Terms and disclaimers

1. What Jiddu does

Jiddu is an automated tool. It uses large language models (currently routed through OpenRouter to OpenAI, Anthropic, MiniMax and Perplexity Sonar) to (a) flag passages that may contain logical fallacies, (b) extract verifiable factual claims and offer a verdict based on a web search, (c) assess how neutral or biased a text is, (d) explain a document paragraph by paragraph at a chosen reading level, (e) flag AI-slop writing patterns (templates and stock phrasing — never a claim about authorship), and (f) produce an adversarial, venue-calibrated pre-submission review of one complete paper. All six workflows are available through the web app, public API and MCP; paper review uses an asynchronous persisted-job contract.

Jiddu does not employ human editors. Every output you see is produced by software. The tool is meant to assist critical reading — not to replace it.

2. No claim of truth

The verdicts ("supported", "contradicted", "mixed", "unverified") and the fallacy labels are best-effort automated judgements. They can be wrong — sometimes confidently wrong.

Jiddu does not assert that any statement, person, or document is true, false, lying, or deceitful. The labels are signals to prompt your own verification. Always read the cited sources and form your own conclusion.

Neutrality scores and the paragraph-by-paragraph explanations are likewise automated interpretations, not statements of fact — the explanations describe what a text says and how it is structured, never whether its claims are correct. Audience level changes only the prose; section-role tags are best-effort guesses.

Paper-review findings, novelty comparisons, severity labels, venue fit and the final recommendation are also automated judgements. They are not an official peer review, an acceptance prediction, or a substitute for a qualified human reviewer.

3. Use at your own risk

You are responsible for how you use the analysis and what you share or publish. If you republish a Jiddu result, you adopt that content as your own — Jiddu's authors are not responsible for downstream uses.

Do not submit confidential, proprietary, embargoed or personal data. Submitted content — including selected rendered PDF page images when visual inspection is needed — is forwarded through OpenRouter to the selected upstream model provider and stored under a shareable id.

Use paper review only on manuscripts you own or are authorized to process. Do not upload a manuscript entrusted to you under confidential peer review unless the authors and the venue explicitly permit third-party model processing.

4. Persons and public figures

Analyses about named persons (politicians, journalists, executives) are produced by software searching public web sources. The language Jiddu uses is intentionally cautious ("evidence does not support", "we could not find") — never adopt a stronger framing than the tool emits.

If you believe a verdict is unfair or unsupported, use the "report wrong verdict" button on the relevant claim. Repeated unfair flags against the same person may be removed at the operator's discretion.

5. Privacy

Each analysis or paper review is assigned a short id and stored on a server in Brazil. Anyone with the share URL can view the final result. There are no user accounts; private intermediate review stages are not exposed in the public viewer.

Venue autocomplete sends only the venue query — not the manuscript — to the public DBLP venue search. The literature-research stage searches public scholarly sources using technical claims from the manuscript; prompts prohibit searching by paper title, author name or other manuscripts by the same authors.

We log the request IP for rate-limiting and abuse detection. We do not sell or share that information.

Page traffic is measured via Google Analytics (gtag.js).

6. Methodology

Our claim extraction and verification pipeline is inspired by Claimify (Metropolitansky & Larson, MSR 2025) for the extraction stage, and by the evaluation framework described in Distilling Expert Judgment at Scale (Goldfarb, Hall, Fisher, Salam, Wilde — Forum AI / Stanford, 2025) for the verdict assessment, source-quality tiers and neutrality dimension. Our choice of 4 verdict classes (supported / contradicted / mixed / unverified) rather than the 5- or 6-class schemes used by some fact-checking outlets is informed by Sahitaj et al. 2025, who found that 3-class labeling outperforms 5-class for LLM-based fact-checking — the additional categories in the middle introduce ambiguity without improving accuracy.

Adversarial paper review builds a page-tagged claim/evidence ledger, researches published literature, synthesizes a structured review, challenges that review in a separate self-critique, and then finalizes it deterministically. Verified venue profiles are edition-specific snapshots of official criteria; the exact criteria and source URLs used are stored with each review. Unmatched venues use a clearly labeled general scholarly rubric, and venue policies can change after a profile is published.

Jiddu is not affiliated with those research groups, DBLP, or any venue whose criteria it summarizes. References and official venue pages are cited so users can audit the methodology and the standard applied.

The fact-check pipeline has been benchmarked against the PolitiFact human-labeled corpus (via the LIAR2 dataset, Apache-2.0). On 200 claims from PolitiFact's polar buckets, Jiddu's verdict matched the human verdict in 67.5% of cases overall and 81.3% on the unambiguous true / false / pants-on-fire buckets, with strict polar disagreement in only 4.5%. Full methodology, confusion matrix and disagreement analysis at docs/benchmark-politifact.md.

7. Contact

Operator: Rafael de Menezes Ehlers. For takedown requests, corrections, or other concerns, open an issue on GitHub.