Skip to content

Coarse (coarse.ink)

external · status: active · focus: review · discipline: general · started: 2026

Project page: https://coarse.ink/

Source: projects/landscape/coarse-ink.yml

Positioning

A web-based AI peer-review service: users upload academic papers (up to 50 MB) and receive AI-generated referee reports with 20+ detailed comments. Operates in two modes — web-based using user-supplied OpenRouter API keys, or local processing via Claude Code / Codex / Gemini CLI subscriptions. Sits squarely in the referee-simulation stage of the RISE pipeline.

Distinctive contribution

Open-source (MIT) and explicitly anti-commercial in framing — "Academic peer review runs on unpaid academic labor. Others decided to make a business out of that. We didn't like that." The site reports a blind self-evaluation against refine.ink, Stanford Agentic Reviewer, and reviewer3.com claiming higher coverage, specificity, and depth. Privacy-by-design: API keys remain in the user's browser tab and clear on close.

Evaluation scores

Dimension Score (0–3) Note
Lifecycle coverage 0 Single stage (referee simulation).
Autonomy level 2 Supervised: user uploads paper and receives a report; optional 'reviewer focus' notes steer emphasis.
Architectural transparency 2 Open source under MIT; multiple back-end models exposed (Claude, GPT-5, Gemini, DeepSeek). Internal orchestration not deeply documented in marketing.
Inputs supported 2 PDF + optional focus note, plus a new optional --deep-literature-search mode (Perplexity Sonar Pro / Sonar Deep Research) that pulls in external literature during review — confirmed on the underlying GitHub repo (davidvandijcke/coarse, MIT, 182 stars, v1.9.2).
Outputs / reproducibility 1 Reports are persisted client-side; not deterministic across runs by design (different model choices).
Internal evaluation 2 Self-reported blind evaluation vs. refine.ink, Stanford Agentic Reviewer, reviewer3.com, plus a new independent third-party benchmark (Nguyen et al., 'Benchmarking Agentic Review Systems', arXiv:2606.19749, submitted 2026-06-18) that evaluates Coarse alongside OpenAIReview, Reviewer3 and a zero-shot baseline on a citation/acceptance quality-proxy study and an error-injection benchmark across six LLMs — Coarse is not reported as the top performer, and the paper is an arXiv preprint, not yet confirmed peer-reviewed.
Openness 3 MIT-licensed; BYOK (bring-your-own-key) model; transparent pricing (~under $2/review with OpenRouter).
Maturity / traction 1 Hosted service presumed active in 2026 — direct fetch of coarse.ink is blocked by this session's network egress policy, so liveness could not be directly reverified, but the ecosystem is corroborated by third-party coverage (e.g. manusights.com reviews of refine.ink referencing Coarse) and by the arXiv:2606.19749 benchmark; underlying open-source repo (davidvandijcke/coarse) now shows 191 GitHub stars (up from 182) and release v1.9.4 (up from v1.9.2); hosted-service user base still not publicly disclosed.
Cross-family policy 1 BYOK OpenRouter exposes Claude / GPT-5 / Gemini / DeepSeek as user-selectable executor — cross-family by user choice.
Runtime assurance 1 Single-pass review with focus-note steering; no published claim-audit harness.
Cross-platform portability 2 Multi-provider via OpenRouter + local Claude Code/Codex/Gemini CLI fallback.

Scored on 2026-10-01. See the evaluation rubric.

Tags

Pipeline stages: referee-simulation

Architectural features: tool-use human-in-loop rag-knowledge-base

Inputs: submitted-paper-pdf reviewer-focus-note

Outputs: referee-report

Knowledge sources: web-search

Limitations

  • Single-stage tool; needs upstream paper provenance.
  • Quality varies with chosen back-end model.
  • Self-evaluation against competitors not independently audited.