SECOM yield analysis: sensor screening and forward prediction
On this data a predictive lot screen is not worth deploying. Recall falls from 0.548 on a stratified split to 0.000 on a time-ordered split, catching none of 26 failures.
Ethical AI is often discussed as principle and rarely delivered as something a person can actually use. Our work in this program is applied: tools and methods for anyone who needs to judge whether a claim, a citation, or a generated answer can be trusted.
That audience is deliberately broad. Students, journalists, clinicians, librarians, and researchers face the same underlying problem, and it does not take a technical background for it to matter to them.
Active project
A verification tool for citations and sourced claims, built for general use rather than only for academics. AI systems produce references that look correct and are not. VeritasAudit exists to make checking them fast enough that people actually do it.
Read the July 2026 newsletterNine engineering and data studies, each published as a single self-contained page. Every number on these pages is read from the analysis output at build time rather than typed into the text.
Most of these studies return a negative or qualified result. We publish those in full, because the qualification is usually the finding.
On this data a predictive lot screen is not worth deploying. Recall falls from 0.548 on a stratified split to 0.000 on a time-ordered split, catching none of 26 failures.
Two cases, one gauge. The passing case resolves 19 distinct categories; the failing case resolves 2, and reproducibility is what breaks it.
The verdict is NOT DEMONSTRATED rather than FAIL. Reliability at 3000 hours estimates at 0.9165 and clears the criterion, but the 90% lower bound of 0.7771 does not.
The tray does not pass. It fails two of three load cases and the mesh-convergence gate, and swells past the dimensional spec under all three coatings tested.
A single plug-flow reactor at 133.808 litres wins on cost per mole, not on yield: a two-reactor network makes 8.84% more B but costs more per mole. The optimum sits on the lower temperature bound of the operating window, so the constraint sets it rather than the chemistry.
Reproduces the 1983 published targets exactly, then shows why the choice barely matters. Total cost moves at most 0.80% across any 10 degree window, so the optimum is flat. Drag the slider and watch it stay flat.
Levelised cost lands at 11.34 USD/kg against a 10 USD/kg sale price, so the base case is uneconomic. Capital cost dominates the uncertainty, which overturns the prediction we recorded before running it.
A two-height BDST line predicts every one of 36 held-out 12.5 cm service times within 20 percent, with a 7.9 percent median error. The source omits every 12.5 cm observation at the slowest flow, so 18 of 54 strata cannot be checked.
The frozen Stichlmair correlation predicts flooding at 4.77 m/s against a first observed flood at 6.89 m/s, 30.7 percent early for this one apparatus. Being conservative here is not a general safety margin.
The study reports above are fixed documents: they state what was found and the reader follows the argument. A tool is a different kind of teaching. It lets someone change a condition and watch the answer move, which is how the shape of a result becomes intuition rather than a figure to be taken on trust.
These run in the browser, are free, and need no account.
Bubble and dew points, activity coefficients, flash calculations and phase diagrams for 44 components. Choose a mixture, pick an activity model, and plot the Txy or Pxy diagram. Every number is computed from a physics core and traceable to its model, parameters and literature source.
It is built for a first course in separations, where the ethanol and water azeotrope is easier to believe once you have watched the curve reach it yourself.
Open the VLE explorer
Research at the undergraduate and graduate level. Agent systems are increasingly used to plan syntheses, screen candidates, and read literature at scale, which makes their failure modes a chemistry problem as much as a computing one.
Chemistry
Work emphasizing problems where a result changes practice rather than only the literature.
Chemical engineering
Scaling what works at the bench into something that survives real production constraints.
AI agents
How agent systems coordinate, where they fail, and what verification they need before a lab trusts their output.
Students and collaborators: we take on research partners.
Work with us