In agent development, we often need to evaluate whether a session exhibits some form of structured functional organization. The slatinwine/dsh-sentience-audit plugin for DeepSeek Harness (DSH) provides an audit tool based on trace structure. It rates sessions from L1 to L5 in accordance with Butlin et al. (2023)’s consciousness metrics theory, rather than measuring subjective experience.

Core Features

The core role of this plugin is to audit DeepSeek Harness sessions and provide a structured rating. It does not rely on any large language model or network calls, and performs deterministic analysis based on the session’s trace structure.

  1. Audit 14 consciousness metrics: Audit the session against the 14 consciousness metrics proposed by Butlin, Long, Bengio, et al. (2023).
  2. Report L1-L5 levels: Output a specific level rather than an average value.
  3. Deterministic evaluation: The results are fully reproducible and do not depend on model calls, LLM judgments, or network access; they are also platform-independent.

Installation and Enablement

This plugin has not yet been published to the npm registry and must be installed directly from the release package.

  1. Download slatinwine-dsh-sentience-audit-0.1.1.tgz from GitHub Releases.
  2. Install it using the DSH command:
dsh plugin --profile my-profile add ./slatinwine-dsh-sentience-audit-0.1.1.tgz
  1. After installation, you can confirm that the configuration includes the sentience-audit entry by running dsh --profile my-profile --dump-config.

Typical Usage

In a DSH session, the model can invoke the sentience_audit tool.

sentience_audit()                       // audit the current session
sentience_audit({ sessionId: "..." })   // audit another live session

The tool returns a report containing the L1-L5 level and supporting evidence. For example, it may display “Achieved 8 · Not assessable 3 · Missing 3” and list the determinations and evidence types for specific metrics (such as RPT-1 and HOT-1), including categories such as architectural preset and trace structure.

Applicable Scenarios and Notes

Applicable Scenarios

  • Evaluating whether an agent has structural features such as recursive loops (L2) or a global workspace (L3).
  • Serving as a reference for agent architecture design, rather than as proof of subjective experience.

Notes

  • Not a consciousness scale: The tool measures functional organization, not consciousness itself. The authors explicitly state that current AI systems have not passed the test.
  • Reliance on trace structure: It is based only on replayable trace structures (such as tool call order and file reads/writes), and does not inspect internal representations.
  • Honest ceiling: Because AST-1 (Attention Schema Theory) cannot be assessed from traces alone, the L5 level is effectively unreachable.
  • Live sessions only: It can audit only live sessions in the sessions service, not archived sessions.
  • Provisional scoring criteria: The scoring criteria are based on a 2023 paper, and the author expects them to be updated in the future.

Conclusion

This plugin provides DSH with an objective and reproducible way to evaluate an agent’s structural capabilities. If you need to assess the degree of functional organization in a session, this plugin can be a useful reference.