FeatureBoard U · for universities & research groups

Research that outlives
the people who did it.

Your lab's hardest knowledge lives in one student's head and three years of Slack. Then they graduate. FeatureBoard keeps the decomposition, the evidence and the dead ends where the next person can find them — and makes the difference between proved and assumed something you can point at rather than argue about.

Free for open-source academic work Purchase orders & volume terms Your data stays yours Runs on your machines

Three problems every research group already has

Turnover erases context

Students graduate, postdocs rotate, funding cycles end. What survives is code nobody can situate and a thesis chapter that does not say which approaches were tried and abandoned.

"Why didn't they just use the obvious method?" — because someone tested it in 2024 and it failed. Nobody wrote down where.

AI-assisted work has a trust problem

Your group is already using AI, whether or not there is a policy. The bottleneck is no longer producing results — it is establishing which ones survive scrutiny, and showing your working to a referee who is right to be skeptical.

Reviewers cannot audit a chat transcript. They can audit an artifact.

Status quietly drifts

A result is "basically done" for eleven months. An assumption made in week three is load-bearing by month nine and nobody remembers it was an assumption.

The count of things you are still assuming should be a number, not a feeling.

What that looks like when it works

CASE STUDY · one working day · open research problem

A proof with nothing left assumed — and a full record of how it got there

A chain of analytic estimates from an open number-theory program, formalised in Lean 4 and driven to a machine-checkable artifact. Mathlib had no Fejér kernel, so the hardest lemma was built from scratch across nine files, each one pulled from the board as its own unit of work.

79theorems verified
0admitted assumptions
12+sessions, no shared memory
1working day

The mathematics is not the point. The point is that the work resumed cleanly a dozen times because the board held what was proved, what was still assumed, and which approaches had already been ruled out — and that the finished claim is checkable by a machine rather than taken on trust.

Read the case study → Explain the maths at my level The wider research program

Who it is for

Principal investigators

See where a project actually stands without scheduling a meeting to find out. Open problems stay visible instead of becoming folklore.

Graduate programs

Students learn decomposition by doing it. A thesis that ships with its own failure log is a better scientific document — and a far better handover.

Research software engineering

Provenance from claim to commit. When a reviewer asks how a figure was produced, the answer is a ticket and a log, not an archaeology project.

Labs adopting AI seriously

Put the verification step in the pipeline rather than in someone's memory, so AI-assisted results arrive with evidence attached.

What changes

Academic needWhat the board does
Continuity across turnoverDecomposition, decisions and dead ends persist outside anyone's head. A new student reads the log instead of interviewing an alum.
Reproducibility & provenanceEvery claim traces to the ticket that produced it and the check that confirmed it.
Status that does not driftPhases close individually. Partial progress reads as partial, which is what a progress report is supposed to say.
Artifact-grade evidenceVerification runs in the build and emits a log you can hand a referee, not a claim you have to defend.
Supervision at a distanceStatus is a read, not a meeting.

Academic integrity, in the AI era

Four working rules from the case study. They are the reason the result is worth anything, and they generalise well beyond formal mathematics.

A gap you renamed is still a gap

Splitting one open problem into five smaller ones is not progress. The number of things you are assuming either goes down or it does not.

Absence of warnings is not evidence

A result can inherit an unproved assumption silently. Only a check that walks the whole dependency graph settles it — so put that check in the pipeline, not in your memory.

Test the claim before building on it

In the case study, a numerical check written before the proof killed the intended approach in seconds. It would have cost a day of formalisation to discover any other way.

Record failures at full strength

A disproved conjecture and a retracted claim are in that project's log at the same weight as the wins. That record is what stops a later student walking the same dead end.

Procurement

Start with one project

Pick the one where the plan is genuinely unclear — the problem nobody has an obvious ten-step recipe for. That is where a board earns its keep. If you already know the ten steps, you need a checklist, not this.

Get started free → Book a walkthrough See more work

FeatureBoard U is our name for academic use of FeatureBoard; it is not an accredited institution and confers no degrees. Figures on this page come from a single documented project — the case study links to its build output and audit log so you can check them yourself.