Skip to content

Framework v1.5 — Machine-Readable Evidence Traceability Layer

Latest

Choose a tag to compare

@mj3b mj3b released this 31 May 12:51
· 15 commits to main since this release
5e81799

What's in this release

This release adds a machine-readable evidence traceability layer to the CDFI Framework — the missing link between the seven-step translation documents and the computable artifacts they produced.

New: claims/ folder — 8 JSON evidence packs

Each file proves how a specific Anthropic AI safety research publication became a specific CDFI mechanism. Every claim carries verbatim paper extracts with section/page citations, Direct/Derived/Original Construct typing, and a cdfi_element_produced field linking the claim to its formula artifact (gate constant, column weight, certification threshold, scoring rule).

File Publication → Mechanism
claims/pub1-evaluation-criteria.json Challenges in Evaluating AI Systems → four-column weighting matrix
claims/pub1-rubric-reliability.json Challenges in Evaluating AI Systems → four-part certification protocol
claims/pub2-hallucination-gate.json Auditing for Hidden Objectives → hallucination gate (CAP_VALUE = 40)
claims/pub3-statistical-rigor.json Statistical Approach to Model Evaluations → 95% CI + deployment tier thresholds
claims/pub4-framing-sensitivity.json Discrimination in Language Model Decisions → relativism resistance gate
claims/pub5-confidence-calibration.json CoT Faithfulness + Discrimination (combined) → confidence calibration (original construct)
claims/pub6-categorical-failures.json Sabotage Evaluations → five failure mode taxonomy + cap gate architecture
claims/pub7-adversarial-probing.json Evaluating Feature Steering → adversarial prompt variant + prompt sensitivity drift

Schema reference, file index, and adaptation notes from applied-ai-research-translator: claims/README.md

Extended: Source Evidence Records in all 8 translation documents

Every translation document in docs/translations/ now carries a Source Evidence Record — verbatim paper text anchoring each pipeline step, with claim types marked and inference chains shown for all Derived steps. Translation 8 (confidence calibration) explicitly marks its convergent claim as an Original Construct with no source quote, because none exists.

What is still open

Item Status
Authority level classification (400 prompts) Pending — theological advisors
Human theological review (21,599 scores) Pending — Filip Ponulak + advisors
ISO/IEC 42001:2023 mapping Deferred to v1.6
Security remediations (prompt injection, pipeline integrity, signing) Deferred to v1.6

Full changelog: CHANGELOG.md