Why This Protocol Exists

Reconstructing an automated decision after the fact — sometimes years after — is now a regulatory default, not an edge case. Every regime below reduces to the same question: show what this result was asserted to have been computed from, as of when, and prove the history has not been rewritten. The gap analysis that led to this specification: What “Traceable Lineage” Actually Requires  (August 2026).

Scale of the Problem

The Common Thread

• Derivation is asserted in mutable metadata, not bound to record content and not anchored in a tamper-evident history.
• A retraction of a source is recorded somewhere, but nothing proves it reached every record derived from it.
• “As of” reconstruction depends on trusting the operator’s internal database, not on independently checkable proof.
• Lineage tooling operates at dataset and column granularity; the obligations above are about specific records behind specific decisions.
• Completeness of a revocation sweep is asserted, not verifiable against an enumerable corpus.

Lineage catalogs (OpenLineage, DataHub, Apache Atlas) track derivation between datasets and columns, in mutable metadata stores populated by what pipelines assert about themselves. Record-level provenance exists in database research; cryptographic tamper evidence exists in transparency logs (Certificate Transparency, Rekor) and content credentials (C2PA) — but for certificates, software artifacts, and media, not for records moving through data and ML pipelines. No standardized primitive combines record-level granularity, tamper evidence, and revocation semantics. AHL specifies that combination.

How AHL Addresses Each Failure Pattern

Protocol Capabilities

• Digest-bound statements — seven types (ingestion, derivation, retraction, correction, propagation, manifest, key) covering the full provenance lifecycle, including the governance of the corpus itself.
• Pluggable log-binding contract — any log offering append-only storage, inclusion and consistency proofs, and authenticated enumeration at L3 conforms; no mandated operator or product.
• Witness protocol (L3) — an independent cosigner makes split-view and equivocation by the log operator detectable, not merely assumed absent.
• Bitemporal statements — valid time, issued time, and log-attested incorporation time stay distinct; every ordering rule uses entry index, never a timestamp.
• Revocation closure with a named boundary (L3) — completeness is defined and mechanically checkable at a declared corpus checkpoint, and never claimed unconditionally; L2 carries no closure duty.

Detailed Case Analysis

Each regime from the table above reduces to a concrete evidentiary demand. What follows names the AHL mechanism that bears on each one — mechanisms, not legal conclusions. AHL does not determine legal sufficiency, generate notices, or decide whether an obligation has been met; a mechanism below contributes evidence toward an obligation, it does not by itself discharge it.

EU AI Act (Regulation 2024/1689) Art. 12, 19, 26(6)

Art. 12 requires high-risk AI systems to automatically log events over the system’s lifetime, sufficient to identify situations that may present a risk to health, safety, or fundamental rights, or that may lead to a substantial modification. Art. 19 and 26(6) require providers and deployers to retain those logs. AHL’s derivation statement (Protocol §3.3.2) is a signed, digest-bound assertion of what specific inputs, code, and model version were used to produce a specific output — evidence toward the traceability the Article calls for, anchored so the log-attested history cannot be silently rewritten after the fact; whether a given deployment’s logging practice satisfies Art. 12/19/26(6) is a determination for that deployment, not a property AHL confers.

NIST SP 800-53 AU-9(3) / AU-10

AU-9(3) calls for cryptographic protection of audit information; AU-10 calls for non-repudiation of recorded events. Signatures on every statement (Protocol §3.1) and anchoring in an append-only, cryptographically verifiable log (Protocol §4) produce cryptographically verifiable audit information and a producer signature over digest-bound content with log-attested time — evidentiary contributions toward both controls, not a substitute for them. AU-10 in particular is an organization-defined control implemented across an organization’s broader control environment; AHL supplies one input to it, not its satisfaction.

ECOA / Regulation B, 12 CFR § 1002.9

Adverse-action notices must cite specific principal reasons for a credit decision. AHL’s point-in-time reconstruction (Protocol §5) resolves the asserted input commitments, representations, and transform digests behind a specific derivation statement — material that may support substantiation of the stated reasons. AHL does not generate the notice and does not determine whether the reasons it helps substantiate are legally sufficient.

FCRA § 611, 15 U.S.C. § 1681i(d)

At the consumer’s request, FCRA § 611 requires notice of deletion or dispute notation to specifically designated recent report recipients, within statutory time windows. This is federation-extension territory, not a core AHL capability: recipient-bearing disclosure/receipt statements plus trigger propagation could compute the consumer-requested, designated, time-scoped recipient set, but core AHL does not itself determine whether notice is required, select recipients, or send notice.

GxP data integrity (ALCOA+)

ALCOA+ calls for records that are attributable, contemporaneous, original (or a verified true copy), and accurate. AHL statements are signed (attributable), bitemporal with a log-attested incorporation time distinct from the asserted valid_time and issued_at (contemporaneous evidence, Protocol §3.2), and digest-bound to the exact bytes a transform consumed (original/accurate, Protocol §3.3.2) — mechanisms that map onto each ALCOA+ component, not a determination that a given deployment’s data integrity program is compliant.