Open problems · join
Stated precisely, because that’s how you find the people who can move them. Each one is live in the beta right now — a contribution here ships into a running system, not a simulator.
Whose judgment holds up, in which domain, learned from observed review behaviour rather than declared roles — and kept current as people join, leave, and grow.
A contribution looks likeAn online estimator of per‑person, per‑domain reliability that beats the org‑chart baseline on routing accuracy.
A reviewer approves the diagnosis and rejects the path. What exactly should update? Part‑level feedback gives us a factored signal the field’s scalar‑reward methods can’t use.
A contribution looks likeAn update rule that provably doesn’t degrade an approved part when an adjacent part is corrected.
98 has to mean 98 — while reviewers rotate, standards drift, and the work itself shifts. Static calibration curves die on contact with a real company.
A contribution looks likeA drift‑aware calibration method with bounded error under reviewer turnover, evaluated on the live ledger.
The tech lead and the support lead disagree, and both are right in their own domain. When is disagreement a tie to break, a hierarchy to respect, or a question to escalate?
A contribution looks likeA resolution policy grounded in social choice that reviewers judge fair — measured, not assumed.
What form must learned judgment take to survive a harness swap — Discord to Teams, one executor to another — without retraining, and without dragging the old surface’s assumptions along?
A contribution looks likeA spec‑level representation with a demonstrated round‑trip: learn on surface A, perform on surface B, no measurable loss.
Human approval on the diagnosis; execution outcome on the path. Two independent supervision signals over one artifact — sometimes agreeing, sometimes not. Almost nothing published uses both.
A contribution looks likeA method for reconciling the two signals that outperforms either alone on first‑pass acceptance.
Join
We’re a small team building the founding research group. Research‑assistant positions are open now; the strongest application is a half‑solution to any problem above. We answer everything that engages with a problem seriously.
You’ll work on the live system: real teams, real feedback ledgers, results that ship and publish. Remote‑first. If you’re earlier in your career than you think this requires, apply anyway — precision beats credentials here.