Skip to content

Reviewer (Ingar30)

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

Project page: https://github.com/Ingar30/reviewer

Licence: MIT

Source: projects/landscape/reviewer.yml

Positioning

A reproducible multi-agent reviewer for academic economics papers. Takes a PDF as input, runs a fixed set of mandatory reviewers plus dynamically-selected optional reviewers, validates their JSON output against schemas, deduplicates findings, and assembles a final editor report. Sits firmly in the referee-simulation stage of the RISE pipeline.

Distinctive contribution

Treats peer review as a structured engineering problem with explicit schemas, validation, parser-quality preflight, and editor-bundle assembly — distinguishing itself from chat-style review assistants. Reviewer prompts and orchestration are open and reproducible; review outputs are kept private by default.

Evaluation scores

Dimension Score (0–3) Note
Lifecycle coverage 0 Single stage (referee simulation).
Autonomy level 2 Supervised: user provides a PDF and receives a final editor report.
Architectural transparency 3 Open source under MIT; reviewer prompts, schemas, validation, and orchestration all in the repo.
Inputs supported 1 Accepts paper PDFs; no integration of external literature corpora at present.
Outputs / reproducibility 2 Schema-validated reviewer JSON and structured editor report persisted; deterministic given fixed model + prompts.
Internal evaluation 1 Smoke checks on report structure and traceability; no published systematic benchmark yet.
Openness 3 MIT-licensed; setup scripts for Windows and macOS/Linux; reproducible workflow.
Maturity / traction 1 174 stars and 35 forks by 2026-09-30, with active commits through 2026-09-24 (95 total) and a new experimental Claude Code backend alongside the default Codex CLI — accelerating adoption for a 5-month-old solo project, but still no tagged releases; remains a single-developer research prototype.
Cross-family policy 0 Single-family per run; the new --backend claude path swaps the whole pipeline to Claude Code rather than mixing families within one review.
Runtime assurance 2 Parser-quality preflight + schema validation + smoke checks gate the editor report; no claim-audit pass.
Cross-platform portability 1 Now runs on two agent runtimes — Codex CLI (default, GPT-6 Sol/Luna) and an experimental --backend claude path via Claude Code (Opus 5.5) — plus Windows + macOS/Linux setup scripts.

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

Tags

Pipeline stages: referee-simulation

Architectural features: multi-agent dag-orchestration tool-use artifact-versioning

Inputs: submitted-paper-pdf

Outputs: editor-report reviewer-json parser-artifacts

Limitations

  • Single-stage tool (review only); requires upstream paper provenance.
  • Economics-paper focused; portability to other disciplines untested.
  • Depends on Codex CLI; not a fully local solution.
  • Source PDFs and review outputs are private by design; no shared evaluation corpus.