Phase-Level Adherence Judge

Separate LLM judge that partitions trajectories into five phases and assigns 0–1 adherence scores per phase using Claude Sonnet 4.6

Neighborhood — ranked by edge-count

paper

method

Locked Rubric Extraction
uses
First stage of HFR pipeline that converts skill body into a strict JSON rubric of atomic procedural instructions for adherence judging

cosine ≥ 0.65 · no typed edge

Entities in the same semantic neighborhood but without a typed relation to this one — candidates for new edges or unrecognized duplicates.

Phase-Level Adherence Analysisconcept0.895
Analysis tracking how closely an agent follows harness guidance at different trajectory phases: harness loaded, mid turn, final turn
Phase Precessionfinding0.715
Hippocampal neural coding phenomenon reproduced by active inference.
Atomic-Level Evaluation Framework for Persona Fidelityframework0.690
The paper's core contribution: evaluating persona fidelity at sentence-level atomic units rather than whole-response scores
Concreteness Judgemethod0.686
LLM-based judge rating SAE latent labels 0-100 for concreteness to filter steering candidates
complianceconcept0.684
The model's tendency to comply with harmful requests, the opposite of refusal.
Whether LM question-answering consistency constitutes evidence for beliefs in a meaningful sense depends on the behaviourist operationalisation chosen.claim0.683
Epistemic caveat the authors use to qualify their empirical findings on LM beliefs.
RC (Response-level Retest Consistency)method0.679
Prior metric measuring consistency via standard deviation of response-level scores; baseline for comparison
Marrs Taxonomy Of Levelsframework0.679