Skip to content

Refine (refine.ink)

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

Project page: https://www.refine.ink/

Source: projects/landscape/refine-ink.yml

Positioning

A commercial AI peer-review service that produces reviewer-grade feedback on academic papers within ~20–40 minutes by running multi-hour parallel compute jobs (~2+ hours per review). Targets four error classes: accuracy (statistical / methodological), mathematical reasoning (proof gaps, edge cases), internal consistency (text↔tables↔citations), and general rigor. Sits in the referee-simulation stage of the RISE pipeline.

Distinctive contribution

Positions itself as enterprise-grade with explicit security and privacy commitments (SOC 2 + ISO 27001 in progress, zero-retention contracts, papers never used for training). Markets adoption by Oxford / Stanford / Yale / MIT / Caltech / Cambridge / Brown / OpenAI researchers. Closed-source commercial offering; first document free.

Evaluation scores

Dimension Score (0–3) Note
Lifecycle coverage 0 Single stage (referee simulation).
Autonomy level 2 Supervised: user uploads, system returns a structured review.
Architectural transparency 1 Marketing-level descriptions only; internals not publicly documented.
Inputs supported 1 PDF inputs; no integration of literature corpora or co-author context.
Outputs / reproducibility 1 Reports persisted to user account; not designed for byte-level reproducibility.
Internal evaluation 2 Aug 2026 self-published benchmark (150 economics preprints, 1,349 head-to-head matches vs. 9 competitor systems): Refine won 90.4% overall (94.8% vs. single-shot LLM referees, 85.0% vs. scaffolded review systems), broken out by subfield — a systematic internal evaluation, though still self-conducted rather than third-party.
Openness 0 Closed-source commercial product.
Maturity / traction 3 2026-08-06: formal partnership with the American Economic Association and the Econometric Society for technical-verification checks on revise-and-resubmit papers; AEA-journal pilot reported ~90% author approval for incorporating Refine into the editorial process — production-grade deployment inside a major scholarly body's workflow.
Cross-family policy 0 Closed; single internal stack.
Runtime assurance 1 ~2-hour parallel compute per review implies multiple internal passes; mechanism not public.
Cross-platform portability 0 Closed commercial product; single web surface.

Scored on 2026-09-01. See the evaluation rubric.

Tags

Pipeline stages: referee-simulation

Architectural features: multi-agent tool-use

Inputs: submitted-paper-pdf

Outputs: referee-report issue-list

Limitations

  • Closed-source; cannot be audited, extended, or self-hosted.
  • Per-review compute cost passed to user via subscription; pricing not transparent on landing page.
  • Its head-to-head benchmark against competitor systems is self-published, not third-party or peer-reviewed.
  • Targets the same review niche as coarse.ink and reviewer3.com; differentiation is per-review compute intensity.
  • AEA/Econometric Society deployment is scoped to a technical-verification check only — it does not assess contribution, importance, or screen submissions en masse.