Projects

Project · Live

Ink-on-bone engraving: a magnifying glass and a small balance scale resting over an open page of ruled text with indigo margin marks, and a footpath drawn into the page dissolving into nothing — instruments of discernment reading a text critically.

Jiddu

Truth is a pathless land. Jiddu is six tools in one for reading critically: a fallacy detector, a fact-check pipeline, a neutrality assessment, an AI-slop detector, and a paper explainer that unpacks a dense paper paragraph by paragraph at the reading level you choose — plus an adversarial review that treats one manuscript as a real submission, audits its claims and experiments, checks published prior work, and tests it against the standards of its target venue. Drop a URL, paste text, or upload a PDF and get a structured read on how the piece argues, what it claims, how it frames its sources, whether it reads machine-assembled, what each paragraph is actually saying, or what would keep the paper out of publication. All six run in the browser; the five text pipelines and the multi-stage review are also available over a REST API and through a hosted MCP server.

The name is a nod to Jiddu Krishnamurti, who declared in 1929 that truth is a pathless land — no creed, authority, or method can hand it to you; you have to get there yourself. That is the whole point of the tool: it doesn't rule a text true or false for you, it gives you the instruments to read it critically and decide on your own.

Jiddu adversarial-review screen at jiddu.app: six critical-reading lenses sit above a PDF-only review form. The selected Adversarial review lens asks for a manuscript, target venue, edition, track, paper type, review date, and model before running the review.
Open jiddu.app ↗rafaehlers/jiddu

What it does

  1. Fallacy detector. Highlights every passage that contains a logical fallacy, explains each in a Word-style margin comment, and assigns the whole piece a fallacy score from 0 to 100. Streams results progressively while the LLM is still writing. 46-fallacy catalog with a per-fallacy modal and a guide page at /fallacies.
  2. Fact-check. Pulls every verifiable claim (numbers, dates, quotes, causal and categorical assertions) out of the same text using a Claimify-style 4-stage pipeline. Each claim can then be sent through Perplexity Sonar Reasoning Pro for a verdict (supported / contradicted / mixed / unverified ), with citations classified into a 6-tier source-quality taxonomy.
  3. Neutrality assessment. Flags the framing, attribution and source-selection patterns that make a text partisan even when it doesn't commit any explicit fallacy and doesn't lie about facts. Eight issue types: loaded language, asymmetric attribution, false equivalence, selective omission, source asymmetry, both-sidesing, steering, genetic framing. Deliberately does not classify partisan direction; the goal is to surface patterns the reader can audit.
  4. AI-slop detector. Flags the stock phrases, manufactured drama, and templated structures that make prose read machine-assembled, and scores the whole piece from 0 to 100 (from reads human to assembly-line prose). 18 patterns across five families — stock wording, manufactured drama, templated structure, hollow substance, and format tics. Every finding quotes the exact span, names the pattern, and says what cleaner writing would do, without rewriting your text. It flags patterns, not authorship, since humans write slop too. The catalog is adapted from Peter Yang's MIT-licensed no-ai-slop, rebuilt as a trilingual catalog with a guide page at /slop-patterns and results at /s/[id].
  5. Paper explainer. Turns a dense paper or technical document into a plain-language read. Generates a whole-document overview first, then walks it paragraph by paragraph, tagging each with its role in the argument (Background, Objective, Method, Result, Interpretation, Limitation, Conclusion) and rewriting it at the reading level you pick — from child (ELI5, everyday analogies, no jargon) through high school and undergraduate up to expert (concise, peer-level, full terminology retained). Results live at /e/[id].
  6. Adversarial paper review. Treats exactly one final PDF as an autonomous submission. It maps claims, evidence, experiments, and visually complex pages; checks published literature; synthesizes a draft; attacks that draft for missed evidence, unsupported novelty, and inconsistent severity; then deterministically finalizes a gatekeeping review. The public result includes the venue standard applied, summary, stage findings, strengths, page-localized weaknesses, questions for the authors, reproducibility issues, a short path to publication, the final recommendation, and the literature consulted. Verified 2026 profiles apply official criteria for NeurIPS, ICML, ICLR, AAAI, IJCAI, and CVPR; other DBLP venues fall back to a clearly labeled general scholarly rubric.

API & MCP access

The five text pipelines and the adversarial review are available programmatically, not just in the browser — so you can wire Jiddu into your own workflows or hand it to an AI agent. Full docs at jiddu.app/api-access.

  • REST API. Five synchronous JSON endpoints — /api/v1/detect-fallacies, /api/v1/fact-check, /api/v1/assess-neutrality, /api/v1/detect-slop, and /api/v1/explain. Each takes a { text } body (40–2000 chars) with optional lang (en / pt / es) and, for the paper explainer, a level.
  • Asynchronous review API. Adversarial review has a separate job contract because a full review takes several minutes and runs in stages. POST /api/v1/adversarial-reviews accepts one uploaded PDF or a public pdf_url and returns a 202; GET /api/v1/adversarial-reviews/{id} reports progress and eventually returns the structured review.
  • Key-based auth. Keys are admin-issued (hashed, shown once) and passed in the X-Jiddu-Key header. Rate limits are per-key and return 429 with retry timing; fact-check additionally caps claims verified per request.
  • MCP server. A hosted Model Context Protocol endpoint at /api/mcp exposes the five text tools plus start_adversarial_review and get_adversarial_review to any MCP client via the public mcp-remote bridge — nothing to install. A standalone packages/jiddu-mcp package is available for local/offline use.

Methodology

The fact-check and neutrality pipelines are grounded in two recent papers:

  • Claimify (Metropolitansky & Larson, Microsoft Research 2025) for the 4-stage claim extraction.
  • Distilling Expert Judgment at Scale (Goldfarb, Hall, Fisher, Salam, Wilde; Forum AI / Stanford 2025) for the verdict assessment, the 6-tier source-quality taxonomy, and the neutrality framework.

Jiddu is not affiliated with either group.

Adversarial review uses a different methodology. LiteParse's academic mode preserves page-scoped Markdown, spatial reading order, selective OCR, complexity signals, and screenshots so criticism can point back to the manuscript instead of floating above it. Venue autocomplete uses DBLP for discovery, while verified venue profiles transcribe official, edition-specific criteria and keep the source URLs with the review. The five-stage pipeline separates evidence collection, literature research, synthesis, self-critique, and final rendering so the model must challenge its own novelty, severity, and venue-fit judgments before a draft becomes public.

Stack

  • Next.js 16 (App Router, Turbopack) + React 19 + TypeScript + Tailwind 4.
  • Prisma 7 with @prisma/adapter-better-sqlite3. SQLite file persisted in a Docker volume.
  • Multi-model via OpenRouter: GPT-5.4 mini as default, MiniMax M3 as a one-click alternative, Sonar Reasoning Pro for fact-check verification. Client-side fallback retries across the model chain if no SSE events have been emitted yet.
  • Layout-aware PDF review with LiteParse academic mode, selective page-image inspection, DBLP venue discovery, and versioned criteria profiles for major 2026 AI and computer-vision conferences.
  • Streaming via server-side ReadableStream emitting SSE; partial-json for incremental finding extraction so comments appear as the model writes them.
  • Trilingual UI (EN / PT-BR / ES) with browser-language detection on first visit and localStorage persistence. The LLM is also instructed to write in the chosen language.

Things worth noting

  • Verdict cache with per-type TTL. Claims are hashed (lang-prefixed sha256 of the normalized statement) and persisted. Numeric claims expire after 30 days, quotes and causal claims after 90, dates and categorical claims after 365. Re-verification is free for cached claims.
  • SSRF-protected URL ingestion. Every fetch hop is DNS-resolved and rejected if it lands on RFC1918, loopback, link-local, CGNAT, multicast, or reserved IPv4/IPv6 ranges (covers the cloud metadata 169.254.169.254). Body size capped at 5 MB.
  • Print-optimized companion routes. Every public viewer (/a/[id], /f/[id], /n/[id], /e/[id], /s/[id], /r/[id]) has a /print sibling that auto-triggers the browser print dialog so the user can save as PDF.
  • Bounded, inspectable paper review. Review accepts exactly one final PDF and refuses to silently truncate past 300,000 extracted characters. Only selected pages are rendered for visual inspection. The ledger, literature research, draft, and self-critique remain private intermediate artifacts; the final review is the single stored source for the web, Markdown, and PDF versions.
  • Admin panel with a feedback queue ("report wrong verdict"), metrics (verdict distribution, per-pipeline volume, cache state), runtime-tunable rate limits and cost caps, and a one-click "mark as reviewed" toggle that removes the yellow "not yet human-reviewed" banner from public viewers.

Status

Live at jiddu.app. Source is open under AGPL-3.0 at rafaehlers/jiddu. Try the demos in EN, PT, or ES without spending a token: /a/demo-en, /f/demo-pt, /n/demo-es, or the complete adversarial-review demo at /r/demo-en.

Built with

Next.js 16React 19TypeScriptPrismaOpenRouterLiteParse