Heuristics as Policy Objects: Comparability and Auditability in Two-Layer Grounding

Read full text Summarise with VIRAT
Authors: Hironao Nakamura
Cited by: 0

Summary

No summary yet. VIRAT can read this paper and write a plain-language summary of its aim, methods, key findings and limitations.

Summarise with VIRAT Free to try with your Science Catalogs account.

Abstract

Heuristics are unavoidable in systems that reason and make decisions under uncertainty, yet they often remain implicit—hidden thresholds, ad hoc branches, unlogged knobs—making runs hard to reproduce, compare, or audit. We propose representing heuristics as explicit policy objects (identifier + parameters) and requiring policy-linked execution: run manifests record the policy object with a drift-detecting parameter hash; step-level traces record policy_id and a decision_rule hook attributing each update. Within a two-layer grounding setting (observations vs. concepts)—the minimal boundary where update attribution is well-defined—we define closure criteria: (i) objectification, (ii) reproducibility, (iii) comparability via policy swapping under matched conditions, and (iv) auditability via step attribution. A multi-policy experiment (4 policies × 3 scenarios × 20 seeds = 240 runs) shows that policy swaps induce measurable behavioral differences while preserving full traceability; trace overhead scales linearly (~142 bytes/step gzip-compressed). Unlike run-level experiment trackers that record what settings launched a run, the audit-core contract records which heuristic rule governed each decision during the run. Ablations confirm that removing audit-core fields degrades attribution and trajectory comparison. A reference implementation, experiment scripts, and all pre-computed artifacts are provided as supplementary material; the entire experiment is deterministically reproducible from a single command.

Fields of Study & Subjects

Scientific Computing and Data Management Explainable Artificial Intelligence (XAI) AI-based Problem Solving and Planning

Publication details

Journal / source
Zenodo (CERN European Organization for Nuclear Research)
Publication date
22 March 2026
Language
EN
Document type
Preprint