dsh-sentience-audit

Assessing whether an AI session exhibits functional organisation associated with consciousness has long lacked a reproducible method. Early heuristics counted vocabulary hits in a transcript, which measured phrasing rather than architecture and could not be diffed. dsh-sentience-audit takes a different route: it reads only replayable event-trace structure, scores it against the 14 indicator properties proposed by Butlin, Long, Elmoznino, and Bengio et al. (2023), and reports an L1-to-L5 level with per-indicator evidence.

The method is deterministic and platform-independent: the same event log always produces the same result on any operating system, with no model call, no network access, and no LLM judge. Three properties that require inspection of internal representations (HOT-1, HOT-4, AST-1) are honestly reported as not-assessable rather than guessed, which means L5 is effectively unreachable by design rather than by oversight.

Who is it for? Research teams auditing AI consciousness in a structured, citable way; platform engineers who need a diffable functional-organisation report for Harness sessions in production; and evaluators who want to avoid substituting adjectives and word-frequency counts for architectural analysis will all find a reproducible baseline to build on.