HIPAA-compliant — that's operator-side data engineering.
  • That the upstream LLM provider (Anthropic, OpenAI, Azure OpenAI) doesn't see the raw input — that's governed by the operator's BAA + the provider's retention terms.
  • Forensic replay without operator cooperation — if the operator's source data is deleted, the chain entry still verifies but the original content is gone.

Regulations that engage this kind

GDPR Art. 5(1)(c)

Data minimisation

HIPAA §164.502(b)

Minimum necessary use + disclosure

COPPA + FERPA

Child-data minimisation

Sectors that rely on this kind

Related evidence kinds

L12 receipt chain

Hash-chained, Ed25519-signed log of every LLM-in-the-loop decision a Promethean-built product makes.

Spec hash commitment

Every L12 entry includes a SHA-256 of the active RuntimeAISpec, making silent spec drift detectable.

FAQ

Can you replay a decision from the chain alone?

No — by design. The chain records the SHA-256 hash of the input, not the input itself. Replay requires the operator to supply the original input from their source data store. When source data is supplied, the chain validates the input matches what was originally recorded (re-hash and compare). This is the right tradeoff: PII safety + verifiable consistency.

What about the upstream LLM seeing the input?

Promethean does NOT proxy LLM calls. The operator's code calls the LLM provider directly; Promethean records metadata of that call into the L12 chain after the fact. Whether the LLM provider (Anthropic, OpenAI, Azure, AWS Bedrock) sees raw PHI/PII is governed entirely by the operator's contract with that provider — not affected by Promethean.