Back to AI Research

AI Research

KLM makes AI evidence records inspectable while documenting what its tests do not prove

Key Takeaways

  • An open inference-record specification reports internal diagnostics and negative results, with end-to-end validation still outstanding.
  • ### Make the recorded basis of an answer visible
  • The [Knowledge Layer Model paper](https://arxiv.org/abs/2610.42066) proposes an implementation-neutral format for recording AI inferences, evidence relationships, versioned assessments and explicit unknowns.
  • SEDIM, the evidence architecture of Nage AI OS, is presented as a reference implementation.
  • The objective is inspectability: a reviewer should be able to examine what basis was recorded for an output.

Make the recorded basis of an answer visible

The Knowledge Layer Model paper proposes an implementation-neutral format for recording AI inferences, evidence relationships, versioned assessments and explicit unknowns. SEDIM, the evidence architecture of Nage AI OS, is presented as a reference implementation.
The objective is inspectability: a reviewer should be able to examine what basis was recorded for an output. That is a different claim from establishing that the output is correct, that the model causally relied on the cited evidence or that the surrounding system is secure.

Read the negative results alongside the specification

The empirical material consists of separate internal experiments rather than an end-to-end validation. The paper reports a corrected two-source test in which a native-model weight-path configuration produced zero successful answers out of fifteen.
Another result concerns context packing. A legacy process lost text containing the expected answer in 118 of 221 selected cases; repaired packing retained that text in all 221. The authors explicitly distinguish retaining text from producing a correct answer.
A historical routing result is also qualified: its numerator was recovered from a rounded summary, and historical example membership is unverified. A capacity-growth comparison could not establish a reliable effect because the older scorer was defective and original outputs were missing. These limitations should remain part of any account of the work.

Evaluate the record separately from the answer

For a team considering an evidence-record format, a useful pilot would inspect whether a reviewer can follow sources, revisions and unknowns without losing the original context. Answer quality needs its own evaluation against suitable tasks and outcomes.
The paper leaves independent replication, a complete public audit package and several proposed evaluations to future work. Its practical contribution is a specification and a candid account of internal diagnostic results, rather than demonstrated universal reliability for evidence-visible AI.

Comments