Skip to Content
Motivation

Why This Protocol Exists

A cryptographically authentic record is not the same as a correctly-read record. Every transparency substrate that proves existence, integrity, and history — ATL, Certificate Transparency, timestamping services, append-only ledgers — leaves an unaddressed gap: the gap between what the bytes are and what the bytes claim about the world. That gap is where decontextualization, frame laundering, and semantic overclaim operate. APL closes it.

ATL made history tamper-evident. APL makes interpretation tamper-evident.

Scale of the Problem

The failure modes APL addresses are not theoretical. They recur across every domain where a narrow observation is published, reused, and compared to another narrow observation under a different set of assumptions.

Attack PatternWhere It AppearsObservable Effect
DecontextualizationAI eval dashboards, benchmark citation, model cardsA result from one narrow observation context is presented as a general property of the subject
Frame LaunderingAggregated benchmarks, cross-harness comparison, provenance portalsA common label (e.g. “MMLU”) masks differences in harness, grader, scope, or protocol
Semantic OverclaimSafety claims, regulatory filings, deployment-readiness statementsA narrow benchmark score is extrapolated to overall quality, safety, or readiness
Silent TransformationImage pipelines (raw → denoised), document redaction, unit conversion, data cleaningA transformation is treated as semantics-preserving while silently altering what is claimed
Mutable Frame IdentityFrames identified by URL, alias, or registry path onlyA frame’s bytes can be swapped without the verifier noticing
Implicit Cross-Frame ComparisonLab-to-lab measurements, cross-instrument readings, cross-runner evaluationsNon-equivalent observations are compared as if equivalent

Each of these is a silent failure: the artifact appears authentic and the carrier proof is valid, yet the meaning of the artifact has drifted — and there is no protocol-level signal that anything is wrong.

The Common Thread

Every pattern above has the same root cause:

  • Claims are published without the observational frame that made them meaningful.
  • Frames are identified by name, URL, or registry path rather than by content hash — inviting silent substitution.
  • Two claims are compared without a declared bridge licensing the comparison.
  • Transformations are asserted to be semantics-preserving without enumeration of what is actually preserved and what is lost.
  • Verifiers treat “cryptographically authentic” as synonymous with “correctly interpretable”.

Carrier-layer cryptography (hashes, inclusion proofs, timestamps, anchors) cannot fix this. A receipt can be perfectly anchored and still license a misreading, because the cryptography does not reach the semantic layer.

How APL Addresses Each Failure Pattern

Failure PatternWhere It AppearsHow APL Prevents It
DecontextualizationAI eval, media citation, benchmark reuseA claim MUST carry frame_ref by content hash; a claim without a resolvable frame is apl-invalid
Frame launderingAggregated benchmarks, cross-harness comparisonFrame identity is SHA-256(JCS(frame)); any alteration of the frame bytes produces a different identity and a hash mismatch
Semantic overclaimModel cards, regulatory reportsFrames MUST declare exclusions explicitly; an overclaim sits outside the declared scope and is protocolically not licensed by the frame
Silent transformationImage and document pipelinesAPL Transformation MUST enumerate preserves and losses; an empty losses array means “nothing is lost” — never “unknown”
Mutable frame identityURL-only, alias-only, registry-path-only referencesResolver hints and aliases are NOT identity; a verifier MUST verify the resolved frame’s content hash against frame_ref.hash
Implicit cross-frame comparisonLab-to-lab, runner-to-runner, harness-to-harnessincomparable is the default across frames; cross-frame comparison requires an explicit, content-addressed Bridge declaring comparison_scope, assumptions, and losses

Protocol Capabilities

Each failure mode above maps directly to an APL mechanism:

  • Frame-bound receipts — every claim carries a pinned frame_ref. The claim cannot be validly detached from the frame it was made under.
  • Content-addressed frame identity — frames are identified by SHA-256 of JCS-canonicalized bytes. Frame substitution is detectable at verification time.
  • Aspect linkage — every claim.aspect_refs entry MUST appear in frame.aspect. A claim cannot silently make assertions the frame does not cover.
  • Explicit bridges — cross-frame comparison is a protocol-level declaration, not a reader’s inference. Bridges are directional, content-addressed, and scoped.
  • Declared invariants and losses — transformations are never implicitly semantics-preserving. preserves and losses are REQUIRED fields.
  • Deterministic verifier contract — a verifier returns exactly one of apl-valid or apl-invalid. There is no “probably valid”.
  • Substrate-agnostic layering — APL does not duplicate carrier cryptography; it composes with any substrate meeting the minimal contract, and inherits the substrate’s security properties without re-proving them.

Detailed Case Analysis

The canonical worked example of what APL refuses to license is The Two MMLU Scores: two cryptographically-authentic benchmark scores for the same model, each produced under a different harness, grader, and scope, publicly compared as if they were comparable. A carrier-only verifier can prove the existence, integrity, and history of each record individually — but carrier alone does not prohibit the silent comparison between them. APL does not declare either score false; it refuses to license the comparison without an explicit bridge.

The full normative treatment of the scenario is in The Protocol § 5.8. The adversarial pattern breakdown (decontextualization / frame laundering / semantic overclaim) is formalized in The Protocol § 6.5.

Last updated on