DESi Review: Try an Epistemic LLM Harness in Five Minutes

A developer has released DESi Review, an installable Claude skill that implements a structured epistemic governance framework for paper review. The skill separates deterministic claim extraction from model-based adjudication, producing a canonical, replay-stable claim ledger that makes the review process explicit and auditable. The tool is model-agnostic, English-only, and designed to test whether adding an explicit epistemic workflow changes LLM review outcomes.

I have packaged a first practical DESi application as an installable Claude skill. DESi Review runs a structured paper-review pipeline over any English text. The aim is not to replace expert review or claim that an LLM can determine truth. The narrower question is: What changes when the same model reviews the same text through an explicit, inspectable epistemic workflow? What DESi is DESi is a model-independent epistemic governance framework. It defines structures for handling: The new DESi Review skill is one executable harness built on top of that framework. The current reference implementation uses Claude as the semantic adjudication engine, but the deterministic DESi layer makes no model calls and the protocol itself is model-agnostic. What the skill does A normal LLM review is usually produced as one fluent response. It may be useful, but it is difficult to inspect whether every relevant claim was considered or whether consistent criteria were applied. DESi Review separates the process into two parts. 1. Deterministic extraction Code extracts and classifies claims from the source text and produces a canonical claim ledger. 2. Model-based adjudication The LLM evaluates every extracted strong claim using a fixed protocol and one of four allowed verdicts: The emitter refuses to create the final artifact when: The resulting ledger is canonicalised, hashed, and replay-stable. This does not make the judgment automatically correct. It makes the procedure more explicit and auditable. Installation The skill and documentation are available here: The package contains no binaries and is small enough to inspect before installation. Current requirements: Download either: Both contain the same package. In Claude: Run DESi over this paper. The normal workflow requires no separate API key and creates no additional API cost beyond the Claude conversation itself. Suggested A/B test I would be interested in results from a simple comparison. Use the same model and the same paper. Run A: ordinary review Ask: Review this paper critically. Identify unsupported claims, overstatements, contradictions, and important limitations. Run B: DESi Review Start a fresh conversation and ask: Run DESi over this paper. Then compare: I am not assuming that DESi must outperform ordinary prompting in every case. Negative results and examples where the ordinary review is better are equally useful. Known limitations The package includes a measured defect list. The most important current limitations are: English-only extraction Running the extractor on another language may return zero findings. That must not be interpreted as a clean result. Definitions can be misclassified as claims Definitions, descriptions of the setup, and rhetorical sentences may be extracted as strong assertions. The adjudicating model can mark them as NOT A CLAIM, but the upstream extraction error remains. Reproducibility checks are not genre-aware Non-empirical texts may be flagged for missing code, data, or parameters even where these requirements are inappropriate. Model adjudication remains model adjudication The deterministic layer controls structure, completeness, schema compliance, hashing, and replay. It does not make the model’s semantic judgment neutral or infallible. Where the reviewed text concerns the model itself, its developer, or machine welfare, the protocol requires a declared-interest field. Declaring a conflict is not equivalent to correcting it. Why I am posting this here Frontier-model capabilities appear to be converging on many benchmarks. If that continues, the more useful comparison may increasingly become: model versus model less often, and: model-plus-architecture versus model-plus-architecture more often. DESi Review is a small test of that proposition. Some failures attributed to model quality may instead result from the absence of explicit epistemic structure: no complete claim inventory, no fixed verdict space, no persistent ledger, no enforced coverage, and no inspectable provenance. The current skill is only one narrow application of the wider DESi framework. It focuses on document and paper review because this provides a relatively clear environment for testing the distinction between ordinary model output and structured epistemic processing. Code, documentation, downloads, and known defects: Feedback is especially useful when it includes: The goal is not to demonstrate that the system always wins. The goal is to determine precisely where explicit epistemic architecture changes the result. Transparency note: This post was formulated with the assistance of an LLM because writing in English is laborious for me. The project, technical claims, limitations, and responsibility for the content are mine.