How Jiddu works

Jiddu offers six independent ways to examine a text or paper: fallacy detection, fact-checking, neutrality assessment, paragraph-by-paragraph explanation, AI-slop pattern detection and adversarial paper review. Live mode adds real-time transcription with selectable context, fallacy, fact-check claim and neutrality lenses. Jiddu Context applies all four lenses in one workspace with an entity graph and saved session timeline.

Analyses return structured, source-linked findings as they become available. Output follows the language selected in the interface, and each feature can be used independently.

The public form does not expose model selection. An operator chooses the primary model, automatic fallback and compatible OpenRouter provider routing in Admin Settings, keeping one consistent analysis policy across Jiddu.

Jiddu can fall back to the other model family when the selected provider is unavailable. Work already returned is preserved rather than discarded, so a visibly partial result should be treated as partial.

Perplexity Sonar is used exclusively for fact-check verdicts because it has built-in web search. Operators can route fast and deep checks separately in Admin Settings; it is not exposed in the public form.

Catalog of 46 fallacies covering relevance, ambiguity, presumption, emotional appeals, causal reasoning and evidence. Each fallacy has a stable id and a distinct color (golden-angle hue distribution) used to highlight passages. The full list is at /fallacies.

The analysis is deliberately conservative: a finding must quote the relevant passage, identify the pattern and explain the reasoning gap in context. Ambiguous wording should not be flagged unless the surrounding text resolves it.

Each finding receives a severity and contributes to a 0–100 document score. The score is a navigation aid, not the probability that a document is fallacious.

Jiddu tests representative detections and false-positive traps, but has not yet published a human-labeled accuracy benchmark for this feature. Every highlighted passage still requires human judgment.

Modelled on Claimify (Metropolitansky & Larson, MSR 2025). The prompt asks the model to walk every sentence through four stages: selection (keep only sentences with a verifiable claim — opinions, value judgements, rhetorical questions get dropped), disambiguation (drop sentences with multiple plausible readings when context doesn't resolve them), decomposition (split multi-fact sentences into separate claims without chasing infinite atomicity), and decontextualization (every emitted claim must stand alone — resolve pronouns, add the year for a number that depends on it, attribute a quote to its source).

Each emitted claim is typed: numeric, date, quote, causal, or categorical. The type informs how the verifier later searches for evidence.

The selected GPT-5.6 Luna or Gemma 4 31B performs extraction only; truth assessment is reserved for the separate verification step. Usage safeguards limit how much work a single document can fan out.

Each claim retains its source sentence and a self-contained statement so the reader can audit what will be checked. Extraction can still miss eligible claims.

Verification is opt-in. When the user chooses Verify claims, each claim is independently researched with the Perplexity Sonar model configured by the operator through OpenRouter.

The result is supported, contradicted, mixed or unverified, accompanied by confidence, rationale and cited sources. Low-confidence evidence is handled conservatively rather than forced into a polar verdict.

Every cited source is classified into one of six tiers (per Goldfarb et al. 2025): primary (official data agencies, court rulings, peer-reviewed papers), scholarly, think_tank, journalistic, commercial_content, informal. A separate boolean flag marks state-controlled outlets (RT, CGTN, Xinhua, etc. — but not BBC/NPR/Deutsche Welle, which are state-FUNDED but editorially independent).

Verdict language is intentionally cautious about named people and describes what the available evidence supports rather than accusing someone of dishonesty.

How accurate is this in practice? We benchmarked the pipeline against PolitiFact's human-labeled corpus. On 200 claims drawn from PolitiFact's four polar buckets (true / mostly-true / false / pants-on-fire, 2020-onward), Jiddu's verdict matched the human polar verdict in 67.5% of cases overall and 81.3% on the unambiguous true / false / pants-on-fire buckets. Strict polar disagreement — Jiddu calling a claim the opposite of the human — happened in only 4.5% of cases (9 of 200). The remaining ~28% were Jiddu returning mixed or unverified instead of a polar verdict, concentrated on PolitiFact's mostly_true bucket (where 58% became mixed — a structural overlap, not a miss, since both labels describe partial truth). Full methodology, confusion matrix and disagreement analysis at the link.

Recent verdicts may be reused for equivalent claims to reduce latency and unnecessary search cost. Cache lifetimes vary by claim type so changeable facts expire sooner than stable historical facts.

Users can request a fresh check from the claim card. Refresh and usage safeguards prevent repeated searches from creating unbounded cost.

Distinct from fallacy detection: a text can be fallacy-free and factually accurate while still using loaded framing, asymmetric attribution or one-sided sourcing. Neutrality surfaces those patterns.

Eight issue types are flagged at the sentence level: loaded_language, asymmetric_attribution, false_equivalence, selective_omission, source_asymmetry, both_sidesing, steering, genetic_framing. The model also classifies the document into one of seven text types (informational / opinion / political speech / advocacy / analysis / marketing / social_media) which informs how strict the audit is.

The document-level verdict is one of neutral / partisan_lean / explicitly_partisan / manipulation. Jiddu deliberately does not classify partisan direction (left/right, pro-X/anti-Y). The Forum AI paper that inspired the methodology requires a bipartisan expert panel for that step; without one, partisan labels are unreliable and harmful. We surface specific patterns the reader can audit instead.

There is no published human-panel benchmark for neutrality yet. The verdict is a navigation aid into auditable passages, not an objective measurement of political balance.

Explain takes an academic paper or technical document and produces a Paper Map plus a paragraph-by-paragraph plain-language explanation. The map identifies the paper type, contribution type, central question, contribution, approach, main answer, evidence and any limitation explicitly stated by the authors. Each map statement links back to supporting paragraphs. Each paragraph also gets a role tag (Background / Objective / Method / Result / Interpretation / Limitation / Conclusion) and an explanation written at the chosen audience level.

Audience levels: Child (short, concrete explanations with everyday words and no assumed background), High school (explicit reasoning with every field-specific term defined), Undergraduate (standard academic terminology and introductory research methods assumed), Expert (concise peer-level analysis that foregrounds assumptions, evidence strength and limitations). The level changes assumed knowledge, terminology, explanatory depth and prose structure; it does not filter evidence, fact-check claims or validate content.

Role tagging and the Paper Map are algorithmic. Paragraphs are explained in chunks, then a bounded final synthesis samples all semantic roles and uses every chunk summary to build the document-level map. Interdisciplinary papers or documents with non-standard structure may receive incorrect tags or map entries.

Unlike the general-text pipelines, Explain is optimised for structured academic papers. Results on blog posts, legal filings, or unstructured prose will be lower quality.

The current evals check role coverage, audience calibration and schema reliability, not factual faithfulness against expert annotations. Explain can simplify a caveat away or confidently misread a method; it should make the source easier to enter, never substitute for reading it.

The slop detector uses a dedicated catalog and returns highlighted passages, short explanations and a 0–100 density score.

The catalog covers 18 AI-slop writing patterns in five families: stock wording, manufactured drama, templated structure, hollow substance, and format tics. The full list with definitions and examples is at /slop-patterns. The pattern selection is based on Peter Yang's MIT-licensed no-ai-slop editing skill; the catalog wording, trilingual content and detection prompt are Jiddu's own.

It detects patterns, not authorship: humans can write formulaic prose too. Every finding quotes the exact span, names the pattern and suggests what clearer writing would do without rewriting the source.

The eval set covers representative positives, clean prose and the no-authorship guardrail. This feature answers “which stock patterns are present?”, not “who or what wrote this?”

Adversarial review accepts one final PDF plus target venue, edition, track, paper type and optional review date. The result includes a recommendation, summary, strengths, localized weaknesses, questions, reproducibility checklist, references and the venue standard applied.

Methodologically, Jiddu adapts the SPECS review dimensions — Story, Presentation, Evaluations, Correctness and Significance — and the staged synthesis/self-critique pattern described in AI-Assisted Peer Review at Scale: The AAAI-26 AI Review Pilot. Jiddu is an independent implementation with its own claim/evidence ledger, venue calibration, parser, output contract and safeguards; it does not reproduce the AAAI-26 system.

The review examines Story, Presentation, Evaluations, Correctness and Significance, checks published literature and challenges its own synthesis before returning retained findings. It does not execute manuscript code.

On Jiddu's 30-case synthetic SPECS benchmark, GPT-5.4 mini, GPT-5.5, GPT-5.6 Luna and GPT-5.6 Luna Pro detected every injected defect. Luna Pro produced 8% fewer findings than standard Luna, but the benchmark reached a ceiling and does not prove superiority on complete papers.

Venue discovery uses DBLP and verified profiles for selected conferences and journals. Verified profiles apply cited official criteria; unmatched venues use a clearly labeled general scholarly rubric rather than guessed requirements.

Every substantive weakness must carry a manuscript page or section and a MAJOR, MODERATE or MINOR severity. The review is still automated: it can miss problems, invent comparisons or misapply a criterion, and it is not an official peer review or acceptance prediction.

At /live, Jiddu transcribes a microphone or, on Chromium browsers, a browser tab with gpt-live-transcribe. During capture, audio travels directly from the browser to OpenAI over an encrypted Realtime connection; Jiddu receives the transcript needed for analysis, not the raw audio. After capture, the user can download the browser-local recording or, after confirming the additional cost, explicitly send a temporary copy through Jiddu to OpenAI's gpt-4o-transcribe-diarize for speaker labeling. Jiddu does not retain that audio file. The user may separately save the labeled text and edited speaker names to private account history.

GPT-5.6 Luna or Gemma 4 31B periodically analyzes the growing transcript through the selected Live lenses: context guide, logical fallacies, check-worthy claim extraction and neutrality assessment. The user chooses a 10, 20, 30 or 60-second interval; longer intervals give each analysis more context. Results stream into an All feed and dedicated panels, with a badge identifying each type.

After finalization, each selected lens receives a shareable full viewer. Detected claims can also be verified against web sources directly inside Live, with evidence-backed verdicts and sources, or from the full fact-check viewer.

Sessions are capped at 60 minutes, use Jiddu Credits or included access, and store only the transcript and selected analysis results. Optional speaker identification is billed separately by audio duration. Tab audio is supported in Chrome and Edge; microphone capture works in other compatible browsers.

At /context, Jiddu turns a microphone or Chromium browser tab into a live workspace. gpt-live-transcribe provides the shared transcript, while GPT-5.6 Luna or Gemma 4 31B analyzes each interval so the graph and panels update as the conversation continues.

The context graph maps people, institutions, organizations, countries and events explicitly mentioned in the transcript, along with the relationships stated between them. Each card links back to the relevant passage. Listeners can correct a name, merge duplicates or remove an incorrect entity; later analysis preserves those manual corrections.

Topics, entities, check-worthy claims, sources, fallacy findings and neutrality findings remain separate inside the workspace. Public profiles and recent related coverage add background, but related coverage is labeled as context rather than proof. Claim verification is an explicit step and returns a verdict with cited sources.

When the session ends, Jiddu recalculates the final neutrality assessment from the complete transcript and saves the transcript, graph, findings and event timeline to the user's private account history. Optional speaker labeling sends a temporary audio copy to gpt-4o-transcribe-diarize; Jiddu does not retain the audio file.

WorkOS AuthKit provides user accounts and sessions. Completed results remain available through unlisted share URLs, and signed-in users can also see their associated analyses in the account area. Anyone with a share URL can still view its final result. Paper-review ledgers, research and draft judgements are internal intermediate artifacts and are not exposed in the public result.

Jiddu Credits are an application balance for web, account-owned REST and hosted MCP analyses that use Jiddu's server-managed provider credentials. Verified accounts on the operator-managed platform allowlist receive included access instead.

Submitted content goes through OpenRouter to the upstream serving the chosen model. Pasted screenshots are sent for transcription, but Jiddu does not retain the image file; its extracted text is stored with the analysis. Paper Explainer and paper review can include selected rendered PDF page images, not only extracted text, when figures or tables need visual inspection. The review literature stage searches public scholarly sources using technical claims while prompts prohibit queries by paper title, author name or other manuscripts by the same authors. Venue autocomplete sends only the typed venue query to DBLP. We log the request IP for rate-limiting and abuse signalling and do not sell it.

Results are automated and unreviewed by default. A result explicitly marked as reviewed has been spot-checked by the operator; that status is an additional human check, not a guarantee that every finding is correct.

Jiddu Credits pay for analyses across the web app, account-owned REST API keys and hosted MCP. Before most work begins, Jiddu temporarily reserves the estimate shown below so concurrent jobs cannot spend the same balance. Live instead checks that the account has the required minimum balance without holding or charging it. Metered usage settles against the work actually completed, and any unused reservation is returned to the account. Some verified accounts have included platform access and do not consume their personal credit balance.

A reservation is an estimate, not always the final debit. Long documents can contain multiple chunks, fact-check verification is charged per eligible claim, and Live analysis is charged per selected lens and completed analysis window. Live transcription is billed from actual audio duration at OpenAI's configured per-minute rate. Provider retries, reasoning usage or an unusually expensive completion can leave the account negative; new analyses then stay blocked until credits are added.

Paid credit packs are sold as one-time purchases through Stripe Checkout. Credits are granted only after Jiddu verifies the completed payment with Stripe; the credit ledger and transaction history remain the source of truth for purchases and usage.

Per-account and per-request safeguards limit abusive or accidentally unbounded use. Limits vary by feature and may evolve as model capacity and provider pricing change.

Every fact-check claim card includes 🚩 Report wrong verdict. The report records the verdict the user saw and an optional explanation for private review.

Reports do not automatically change a verdict or outweigh the cited evidence. They inform manual review and future quality improvements.

The text analyses are available through a public REST API, while adversarial review runs as an asynchronous job suitable for longer agent workflows. Requests are tied to a verified Jiddu account and use included access or Jiddu Credits.

The hosted MCP server exposes the text tools plus adversarial review. Compatible clients use WorkOS AuthKit MCP for OAuth sign-in and resource-scoped access.

Endpoints, payloads, authentication instructions and setup snippets are documented at /api-access.

Claimify informs claim extraction, and Distilling Expert Judgment at Scale informs verdict assessment, source tiers and neutrality. no-ai-slop supplies the starting taxonomy for writing-pattern detection.

Adversarial paper review adapts the SPECS dimensions and staged synthesis/self-critique described in AI-Assisted Peer Review at Scale: The AAAI-26 AI Review Pilot. Jiddu adds its own ledger, LiteParse extraction, DBLP discovery, venue profiles, source validation and retention gate. It is not affiliated with those groups, DBLP or the venues.

Published evaluation currently consists of the 200-claim PolitiFact benchmark and a 30-case synthetic SPECS benchmark. Both have documented limits and neither guarantees correctness on a new text or paper.

The full disclaimer about what Jiddu does and doesn't claim is at /legal.