Skip to content

ADR 025 Reproducibility is Structural and Local

Jump to bottom
Michael Zargham edited this page May 18, 2026 · 2 revisions

ADR-025: Reproducibility is Structural and Local

Status: Accepted Date: 2026-05-16 Deciders: Michael Zargham Related: ADR-022 External URI References as Open-Source Foundation; ADR-023 Cryptography by Composition of Battle-Tested Standards; ADR-024 Identity by Thin Projection of External Sources; Verifiable Self-Certification; Transcript Replay Semantics; Design Spec

Context

A cert artifact's value depends on its reproducibility — auditors need to verify that the recorded facts are what they say they are. The naïve framing of reproducibility is global: a verifier dereferences the entire RTM graph, re-fetches every external dependency, replays every transcript step, and confirms equivalence. That framing assumes a single verifier with universal permissions, which breaks in real institutional contexts: different parties have different permission slices (the safety reviewer can see safety-scope facts but not security-scope facts; the regulatory submitter can see compliance facts but not internal IP), and no single party has universal access. A reproducibility model that requires universal permissions cannot be exercised in multi-party institutional audits. The question is whether v0.1's reproducibility model is global-and-monolithic or structural-and-local. See Design Spec §9.A.5 (cross-cutting acceptance criteria X6, X7, X8) and Verifiable Self-Certification.

Decision

flexo-rtm v0.1's reproducibility model is structural and local: each fact in the cert artifact is structurally complete for its own local context — the RDF neighborhood, external URI references (see ADR-022 External URI References as Open-Source Foundation), projection-at-cert-time (see ADR-024 Identity by Thin Projection of External Sources), and signatures (see ADR-023 Cryptography by Composition of Battle-Tested Standards) sufficient to reproduce that fact in isolation. Verifying parties need only the permissions for the facts they want to verify, not universal access. Reproduction federates computationally (compute distributes across parties) and organizationally (different parties verify different permission slices, composing to a complete audit). Cross-cutting acceptance criteria X6, X7, X8 (Design Spec §9.A.5) enforce this property.

Consequences

Positive

  • Multi-party institutional verification works without central coordination: each party verifies its own permission slice; the composition of all slices is the complete audit
  • Locality is a property of the cert artifact, not a property of the verifier — the artifact is structurally complete fact-by-fact, so the verifier doesn't need universal access to verify any single fact
  • Refresh policy (see ADR-024 Identity by Thin Projection of External Sources) and reproducibility are not in tension: refresh is an authoring-time freshness setting; reproduction always operates against the recorded projection-at-cert-time, never against fresh data
  • Federated computational verification: parties can verify in parallel; no single computational bottleneck
  • Forward-compatible to any downstream-analysis path (per ADR-032 Methodology Agnosticism as Foundational Axiom): an adopter who chooses to run topological analysis as a downstream-analysis mode (see ADR-003 Topological Framework Documented as Future Work) would compute the recursive completeness check over the same structurally-local facts; SLSA, GSN, ARP4754A, or in-house downstream analyses read the same locality property

Negative / Tradeoffs

  • "Structurally complete for its own local context" is more nuanced to document and explain than "globally reproducible"; mitigated by clear documentation in Verifiable Self-Certification and by the acceptance criteria X6/X7/X8 enforcing the property mechanically
  • Cert-artifact size grows with locality completeness — each fact carries its neighborhood, external references, projection-at-time, and signatures; mitigated by the lean-default ontology (see ADR-014 Parsimony Layer Build-Time Extraction) and by external references being URI-sized (cheap)
  • Federation requires permission infrastructure on the verifier side; adopters integrating with their own permission systems must respect the locality property

Neutral

Alternatives Considered

  • Treat reproducibility as a single global property requiring a verifier to re-dereference the whole graph and hold universal permissions: A single verifier with universal access dereferences everything, re-fetches every dependency, replays every transcript, and emits a global pass/fail signal. Rejected: that framing breaks in multi-party institutional contexts where different parties have different permission slices. No single party can hold universal permissions across safety, security, compliance, IP, and personnel boundaries. The chosen framing — each fact is structurally complete for its own local context, verification is local in the traceability graph, and reproduction federates computationally and organizationally — is what makes the cert artifact usable in real institutional audits where no single party has universal access. Refresh policy and reproducibility are not in tension: refresh affects authoring-time freshness; reproduction always operates against the recorded projection-at-cert-time.

Implementation Notes

References

Start here

Foundation

Internal Research

External Research

v0.1 Certification Model

Roadmap and future work

Three-Layer Architecture

Ontology Design

Interface Contracts (normative)

Adapter Contracts (background)

Reproducibility

Open issues

Decision Log

Meta

Clone this wiki locally