A better agentic paper-maker — where the human pilots, and citation fabrication is structurally impossible.
Most AI research tools race toward more autonomy — write the whole paper, invent the whole bibliography, trust the machine. The predictable result is fabricated and mismatched citations slipping into real manuscripts.
Probator races the other way. It is a desktop cockpit for writing scholarly papers where a human stays in the pilot's chair and the tool's job is to make dishonesty impossible, not merely discouraged.
The human pilots, and fabrication is physically impossible.
If a feature does not make the human sharper or make fabrication structurally impossible, it gets cut. That's the whole product.
The name fits the mission: a probator is an assayer — one who tests metal or coin for genuineness — and, in old law, an approver who gives sworn evidence. Probator assays papers and citations for genuineness; the human approves.
One app, four fused modules:
- The Cockpit — the pilot's chair; the frame everything lives in.
- Citation Firewall — assays every reference against live scholarly databases (OpenAlex, Semantic Scholar, Crossref). Fabricated or mismatched citations are physically blocked from the draft, with evidence shown. Not flagged — blocked.
- The Adversary — a panel of adversarial reviewers (skeptic, methodologist, stats-cop, novelty-hawk) that hardens the draft before submission. The panel advises; the pilot renders the verdict.
- The Console — an embedded terminal co-pilot. It drives a real agentic CLI headless on your existing subscription (no API key), single-agent and read-only, so it can see the Lab and advise without ever silently changing it. Claude is live; the backend is provider-agnostic (Codex, Gemini, Kimi, z.ai and more are registered).
Pipeline: Idea → Literature → Experiment → Draft → Firewall → Adversary → Export. Every export ships with a colophon — a full provenance trail of every source, result, and decision. Honest by construction.
A Windows 98 silver window frame wrapped around a black-and-green phosphor terminal.
- Font: Fixedsys Excelsior — the real Win95/98 look.
- Chrome: the native title bar is hidden; Probator draws its own Win98 title bar with working minimize / maximize / close buttons, built from two-tone bevels.
- Palette: silver
#C0C0C0, navy title gradient#000080 → #1084D0, terminal black#000000, phosphor green#33FF66.
- Flet 0.85 — Python UI rendered with Flutter. Real-time desktop app, minimal frontend pain.
- Python 3.14, standard-library-first.
httpxfor scholarly lookups;pytestfor tests. - Free scholarly APIs for the Citation Firewall: OpenAlex + Crossref (no API key).
- Co-pilot backend: a headless agentic CLI (
claude -p) on your subscription — no API key, no per-token bill. Seedocs/adr/for the why.
# from the project root
python -m venv .venv
.\.venv\Scripts\python.exe -m pip install -r requirements.txt
.\run.ps1 # or: .\.venv\Scripts\python.exe -m probatorprobator/
__main__.py entry point: ft.run(main, assets_dir=...)
app.py assembles the cockpit; owns running state + screen routing
core/pipeline.py the Idea → Export pipeline (pure data)
core/models.py Paper + Source (a Source is born unverified)
core/provenance.py the append-only colophon
core/store.py folder-per-project persistence (Documents/Probator/<slug>/)
services/copilot.py provider-agnostic co-pilot over headless CLIs
services/firewall.py Citation Firewall — assay vs OpenAlex/Crossref, fail-closed
ui/… theme, win98 toolkit, rail, workspace host, screens, projectbar, console
assets/ Fixedsys font, the assay-check logo, icon generator
tests/ pytest suite (core spine, co-pilot, firewall)
The design ledger — the North Star and working rules — lives in CLAUDE.md;
the phase plan in ROADMAP.md; the decisions in docs/adr/.
Early build, but the intelligence is real now. See ROADMAP.md for the
full 0.1 → 1.0 ladder.
- v0.1 — Win98 + green-terminal cockpit running; assay-check branding.
- v0.2 — the spine: folder-per-project storage, append-only colophon, a real Idea gate.
- v0.3 — Citation Firewall: assay a DOI/title against OpenAlex + Crossref; real records are verified, fabricated ones are blocked (fail-closed).
- v0.4 (part 1) — Console wired to a real agentic CLI on your subscription, no API key.
- v0.4 (rest) — propose→approve claim editing, co-pilot web search, Literature stage.
- v0.5–v1.0 — Draft, Adversary, Experiment, Export + colophon, hardening.
GPL-3.0. Forks and derivatives stay open source.
