Reproduction
The agentic lab that can prove what it did.
Labrynn is an autonomous research environment for reproducible work: reproduce papers, build datasets, run experiments, train models, verify claims, and preserve every step behind the answer.
Reproduce
papers
Run
experiments
Verify
claims
Memory 5.2 / 16 GBCPU 10 coresSandbox 30s cap
Investigation
read paper.pdf · rebuild method · rerun tests
Can you reproduce this result locally?
Rebuilt the method, reran the analysis, generated the artifact, and found one figure that does not reproduce.
Gate verdict
paper parsedpassed
method rebuiltpassed
claims groundedpassed
artifactcreated
Product
Not a chatbot. A control room for autonomous work.
The app shows what matters during research: files, memory, tool calls, charts, artifacts, gates, and the exact trail behind an answer.
Artifacts
Measured charts stay tied to the rows that produced them.
Search
Moonshot and model candidates are compared from evidence.
Workspace
Files, transcript, sources, and run state remain visible.
Capabilities
From paper to proof, not prompt to prose.
Labrynn is built to do the work around the answer: reproduce papers, create datasets, run experiments, train baselines, compare hypotheses, and keep the evidence trail.
Paper reproduction
Upload, name, or link a paper. Labrynn extracts the method, rebuilds the data path, reruns the analysis, and reports what reproduces.
Research reproducibility
Every run keeps the files, code, tool calls, outputs, artifacts, evidence refs, caveats, and final claims needed to inspect it later.
Experiment engine
Turn hypotheses into local experiments with controls, alternate explanations, falsification checks, and repeatable outputs.
Dataset construction
Fetch, clean, reshape, join, validate, and export datasets instead of only commenting on them.
Model building
Train baselines, compare candidates, preserve metrics, and keep models only when they beat the prior result.
Moonshot search
When no known route exists, generate ideas, test them, reject weak ones, and keep searching for stronger approaches.
Harness
The moat is the system around the model.
Models get better inside Labrynn because the harness owns the boring, critical parts: state, execution, paper-to-code reproduction, verification, recovery, persistence, and replay.
Autonomous investigation
Labrynn plans the work, chooses tools, writes scripts, runs code, checks failures, and keeps moving until the question is actually handled.
Evidence-gated answers
Claims are tied to observations, files, tests, artifacts, and run history. Unsupported conclusions are blocked or marked clearly.
Replayable research paths
The run graph remembers every step so results can be inspected, resumed, repeated, or turned into a durable workflow.
Moonshot
A discovery loop for answers nobody has written down yet.
Moonshot turns an impossible-looking request into a research program: map the frontier, build rival approaches, run local experiments, falsify weak ideas, and promote only what survives evidence.
new algorithmshardware-aware optimizationresearch strategymodel/data methods
Prior-art map
Find what already exists before claiming novelty.
Candidate swarm
Generate diverse approaches instead of one polished guess.
Held-out tests
Separate real improvement from overfit luck.
Evidence package
Preserve the code, data, failures, and promotion reason.
moonshot run
evidence first01
Map the frontier
Read the problem, prior art, baselines, constraints, and known failure modes.
02
Generate candidates
Create competing hypotheses, algorithms, experiments, and ablations.
03
Run the tournament
Execute tests, score held-out cases, falsify weak routes, and retry stronger ones.
04
Promote or reject
Keep only ideas that beat the baseline with evidence, caveats, and replayable artifacts.
promotion rule
No idea graduates because it sounds clever. It must beat the baseline, survive checks, and leave a replayable trail.
Baseline first
Known approaches are rebuilt before Labrynn claims a better route.
Held-out pressure
Candidates are judged on data or cases they were not tuned against.
Replayable trail
Every promoted idea keeps its code, evidence, artifacts, and refusal history.
Model agnostic
Small models should still stand a chance.
Labrynn adapts the handoff, context, tool schema, and state package to the model you choose.
8B
Slow apprentice
Can follow a narrow handoff and reuse durable evidence.
20B
Default worker
Enough reasoning for serious local work when the harness is strong.
120B+
Fast senior
Explores fewer dead ends because the harness already owns the proof path.
Access
Built for researchers who need answers they can defend.
Use it to reproduce papers, audit prior results, run local experiments, build datasets, compare models, and turn successful investigations into reusable research workflows.
Founder email: onkar@labrynn.com
Local-first workspaceResearch reproducibilityModel-agnostic adaptersDurable run memoryEvidence-linked claimsSandboxed executionArtifact recoveryPaper-to-reproductionMoonshot search
request access