We’re training an AI Research Agent
to wrestle with the hardest open problems.

One small theorem, one spectrum, one gradient step at a time.
The goal: generate_paper(agent, human) in ⌈ 1 day ⌉ — an AI-drafted paper, refined by a human expert, on real problems in math & astrophysics.

Latency budget for v0: 24h / paper

Stack: open-source LLMs + symbolic engines + simulators

agent-core: online

node: qpu-math-astro-01

lab@daemon:~$ bootstrap_agent --domain="math, astrophysics" --mode="paper-loop"

[log] scanning open problems… 42 candidates found

[select] focusing on spectral methods for exoplanet atmospheres

[plan] t=0 → t=24h: co-write 1 paper with human expert

[safety] proof-sketch verifier: enabled, simulation sandbox: armed

[status] agent warm-up… sampling ideas…

[warn] humans still required to check the math.

[err] no coffee process detected — add human.

paper_loop_progress 0.23

human_review_required True

Mission objectives

1. Tackle real open problems

Not toy datasets. We care about conjectures, spectral gaps, exoplanet signals, and messy real-world data. If it isn’t painful, it’s not interesting.

functional analysis Bayesian inference exoplanet spectra

2. AI that writes like a collaborator

The agent drafts sections, runs experiments, and cites sources — you do the deep thinking, sanity checks, and “this is actually wrong” corrections.

symbolic + neural tool-using LLM proof sketches

3. One-day AI→human paper loop

Constraint: from zero to shareable draft in ~24 hours. Version 0 will be rough, opinionated, and occasionally wrong — on purpose.

24h deadline human in the loop open-source first

Day-1 loop: from idea → AI draft → human expert iteration_0.alpha

T+0h

Select a concrete math/astro problem and freeze the scope. Define assumptions, dataset/simulation, and success criteria.

in-progress

T+8h

Agent generates experiments, sketches proofs or models, and writes a noisy first draft: abstract, intro, methods.

queued

T+16h

Human expert reviews: fixes math, rejects nonsense, adds insight, and tags failure modes for the agent to learn from.

requires human

T+24h

Share a joint AI↔human paper draft. Even if it’s imperfect, it must be honest, reproducible, and falsifiable.

v0 target

We’re training an AI Research Agent to wrestle with the hardest open problems.

1. Tackle real open problems

2. AI that writes like a collaborator

3. One-day AI→human paper loop

We’re training an AI Research Agent
to wrestle with the hardest open problems.