Open Research Lab OS
Help turn learning into reproducible open AI work.
Continuous Function is building a public research lab workflow: learners move from concept notebooks into browser witnesses, toy scripts, reproducible experiments, evals, model cards, public reviews, and funded next work.
Open Research Lab OSLearn -> reproduce -> evaluate -> contribute
House rule
No claim without source, artifact, reviewer, status, and caveat.
Current wedge
One room first, then the institute.
The first ask is not to join a giant lab. It is to review one flagship research room and help make the evidence ladder honest enough that learners, contributors, reviewers, and sponsors can trust it.
Contribution ladder
You can help in 30 minutes, 3 hours, 3 days, or 3 months.
Leave one precise critique
Read the flagship room and point to the missing source, caveat, experiment, reviewer role, or learner step.
Best for senior reviewers and busy practitioners.Review one RFC or object
Inspect a claim, equation, demo witness, dataset note, eval proposal, or safety boundary and comment on the exact artifact.
Best for researchers, engineers, and educators.Ship one bounded contribution
Add a toy script, source panel, notebook repair, eval case, contamination check, model card section, or route handoff.
Best for maintainers and technical contributors.Lead one research room
Own a small track from learning route to reproducible experiment, benchmark, public report, and next contributor issues.
Best for guest lab leads and funders.First reviewer pilot
The first-10 reviewer protocol is ready; public signup is not open.
Target 10 reviewers for attention serving. Cohort not started; live discussion not enabled; public credit listing not live.
Launch reviewer
Senior engineers, AI educators, and research leads with little time.
Spend 30 minutes on the flagship room and name the single evidence gate that most weakens trust.Source reviewer
Researchers and practitioners who can inspect papers, docs, or benchmark claims.
Check whether a source supports the local claim and whether the caveat is honest.Learner tester
Students and self-learners who can describe where the route becomes confusing.
Use the room like a learner and report the first place the prediction, witness, source, or caveat feels unclear.Eval reviewer
Evaluation engineers, benchmark maintainers, and model reliability reviewers.
Review the candidate model-backed eval RFC and identify the smallest falsifiable task that would preserve caveats.Security reviewer
People who understand public contribution, workflow, supply-chain, data, and agent-run risk.
Mark what must stay private before public discussions, issue intake, or agent review become live.Pedagogy reviewer
Educators, learning scientists, and visual learning tool builders.
Review whether the learner loop makes the invisible system state easier to predict, inspect, and transfer.Each packet names an artifact, review question, evidence requirement, accepted output, and blocked output.
Credit does not grant authorship, funding, project role, merge authority, or agent-dispatch authority.
The form and roadmap API define intake shapes only; no live cohort roster, live server persistence, public signup queue, or public reviewer ledger is claimed.
Contributor tracks
Pick the smallest role that matches your leverage.
Launch reviewer
AI educators, senior engineers, research leads
Review the lab ladder and name the evidence gates that would make it credible.RFC reviewer
Post-training, data, eval, security, or pedagogy specialists
Comment on one proposal before it becomes implementation work.Research contributor
People who learn by reproducing papers and shipping code
Turn one claim into source notes, a toy witness, tests, and a next experiment.Data contributor
Dataset, licensing, deduplication, and contamination reviewers
Make one dataset or benchmark path auditable before anyone trusts the result.Eval contributor
Benchmark builders and applied AI evaluation engineers
Design failure cases, eval cards, prompt transparency, and leaderboard caveats.Security reviewer
Supply-chain, model, agent, and data safety reviewers
Mark what must stay private and what can be safely opened.Bounty sponsor
People funding public goods, compute, and open AI infrastructure
Sponsor one reproducible experiment, review queue, or contributor bounty.Learner
Students who want to become contributors instead of passive readers
Start with a concept, run the witness, write down the prediction, and file one useful issue.Flagship room
Attention to Serving is the first review target.
It already has a concrete route from attention math to KV cache memory, GQA/MQA, FlashAttention, long-context pressure, serving latency, and decoding controls. The review question is simple: what evidence gate is missing before this becomes a serious open lab room?
Open the roomConcept route: attention math, KV cache, GQA/MQA, FlashAttention, long context, serving latency, decoding.
Browser witness: learners manipulate the system before reading the full derivation.
Toy script: a small runnable check that mirrors the claim without hiding behind a framework.
Source boundary: every strong statement names what the source supports and what it does not prove.
Eval sketch: one falsifiable measurement and one contamination or misuse caveat.
Contribution issue: a small next task that can be completed without owning the whole platform.
Advisor map
Ask people for narrow critique, not open-ended commitment.
Practical AI educators and course builders
Does the learner-to-researcher ladder actually help people become capable?
Open model and post-training researchers
What would make the open-model room technically honest and useful?
Open data and reproducibility maintainers
What evidence should exist before a training or dataset claim is promoted?
Evaluation and reliability reviewers
How should eval artifacts avoid becoming misleading dashboards?
Visual learning and tools-for-thought reviewers
What would make the room feel explorable, memorable, and worth returning to?
Outreach packet
Send less. Make the ask easier to answer.
The best note is short: describe the public research lab OS, point to one room, ask for one critique, and make it clear that no operational ownership is expected.
A working demo URL.One public flagship room.One GitHub issue or RFC they can comment on.A one-page vision with no inflated claims.A narrow ask tied to their actual work.
I am building an open research lab OS where learners move through a machine-checkable ladder: concept, browser witness, toy script, reproducible experiment, benchmark, and open contribution. Could you give one critique on the part closest to your work?