The Lab
The Lab
This is where our operations research gets done, and it runs on the rule a good referee would impose: define the problem exactly, state the closest prior art, name the contribution narrowly, and fix the test before the result is opened. Most candidates do not survive that rule. The ones that do become products. The ones that do not are published here too, because the next person to have the idea deserves to know what happened to it.
- 01
1 · Research
A precise problem, a stated novelty threat, and a test fixed before the result is opened. Negative results are kept and published alongside the rest.
- 02
2 · Proof of concept
An exact, internally-reproduced implementation that shows the method survives real structure, with external replay as the bar we are building toward.
- 03
3 · Product
When it repeats across buyers and stays reusable, it graduates into a product with an operable path.
Papers & reports
Papers and reports
Each paper states what is proved, what is a negative result, and where the contribution stops. Novelty is claimed narrowly and defended against the closest prior art, and the boundary gets the same care as the result.
The research frontier ↝When recovering outcomes one at a time is worth the wait
Recoverability-aware adaptive de-censoring: a positive backtested signal that adaptivity pays, an exact criterion for when it reverses the batch decision, and a simple operator-runnable policy that recovers almost all of the value.
Which missing outcomes should you pay to recover?
Decision-focused de-censoring: selecting which already-censored historical outcomes to recover when their only purpose is to improve a later, globally constrained decision.
When a screening certificate costs more than the solve it saves
Decision-stability screening certificates: a negative result on why decision-aware pruning must be cheaper than the exact work it avoids, not merely tighter and safe.
Recovering outcomes when you do not trust your recovery model
Distributionally robust de-censoring. Three early structural findings on choosing what to recover when the recovery model itself may be wrong.
Learning what to recover while you spend on recovering it
Online two-ledger de-censoring. An online policy over two coupled, non-fungible budgets (evidence capacity and intervention capacity) under delayed, censored feedback, and the restricted-setting theorem it still needs.
The higher-order-function architecture of operations research
A structural thesis: that optimisation, sequential decision-making under uncertainty and validation are compositions of a small, shared set of higher-order-function patterns.
Proofs of concept
Proofs of concept
Working code, written to test one claim before anybody was billed for it. Some of it is open source, and some of it is now a product.
Rainbow: a theorem prover that treats proofs as data
A Clojure-native dependent-type prover with a tiny trusted kernel, built to prove operations-research algorithms rather than only textbook mathematics.
Clojure · Dependent types · Content-addressed proof objects
An exact solver for decision-focused de-censoring
An exact solver that decides which censored outcomes to recover before a constrained portfolio is allocated, internally reproduced, with certificates for when the answer is to recover nothing.
Clojure · Exact rational arithmetic · Multiobjective knapsack DP
orhof, operations research as higher-order functions
A dependency-free Clojure library that reconstructs a large body of operations-research algorithms as compositions of a small set of higher-order-function patterns.
Clojure 1.12 · Zero dependencies
chachaml, REPL-first MLOps for Clojure
A Clojure-native experiment tracker and model registry, the Clojure answer to MLflow, where runs, artefacts and pipelines are managed from the REPL and queryable by an agent over MCP.
Clojure · SQLite / Postgres · MCP server
clj-bdh: a Clojure port of the Baby Dragon Hatchling model
A Clojure port of Pathway's brain-inspired "Baby Dragon Hatchling" language-model architecture, where the model is a plain function over a map of tensors and every run is tracked end-to-end.
Clojure · libpython-clj2 · PyTorch
Case studies
Case studies
FAQ
Questions about the Lab
What is the Lab?
Flexiana's operations-research lab. The rule it runs on is the one a good referee would impose: define the problem exactly, state the closest prior art, name the contribution narrowly, and fix the test before the result is opened. Most candidates do not survive that. The ones that do become products, and the ones that do not are published anyway.
How does research become a product?
In three stages. Research: a precise problem, a stated novelty threat and a frozen falsifier. Proof of concept: an exact, internally-reproduced implementation that shows the method survives real structure, with external replay as the bar we are building toward. Product: when it repeats across buyers and stays reusable, it graduates into a product with an operable path.
What does "negative results" mean here?
A candidate that fails its frozen test is kept and published. The falsifier is fixed before the result is opened, so a method that does not survive is a finding, and most do not survive. Each paper states what is proved, what is a negative result, and where the contribution stops.
What is a proof of concept in the Lab?
A working implementation built to prove one specific claim before anyone paid for it — an exact, internally-reproduced build that shows the method survives real structure. Some are open source and some graduate into products; external replay is the bar we are building toward.
Do you publish the method, or only the result?
It depends on the work, and each item says which. Where the work is open source, academic, pro bono or deliberately public, the method is published and you can read it. Where the mechanism is the commercial asset, we publish the result, the evidence grade and the boundary, and say plainly that the mechanism is withheld, instead of writing around the gap and hoping you do not notice. Each paper states what is proved, what is a negative result, and where the contribution stops. Novelty is claimed narrowly and defended against the closest prior art, so the boundary of the method is published alongside the method itself.
Is my problem a research problem?
Possibly, if it is expensive, recurring, and badly served by the tools that exist. Most decisions are not research problems and are better served by an audit. Bring it to us and the first conversation settles which it is, and whether we are the team for it.
Start with the decision
If your problem might be a research problem
If your decision is expensive, recurring and badly served by off-the-shelf tools, it may be a research problem. The first conversation settles whether it is, and whether we are the team for it.