Skip to content

What's new

0.2.2

Released 2026-10-02.

Simulate from a scenario

  • It opens again. Clicking ▷ simulate blanked the whole window in every release since 0.1.2. It now opens the dialog.
  • You can see it work. While a run is under way, the dialog shows the same progress overlay as the rest of Scelo: a table materialising over a rail that fills with the agents that have answered ("30 / 120") and scans while references are resolved and results are scaled.
  • Pause, resume, stop. Pause holds the run where it stands: the calls in flight are cancelled and asked again on resume, and no new agent starts. Stop ends the run and changes nothing. Closing the dialog leaves a run going: it lands when it finishes, and reopening the dialog shows where it is. Leaving the pipeline (for the workspace or settings) or quitting Scelo stops it.
  • Augment matches only what your rows say. A row is matched to the reference cohort on the age, sex and comorbidity columns it actually has. Rows without them used to be treated as a 35-year-old woman with no comorbidities, and all handed one agent's answers; they now get the whole cohort's figures, labelled cohort in sim_bucket_match. The dialog says so before you run it and suggests generating a dataset instead.
  • You land on what it added. When a run finishes, the grid scrolls to its first sim_* column. Running augment again replaces the earlier sim_* values instead of stacking another "+ sim", and a run in which every agent failed says so and changes nothing. See Simulating and exporting.

Tools

  • Switching a model on wires it. A model that arrives switched off, because the data can't feed it, and is switched on later (by its toggle, a new pick or the chat) is now plugged into the models waiting on it, so SHAP beside a GBM no longer says it can't run. Identify models and regenerate wire every model they pick, and sessions and .sce projects saved with a model left unplugged this way are repaired when they open. See Pins and wires.

Council

  • The council card says what comes back. On a Hard Data result, the council card and the deliberation overlay no longer promise a proposed parameter shift: since 0.2.1 a result is judged as stated, so trust, distrust or uncertain come back, with the reasons in the swarm.

0.2.1

Released 2026-10-01.

The swarm judges what it is given

  • No more blanket distrust. The swarm's forecast models a community: people, and the money, time and relationships that carry them. A scenario that isn't one (a fund's allocation, an insurer's capital decision, or a model result sent from Hard Data) used to get the same generic forecast, and the council distrusted it every time. The council now judges such a scenario as stated: trust means it holds up on what is given, distrust that something it states is wrong, and uncertain that there isn't enough to judge. Community scenarios still get the forecast, and the council still votes on it. See Council, society & simulation.

Opening on the pipeline

  • A launch intro. The S₀.₂ mark draws itself in, and the window opens in your theme's colour, with no white flash. A click or a key skips it. See First launch.
  • Straight to the pipeline. Scelo opens on Soft → Tools → Hard. The bar above it carries workspace and reset session, and get started, top left, leads to the welcome screen.
  • Level cards. The three pipeline cards share one height, so their chat boxes line up and the wires between them run straight. The chat's send button is now an enter key.

One typeface

  • The whole interface uses SN Pro, with its own arrows: labels, headings, charts and the swarm alike. The code editor and terminal keep a fixed-width font, so columns line up.

0.2.0

Released 2026-09-30.

Tools: a typed canvas

  • Only the pins that matter. The dataset hub offers only the inputs your data really contains — a claims triangle, a mortality table, model points and so on — each with the evidence it found, and every model shows only the pins it uses: inputs on the left, outputs on the right. See Tools.
  • Every wire carries something. A wire joins only pins whose types agree, and each one carries a result the next model actually uses: a Lee-Carter or CBD projection priced by Life Contingencies, claim frequency into severity for the pure premium, the GBM that SHAP explains, rate scenarios into the SCR. The old decorative arrows are gone, and sessions and .sce projects saved by earlier versions are rewired when they open. See Pins and wires.
  • The canvas says what fits. A dragged wire says why a drop won't take; dropping it on empty canvas, or right-clicking, lists the models that fit; and the Tools chat can wire models too. A model your data can't feed says so on its node, and suggested models the data can't feed arrive switched off. See Connecting.

Hard Data: fitted models, honest results

  • Gradient boosting and SHAP are fitted. GBM fits boosted trees to your data and scores them on rows it never trained on; SHAP gives exact attributions for that same model. See What the models fit.
  • Mortality models are fitted. Lee-Carter and Cairns-Blake-Dowd fit your mortality table, and Life Contingencies prices a wired projection's cohort or the table itself. In Scelo IDE, the bundled R now prices Life Contingencies on Linux and macOS too, wherever the bundled R runs; before, that bridge failed on every run there. See What the models fit.
  • "Not applicable", not a stand-in number. A model whose inputs aren't in your data says so, with the reason, and is counted apart from real failures. When a bundled Python or R run fails, the card shows the real reason. See The results canvas.
  • Reserving methods line up. All four reserving methods now share one forest plot (in Scelo IDE it never formed before), with Mack's standard error and the bootstrap's range, and Bornhuetter-Ferguson uses an independent prior, the book-average ultimate. See What the models fit.
  • A board pack in plain English. The executive summary is written for the signing actuary — the figure that matters, how far the methods agree, the uncertainty in words, what to check before relying on it — and the pack lists every model it couldn't compute, with the reason. See The board pack.

A quieter interface

  • Visible panel toggles. Every side panel in Soft, Tools and Hard has a panel-toggle icon in its header corner, and a collapsed panel leaves a thin rail with its name. Toolbars and scrollbars are quieter, and views, dialogs and menus arrive with a brief, purposeful motion (none, if your system asks for reduced motion). See Side panels.
  • The swarm's pets are alive. They breathe, blink and follow your cursor, lean in when you hover and hop when chosen — and the swarm now follows the IDE's light or dark theme. See The swarm.

Installers

  • The whole Python stack ships. The Linux, Windows and macOS installers now carry the complete bundled Python stack — SciPy, scikit-learn, statsmodels, LightGBM, chainladder and climada — which earlier installers left out. See Bundled Python & R.
  • chainladder that fits. The bundled chainladder moves from 0.8.26, which imported under pandas 3 but could not fit a triangle, to 0.10.1; the RAA chain-ladder IBNR (52,135) now matches R's ChainLadder.
  • R that stands on its own. The bundled R's packages (ChainLadder, forecast, lifecontingencies, lintr, languageserver) now travel with all of their dependencies; earlier Linux installers only worked where those happened to be installed already. On Linux the bundled R runs on Ubuntu 24.04 and later. See Linux.
  • A real terminal. On Linux, macOS and Windows (ConPTY) the integrated terminal runs in a real pseudo-terminal. Earlier installers silently fell back to a plain pipe: commands ran, but without a prompt, colours, line editing or full-screen programs. See Terminal & runtimes.
  • A new mark. The app icon and wordmark are now the S₀.₂ mark.
  • Apple Silicon. 0.2.0 adds a macOS build: a .dmg for Apple Silicon Macs on macOS 14 or later, with the same bundled Python and R. It is unsigned and not notarised, so clear the download's quarantine flag once (one Terminal command) before opening it. See Windows & macOS.
  • Built and tested in CI. The macOS and Windows installers are built on GitHub's own macOS and Windows machines, which check the bundled Python and R and launch the app before anything is published; the Windows installer is also installed there first.