AHL Protocol
Record-Level Provenance.
A derived record — a resolved identity, a risk score, an enriched attribute — is produced from other records by a transform. ATL proves the history was not rewritten. APL proves you are reading it under the same frame. AHL proves what a record was asserted to have been computed from, and what must happen when a source is corrected or withdrawn.
The Problem
The lineage catalogs surveyed for this work record derivation as mutable metadata: dataset- and column-level graphs, asserted by the pipeline about itself, changeable without a trace. In domains where a later reviewer must reconstruct a specific output — sometimes years later — that is not evidence.
“This attribute came from system A.” — Which record, as of when, under which transform?
“The model agrees with the enrichment.” — Independent corroboration, or the same inference recycled through a training set?
“We removed the source record.” — Did every derived attribute and model that consumed it get invalidated?
“This is why the system scored them that way in 2019.” — Can you reconstruct the inputs, or only the current store?
Those catalogs were built for impact analysis and debugging, where a trusted-operator assumption is fine. Audit-grade provenance of individual predictions is an adversarial setting. The assumption does not transfer. Record-level granularity exists in research systems without integrity protection; integrity protection exists for certificates, software, and media — not, in the survey, for records in data and ML pipelines. Revocation propagation is not defined in the catalogs surveyed.
The Solution
AHL binds derivation to signed, digest-bound statements over individual records, anchored in an append-only, cryptographically verifiable log. The core insight: an assertion about a pipeline is not evidence about a record.
Record-Level Statements
Seven statement types — ingestion, derivation, retraction, correction, propagation, manifest, key — name the exact input records, code, model, and parameters that produced an output. The unit is the record, not the dataset.
Pluggable Log
Statements anchor into any log meeting the binding contract: append-only, inclusion and consistency proofs, authenticated enumeration at L3. AHL mandates no operator or product. ATL is one conforming log, alongside RFC 9162-class transparency logs.
Bitemporal Statements, Index-Ordered
valid_time, issued_at, and log-attested incorporation time stay distinct. Every “before”, “after”, “latest”, and checkpoint-membership rule uses the entry’s position in the log — never a timestamp.
Revocation With a Named Boundary
At L3, a retraction or correction obliges computation of the transitive closure over the statement graph, disposition of every affected record, and an anchored propagation statement — completeness defined at the trigger’s declared corpus checkpoint, never claimed forever after. L2 records the trigger and states plainly that revocation was recorded, not propagated.
Independent Verification
An Evidence Receipt is a portable artifact: offline-checkable, years later, without the producer or the log. Every receipt names the boundary of its claim and proves exactly that.
Who Needs This
Data & ML Pipelines
Enrichment writes derived attributes back into the store that trains the next model. Without record-level edges the feedback loop is invisible. Where those edges were captured, AHL can make a cycle detectable and the asserted inputs of a logged output reconstructible — at L3, relative to a declared corpus.
AI Accountability
A score, a link, a recommended next system — each is a derivation. When someone asks why a logged output existed at time T, a receipt can answer for the statements that were anchored: asserted inputs, code, and model version. It does not prove that every real-world input was captured.
Regulated Records
When a later reviewer must reconstruct what a system asserted about a record, that is a provenance problem. AHL specifies a graph of signed, digest-bound statements. Whether that graph satisfies a given statute or control is a deployment question, not a property of the protocol.
Credit & Consumer Reporting
Showing what a scored decision was asserted to rest on is a provenance problem. AHL addresses that piece. It does not write notices, choose recipients, or determine whether a reason is legally sufficient.
Cross-Organization Exchange
When a record or its provenance crosses an operator boundary, the receiving side can store a claim or verify one. Catalogs offer the first. AHL’s log-binding and receipts offer the second.
Technical Foundation
Open Specification (v0.3-draft)
AHL is protocol infrastructure, not a product. Apache-2.0 on publication of the reference implementation. No novel cryptography — signatures, digests, and RFC 9162-class proofs.
Reference Implementation
- ahl-core: Statement construction and validation, closure computation, Evidence Receipts, deterministic test-vector corpus. Pre-release; source publication is a gate for the public draft.
- Log adaptors are versioned, content-addressed profiles. ATL is the expected first adaptor.
Need record-level provenance in a pipeline you already run? Binding derivation to signed statements is contract work I take on.
Learn More
Full protocol specification, data structures, test vectors, and conformance rules: