Paper Review Automation

What it does

A scheduled, file-driven pipeline built around one hard rule: it never deletes, moves, or renames a client's files.

🕘

Fully scheduled

Fires daily via Windows Task Scheduler. Walks every employee and client folder, decides what's eligible, and processes it — no manual trigger needed.

🔒

Zero delete/rename/move

Enforced by an AST-checked test, not just a convention — tests/test_safety.py fails the build if that invariant is ever broken.

🧪

Scientific integrity checks

Every revision is diffed against the original and flagged for changed metrics, invented numbers, leftover placeholders, or suspicious truncation.

👥

Role-based LAN dashboard

Admins see every employee's papers and settings; Users see only their own — enforced server-side, not just hidden in the UI.

📜

Full audit trail

Every paper's timestamp, phase, model, outcome, and any error is recorded in SQLite and queryable directly.

♻️

Idempotent & crash-safe

Eligibility is recomputed from disk every run. A crash halfway through just picks back up — it can't double-process a paper.

🖥️

Tray app + desktop icon

A system tray icon and one-click desktop shortcut keep the dashboard always reachable, without a console window.

📦

Portable installer

A PyInstaller + Inno Setup build packages five executables under one per-user installer — no Python environment required on the target machine.

How it works

Two phases, both driven by CLI subprocesses reaching your Codex and Claude Code subscriptions — never a direct API call.

09:00 daily

Scan

Walk every employee → client folder for the current month, apply the file-count rules to decide what's eligible.

→

Review · Codex

paper.docx → review

Converted to markdown, reviewed to a Q1 standard by the Codex CLI, written back as <Client>_review.docx.

→

Revise · Claude

paper + review → revision

Re-scanned from disk, then paper and review are sent to Claude Code, producing Correct_<Client>_paper.docx.

→

Validate

Check & log

The revision is validated against the original; anything suspicious is flagged REQUIRES_HUMAN_REVIEW, never silently accepted.

Screenshots

The LAN-only control panel — reachable only from your office network, never the public internet.

Sign inAccount requests need admin approval before first login.

DashboardThis month's papers by status, with a one-click Start.

Which AI is usedCodex finds the problems, Claude writes the corrections — plus the daily schedule.

System statusStorage, automation, schedule, workers, and network state at a glance.

InstructionsThe exact prompts sent to Codex and Claude — editable, with integrity rules always appended.

SettingsReview mode, model access, and per-stage model selection.

UsersEvery account and role, with disable/reset/rename/delete controls.

Built around not trusting the model blindly

Both prompts forbid inventing results, datasets, citations, or statistics — and the pipeline checks that, rather than taking the model's word for it.

🚫 A revision fails validation when it:

Reports a metric whose value changed from the original — the signature failure this guards against (an inflated accuracy, say)
Introduces a metric that wasn't in the original at all
Leaves placeholder text behind — [TODO], [INSERT RESULT], [ADD CITATION]
Comes back less than half the length of the original — likely truncated

⚠️ Warned, not failed

A disappearing dataset or model name, or a missing expected section, is flagged but doesn't block the file — every flagged paper is recorded REQUIRES_HUMAN_REVIEW, and the file is still written so a human can inspect it directly.

	Admin	User
See papers	Every employee	Own employee only
Start a run, change settings/schedule/prompts	Yes	No
Approve/reject sign-up requests	Yes	No
Open a file on the server	Any paper	Own only

Setup

Both models are reached through the Codex and Claude Code CLI binaries, authenticated with your existing subscriptions — no API keys anywhere.

Install & configure

py -m pip install -r requirements.txt
copy config.example.toml config.toml
# set research_papers_root in config.toml

codex login
claude

Try it safely

py run.py --dry-run      # reports decisions, changes nothing
py run.py --test-mode    # mock model, checks file naming
py run.py --provider real   # one real paper

powershell -File scripts\register_task.ps1  # schedule it

Drop a paper in a folder.
Wake up to a reviewed, revised draft.