Paper
A control-system study open in Paper, mid-run: an outline rail beside the document, a notebook cell whose heading a co-author is selecting as they read, the prose setting up a PID controller for a 50 L stirred tank, and the code cell the agent has just written underneath it. The agent panel lists the four calls it made — thinking about the IMC approach, writing 204 lines to controller.ts, rebuilding to verify the settling time, reading the response plot — and a compute card in the corner reports 60 per cent of a Frankfurt GPU pair while the run is still going.
What Paper can do
Engineering knowledge dies in two places: a PDF nobody can re-run, and a script nobody can read. Paper collapses both into one artifact — prose, Modelica models, and live compute in a single document that still executes years later.
Notebook
Prose and computation, together
Write in blocks — text, LaTeX equations, Modelica models, parameter tables and plots. Every figure is the output of code that sits three lines above it, so a reviewer can question the number and re-derive it in the same breath.
A notebook of four kinds of block — prose, a parameter table, a three-line cell and the plot it produces. A reviewer drags the mass-flow parameter from 1.50 to 2.00 kg/s, the cell re-runs, and the figure redraws while the outlet value quoted in the prose above it moves with it, to 83.2 °C.
Compute runtime
A runtime dock for every run
Attach compute to the document, not to your laptop. The runtime dock shows solver status, step size, residuals and wall time as the run proceeds — and pins each result to the exact model revision and parameter set that produced it.
A runtime panel mid-solve: dassl on four cores in Frankfurt, simulated time marching to 120 seconds while the dock counts the run down, the step size varying and the residual falling through orders of magnitude to 8.4e-9. When the run lands, the model revision and parameter set it is pinned to come forward below a rule.
Reproducible
Reproducible by construction
Models, solver traces and datasets are addressable records inside the project, not files in a folder someone renamed. Reopen a paper eighteen months later, change one coefficient, and every dependent figure updates — or fails loudly, which is equally valuable.
A rail of dated revisions grouped by year, beside the figure they produce. Reopening the 2026 revision — one coefficient changed — replaces the plotted curve with the recomputed one; two dependent records come back updated and a third reports itself stale and needing a re-run.
Publish & cite
From working document to publishable record
Compile to journal-specific LaTeX templates without reformatting by hand. Export the executable version alongside the PDF so reviewers receive the evidence, not just the claim — and cite a run by its identifier, the way you would cite a dataset.
A working document beside the two-column record compiled from it, assembling block by block. The same figure is carried across, and the compiled page closes on the run identifier — printed at its foot and shipped alongside the PDF as its own artefact — so the run can be cited the way a dataset is.
Use cases
01
Design justification
The document that proves a sizing decision was correct — and can be re-run when the requirement changes.
A cooler-sizing section of an executable paper: the ambient requirement moves to 45 °C, and the sized cooler face, the outlet peak the paragraph quotes and the headroom left against the 85 °C limit all re-derive from it — 3.3 K of margin down to 0.4 K.
02
Verification reports
Model-vs-test correlation studies where the comparison plot regenerates against the latest measurement set.
A model-versus-test comparison where the model curve holds still and the measurements change: a newer set of 22 test points replaces the previous one, which stays on faintly, and the fit re-computes to an R² of 0.991 at 0.6 K RMSE.
03
Trade studies
A written argument over an Exploration result set, with the Pareto front embedded live rather than pasted in.
A written trade study over an Exploration result set. Twenty-four more candidates land in a cloud of 180, the Pareto front re-fits over them, and only then does the sentence's named design change — the knee moving from 12.8 kg at 11.0 K to 12.9 kg at 10.0 K.
04
Teaching material
Course notes students can perturb — change the damping ratio, watch the argument change with it.
A page of course notes with the damping ratio set inside the sentence it governs. Stepping it from 0.35 to 0.70 swaps the step response for the one that ratio produces, and the claim about overshoot re-counts from 31 % to 5 % with it.
