Most reproducible computational claims fail in the hand-off: missing checkers, unreproducible scripts, or unverifiable randomness. V3R1T4S reframes agent-led discovery around "proof or it didn't happen": every run is designed to either produce a fully checkable artefact or to report a validated negative.
What Sets It Apart
- Automated verification-first pipeline: a hunting run pairs a null-test on hundreds of decoys with three independent skeptic re-checkers before a claim is accepted, so wins include the object, the generator code and a tiny verifier.
- Designed for reproducible table-breaking and record-search work: shipped checkers can be run locally (C checkers included for large graphs), and the repo stores snapshots of the beaten table entries and the exact code used to find them.
- Agent-agnostic integration: works by invoking agents that can run shell commands (Claude Code, Codex, Gemini CLI, Grok, Hermes and others); setup is local and nothing posts externally without an explicit final confirmation.
Who It's For & Tradeoffs
Great fit if you want to: run automated computational hunts that must survive independent verification; add agent-driven search to leaderboards or record tables; or produce reproducible computational artefacts for cryptography, graph theory, coding theory or ML benchmarks. Not a good fit if you need a turnkey cloud service—runs consume your local/cloud credits and you remain responsible for agent permissions, data use, and legal constraints.
Where It Fits
Positioned between agent orchestration tooling and reproducible-research kits: faster to start than building a bespoke agent pipeline from scratch, but more conservative than hosted auto-experiment platforms because it forces local control, null tests and triple skepticism before publishing a claim.
How It Works (brief)
- Setup creates a folder and asks six short questions (credit name, target domain, time per run, schedule, research charter sharing, permissions).
- The agent hunts a chosen problem, runs null tests on many decoys, spawns three skeptics (fresh contexts) to re-check promising leads, runs a mechanical checker and then asks the human owner for final approval before adding a trophy-case entry.
