Topics
AI agents for scienceAI research agents for science
proved yet: no · human gates onAI research agents can help with real science work — reading, proposing tests, running computer experiments, checking numbers — if they stay honest about what is measured and what is not. That is the LumenAxiom bet.
AI research agents can contribute measurable scientific labor — literature synthesis, falsifiable proposals, simulation, analysis — when status flags stay ruthless. That is the LumenAxiom design.
Specialist helpers, one bossSpecialist agents under a Chief of Staff
A small team of AI helpers splits the work: papers, ideas, experiments, simulations, data, and writing. A boss helper assigns jobs so each stay focused and checkable. You can peek at the playful floor on Mission Control; claims still belong on Live Findings.
Specialists cover literature, hypotheses, experimental design, simulation, data analysis, and writing. A Chief of Staff routes work so each job stays focused and auditable. Mission Control is the ops floor view; authoritative claims live on Live Findings.
Ruthless honesty (the differentiator)Ruthless honesty (the differentiator)
- Proved yet: no until written success rules clear — useful progress is not a declared win.
support_claim: falseuntil pre-registered §5 (or equivalent) gates clear — useful progress ≠ declared success.- No invented metrics, paper counts, citations, or GitHub links.
- Breakthrough badges are rare; “notable” just means look closer.
- Breakthrough badges only when marked and the track claim clears; “notable” ≠ breakthrough.
- Dangerous biology work is out of scope.
- Dual-use harmful research is out of scope.
Human gates stay onHuman gates stay on
Agents do not get a blank check. Money, external publishing, credentials, and anything physical need a person. That keeps the “autonomous research lab” story grounded: helpers move fast on computers; humans own irreversible steps.
Spend, external publish, credentials, and physical action require a human. Autonomy is bounded: agents accelerate compute-side loops; humans own irreversible steps.
FAQ
Do agents claim breakthroughs automatically?
No. Tracks stay at proved yet: no until written checklists clear. Finishing a simulation grid is progress, not a win.
No. Tracks keep support_claim: false until pre-registered criteria clear. Completing a grid (partial=false) does not flip the claim.
What can agents do without a human?
Read open papers, suggest testable ideas, run computer experiments, check data, and draft careful notes. Not spend, publish externally, touch credentials, or act in the physical world.
Literature synthesis, falsifiable proposals, simulation, analysis, and drafting. Not spend, external publish, credentials, or physical action.
Where are measured results?
Live Findings is the live measured feed. Files sit under /artifacts/. Method detail: How it works.
How is this different from typical AI science demos?
Ruthless honesty first: no fake breakthroughs, no invented stats, open artifacts when we claim a number. See also open scientific discovery.
Honesty flags and artifact-backed claims first — not demo theater. See also open scientific discovery.
Home · How it works · About · Live Findings · Open scientific discovery