An optimizer that reads the rules — in plain English
Assign 72 real papers to 16 reviewers, three reviewers each. The objective can be a semantic-embedding similarity (the incumbent) or a calibrated expertise-fit read from the abstract. And each reviewer can carry a rule written in plain English — “no theory papers” — that the optimizer must honor. Toggle one and watch the assignment re-solve, exactly, in your browser.
Understand exactly what you're looking at
The theory, how the data is acquired and cleaned, how to read the result — and what BIS would do for you next.
Assigning papers to reviewers under load limits is a classic capacity-constrained assignment — solvable exactly (here, by min-cost max-flow). The hard part was never the solver; it was the objective. You can only optimize what you can quantify, and “how well does this reviewer fit this paper?” lived in prose. The incumbent (the Toronto Paper Matching System) approximates it with text similarity — a heuristic, not a judgment.
A calibrated decision model changes what the objective can be. TypeSafe’s Jev reads the abstract against the reviewer’s expertise and returns a calibrated expertise-fit — and, just as importantly, it reads a reviewer’s rule written in plain English and returns a calibrated per-paper compliance. That turns an English sentence into a hard constraint the optimizer can honor — which no similarity metric can do.
Real arXiv papers across four areas (cs.LG, cs.CL, cs.CV, cs.CR). Each reviewer is a real researcher’s expertise; conflicts are real co-authorships. Every number you see — the expertise-fit, each rule’s compliance per paper, and the embedding-similarity baseline — was precomputed offline and shipped as static data. The solver runs in your browser; there is no model call at runtime.
Being exact about what this is:
- The objective is a reading, not ground truth. Jev’s expertise-fit is a calibrated judgment; we validated it independently (it identifies a true same-area expert better than the embedding does), but it is still a model’s reading.
- Rules are precomputed. The five English rules were scored against all papers offline, so you can toggle them live; you can’t type a brand-new rule here (that would need a model call).
- Exact, small. 16×72 solves instantly by min-cost flow; a real conference is larger and would run the same method server-side.
Two controls, and the plan re-solves on every change:
- Objective — switch between the embedding similarity (incumbent) and Jev’s calibrated expertise-fit.
- Rules — turn each reviewer’s plain-English rule on or off. A red chip is a paper assigned to a reviewer against their rule.
Watch the tiles: “English rules broken” goes to zero when the rules are on (only the calibrated setup can honor them), while “papers with a domain expert” and “mean expertise fit” hold or improve. That is the whole thesis — you can optimize a quantity, and honor a constraint, that used to exist only as words.
Most real allocation problems have this shape: a clean combinatorial core, wrapped in objectives and constraints that live in text — a customer’s note, a proposal, a policy, a preference. Classical optimization stops at the numbers; the interesting value is in the words it throws away.
BIS builds the version that reads them: a calibrated model turns the prose into objective coefficients and constraints, and a real solver optimizes it — against your data, your rules, behind an API your systems call.
This preview runs on real public data. Yours would run on your data.
We start with a conversation, prove the value on a focused POC with your real data, then ship it — in our environment or deployed into yours.