Method
How a work is reviewed
Elenchus asks one question of a text: does the conclusion follow from what it says? It does not decide whether the premises are true.
The panel
Each work is examined by five agents in sequence. Every agent returns structured output, and every review publishes all of it.
- CartographerMaps the argument: the thesis, each claim with its role and a quotation anchoring it in the text, the premises the text needs but never states, the inferential steps, and the key terms.
- LogicianAudits every step as valid, strong, weak or a non sequitur, and names formal and informal fallacies.
- SocratesReads across the whole text for contradictions, tensions, equivocation, circular reasoning and conclusions that drift beyond their evidence, then writes the questions he would put to the author.
- AdvocateDefends the work under the principle of charity. Each objection is upheld, mitigated or rejected, and the Advocate writes the strongest version of the argument.
- ArbiterWeighs only the objections that survived the defence and returns the verdict, a coherence score from 0 to 100, and ranked issues with suggested repairs.
The grade
The letter grade is derived from the coherence score.
| Grade | Score | Meaning |
|---|---|---|
| A | 90–100 | Airtight |
| B | 80–89 | Sound, minor gaps |
| C | 70–79 | Sound, real gaps |
| D | 40–69 | Seriously flawed |
| F | 0–39 | Does not hold together |
Issues are rated critical when the thesis does not follow or the text contradicts itself on something central, major when an important supporting line fails, and minor for local gaps, overstatements or loose terms that do not threaten the thesis.
What a grade does not mean
- It is not a fact-check. A perfectly coherent argument can rest on false premises, and a weak argument can reach a true conclusion. Each review lists the empirical claims to verify elsewhere.
- It is not a judgement of the research. Data quality, originality, methodology and significance are outside its scope.
- It is not a verdict on the author. It reads one text, as written, at one point in time.
Limits
Language models make mistakes, including confident ones. Treat each review as a careful second reader, not an authority. That is why every review publishes the full record: the argument map, the step audit, the findings and the Advocate's rulings, so any reader can check how the grade was reached. Very long works are truncated before examination, and the review says so when this happens.
Right of reply
Authors are invited to respond. If a review misreads your work, misquotes it, or misses a step that answers an objection, write to contact@baris-atik.com with the review's address and your response. Substantiated corrections are applied and noted on the review, and a revised work can be re-examined on request.