Commission
Research
All research areas →

Agent Evaluation and Coordination

Generated-Code and GPU-Kernel Correctness

Federated Learning and Privacy

Decentralised Systems and Protocol

Methods
Papers
Collaborate
About Contact Bring a technical claim

Collaboration · Industry and academia

Share the question, not the constraints

Labs and industry teams often want the same answer but cannot work the same way: one needs to publish, the other controls the data. I help design the collaboration around those differences before anyone writes code, and can contribute the systems and evaluation work myself.

The situation

We share a research question but have different access and publication constraints. What collaboration design could work?

What you leave with

A joint question map, a contribution split with named owners, and a brief listing every data and publication dependency and how it will be resolved.

For: Principal investigator; industry research lead

Where these collaborations usually break

Rarely on the question. They break on dependencies discovered late: data that turns out not to be releasable, a publication deadline that does not match the partner’s review cycle, or an authorship dispute because one side built the infrastructure and the other wrote the paper. All three are predictable, so the design work happens first.

Three documents before any code

  1. Joint question map. The shared question is split into sub-questions. Each is marked by what it needs (public data, partner data, partner compute) and who needs the answer (the paper, the product team, both). Sub-questions nobody needs are dropped.
  2. Contribution split. Every planned output gets an owner for design, build, run and write-up, and the authorship rule is applied to that record.
  3. Data and publication dependency brief. For each result: which data it rests on, who controls it, what can be released, how long partner review lasts, and what is published if the data is withdrawn.

Illustrative example — not a client engagement. A university group studies how retrieval-augmented systems handle conflicting sources; an industry team runs such a system on internal documents it cannot release. The question map separates “does the detector’s trade-off depend on input distribution?” (answerable on public probes and human-labelled public pairs) from “which regime does our corpus fall into?” (answerable only internally). The split gives the lab the public study and paper, the industry team an internal run of the same harness, and the dependency brief states that internal numbers appear in the paper only as an aggregate, after a two-week review limited to confidentiality.

What I can contribute

My work sits on the systems side: building the instrument that makes a question measurable. The papers linked below are my own preprints and manuscripts, not past joint projects, but each shows the kind of contribution a lab partner could build on:

  • An evaluation corpus and harness. In The Correctness Illusion, seeded fuzzing over a 26-op GPU-kernel corpus caught 10 of 10 seeded illusions with 16 of 16 controls clean on five GPU classes. The same corpus now drives four papers.
  • A measurement substrate. In Before the Pull Request, grite’s append-only log is both the coordination mechanism and the data source; duplicate work fell from 78% to 0%. grite is a Neul Labs project, and I founded Neul Labs.
  • An input-regime audit of an existing component, as in the RAG conflict-detection audit, where the detector trade-off turned out to depend on the input distribution.
  • A partial but verifiable release. For the on-device auction study the simulator source is withheld, but its hashes are recorded and the results, calibration data and analysis scripts are public: a pattern that suits partners who cannot release everything.

Independence, data rights and publication

I take part as an independent researcher. Conflicts are checked before scoping, and data rights and publication terms are agreed in writing before work starts. See publication, IP and independence for the defaults I start from.

Commission this if

  • A lab and an industry team share a research question, and at least one side needs the results to be publishable.
  • Your group owns the question but needs complementary systems engineering — a harness, simulator, corpus or measurement substrate — to answer it.
  • Data, compute or model access sits on one side and the publication incentive on the other.

Not the right fit if

  • You need engineering capacity with no research question attached. That is delivery work.
  • The industry partner wants the right to suppress unfavourable results. Review windows can be agreed; vetoes cannot.
  • You need a partner who can hold a grant or sign as an institution. See consortium contributions for what an independent researcher can and cannot do there.

What you receive

  • Joint question map. The shared question broken into sub-questions, each marked as answerable from public data, answerable only with partner data, or out of scope for this collaboration.
  • Contribution split. Who designs, builds, runs and writes each part, with the authorship and acknowledgement rule agreed before any work starts.
  • Data and publication dependency brief. For each planned result: the data it needs, who controls that data, what may be released, and the review window that applies before submission.
  • Fallback plan. What can still be published if partner data is withdrawn or a result is unfavourable — usually the method plus a public-data version of the experiment.

What the study needs from you

  • The question. A paragraph stating the research question and why each side cares about the answer.
  • Each side's constraints. Publication needs (thesis deadlines, venue cycles, open-access mandates) and confidentiality or data-governance limits.
  • Named leads. One person per side who can agree contributions and authorship.
  • An access description. What data, compute and models exist — described, not shared, at this stage.

How the work runs

  1. 1
    Question call. Understand the question, check conflicts of interest, and decide whether a collaboration is the right shape at all.
  2. 2
    Question map. Draft the map; each side marks the sub-questions it needs answered and the ones it can contribute to.
  3. 3
    Contributions and authorship. Agree the split and the authorship rule while nobody yet knows whose part will produce the headline result.
  4. 4
    Dependency brief. Settle data rights, release conditions, review windows and the fallback plan; the partners' own legal offices turn this into agreements.
  5. 5
    Work and review. Run the agreed parts. Changes to the split or the plan are logged with reasons and agreed by both leads.

Limits and unfavourable results

  • As an independent researcher I cannot act as principal investigator on a grant or sign agreements on behalf of a university. The academic partner's institution handles those.
  • The design documents are not legal agreements. Data-sharing, IP and publication terms are drafted and signed by the partners' own offices.
  • If the constraints are genuinely incompatible — a result that must be both published and confidential — the question map will show it, and the honest recommendation may be not to collaborate.

An unfavourable result is still a result. The dependency brief fixes in advance what happens to it, and partner review windows exist to check confidential data and factual accuracy, not to remove findings.

Engagement terms

Model
Agreed per project: co-authored research where the question warrants it, a costed contribution inside a funded project, or a fixed-scope contract for the design work alone.
Who does the work
Dipankar Sarkar personally does the design work and any contributed research. Any specialist help is disclosed and agreed in advance.
Commercial basis
Agreed per project. Where a fee is paid, it is fixed at scoping and does not depend on the result.
Availability
Checked per enquiry.

Conflicts are checked before scoping. See publication, IP and independence.

Evidence behind this page

  • gpuemu-corpus (GitHub) — A shared evaluation corpus with drivers for four papers, stored seeds and replay scripts — the kind of infrastructure a lab partner can build on. MIT or Apache-2.0.
  • grite (GitHub) — Git-native coordination log used as both mechanism and data source in a multi-agent study. MIT. A Neul Labs project; I founded Neul Labs.

Questions

How should contributions and authorship be agreed when one side builds infrastructure and the other writes the analysis?

Before the work starts, and per output rather than per project. The contribution split names who designed, built, ran and wrote each part, and the authorship rule is applied to that record. A recognised contributor-role taxonomy helps, but the main protection is agreeing it before anyone knows which part will carry the headline result.

Can an independent researcher collaborate with a university lab?

Yes, as a co-author or contributor on a defined part of the work. I cannot be the principal investigator, sign institutional agreements or hold grant funds; those sit with the academic partner. My connections are disclosed up front: I founded Neul Labs, which builds AI agent infrastructure in Rust, and my research affiliation is Skelf Research.

What if the industry partner cannot share its data?

The question map marks which sub-questions need that data. Often the method can be developed and published on public or synthetic data, with the partner running it internally on their own data. If the publishable claim depends entirely on confidential data, that is settled in the dependency brief before work starts, not discovered at submission.

What does the industry side get out of a research collaboration?

Typically a working harness or corpus they can keep running, an internal report on their own data, and an external paper that tests the method in public. Which of these applies is written into the contribution split, so neither side assumes the other's output.

See also

Send a research question

Describe the question, who is on each side, and the constraints you already know about. A paragraph is enough to start.

Last reviewed 2026-10-07.