Hard Data¶

The readout desk. Scelo runs every switched-on model and lays the results out on a canvas, with a board pack you can print and a bridge to the swarm.

The Hard Data workstation: result nodes on the canvas with a board-pack hub. The animated illustration needs JavaScript.

The results canvas¶

When you arrive, Scelo runs the models in wire order — a model runs after the models wired into it and receives their results — and shows:

Result nodes — one per model, coloured by family. Each shows a headline number (e.g. Survival @ horizon · 0.514), a sparkline or a small table, and a confidence-interval strip when the model carries uncertainty. Tables and long content scroll inside the node (with a soft fade at the clipped edge) so nothing is cut off. A badge says where the figures came from: R or python (Scelo IDE's bundled runtimes) or in-browser. If a bundled run fails, the card says bridge failed with the reason and shows the in-browser figure instead.
Not applicable cards, for models whose inputs aren't in your data. The card stays neutral grey and prints the reason — Needs a mortality table … — and it's counted apart from failures in the run stats, on the macro card and in the board pack. A run that actually broke says failed, in red, with its error.
The wires from Tools, drawn between the result cards as dashed brackets labelled with what flowed (projected mortality, fitted GBM, …).
A Board Pack hub node that aggregates the run.

Click a result node's body to focus it in the side panel; click the ⤢ icon to open its detail dashboard.

What the models fit¶

These models are fitted to your data; where the inputs aren't there, the card says not applicable rather than showing a stand-in figure.

Model	What Hard computes	Not applicable when
GBM (LightGBM)	Gradient-boosted trees fitted in the app (Newton boosting on binned features — the LightGBM recipe, not the LightGBM library), with a squared-error, logistic or Poisson loss to suit the target. A fifth of the rows are held out and the headline is scored on them: AUC for a 0/1 target, R² for an amount, deviance explained for a count. The chart is a lift chart on the holdout — the average actual outcome in each band of predictions, lowest to highest.	No numeric column to predict, fewer than 50 rows with a usable target, or no feature columns
SHAP explainability	Exact TreeSHAP of the wired GBM's own trees on up to 500 of its held-out rows: each feature's share of the mean absolute SHAP value, and whether higher values raise (↑) or lower (↓) the prediction	No switched-on GBM is wired into its model to explain pin, or that GBM didn't fit
Lee-Carter	log q = α(x) + β(x)·κ(t) fitted to the dataset's mortality table, κ projected ten years as a random walk with drift, with a 95% band. The headline is q at 65 (or the nearest age) in the last projected year.	No mortality table, or fewer than 3 years × 2 ages with a rate in every year
Cairns-Blake-Dowd	logit q = κ₁(t) + κ₂(t)(x − x̄) fitted year by year over ages 50 and up (every complete age, if fewer than three of those are), both κ projected ten years on their drifts	As Lee-Carter
Life Contingencies	Annuity-due, term assurance and pure endowment for a life aged 65 (or the nearest age) over up to ten years at 4% — priced on the cohort of a wired Lee-Carter or CBD projection or, with nothing wired, on the latest year of the dataset's life table	No mortality table in the data

A mortality table means an age column with qx, mx, or deaths + exposure — and a year column, to project. The GBM's target is chosen from the data: a claim amount, a claim count, a 0/1 outcome, a column named target, label, y, response or outcome, otherwise the last numeric column — and the card says which, and why. For a claims target, the other claims outcomes (paid, incurred, …) are kept out of the features so they can't leak the answer.

In Scelo IDE the bundled runtimes compute two of these on the same basis: Lee-Carter in Python (numpy and statsmodels) and Life Contingencies in R, with the lifecontingencies package. Elsewhere on the canvas:

GLM · severity — with GLM · frequency wired in, the card adds the pure premium (frequency × severity). Its in-browser headline is the claim-weighted mean severity.
Bornhuetter-Ferguson takes the book-average ultimate as its prior, not chain ladder's.
In Scelo IDE, when CLIMADA can't run on real hazard data, the bundled-Python fallback is labelled synthetic loss model, not CLIMADA.

The model detail dashboard¶

The ⤢ on a result node opens a full-screen detail view:

Theory · assumptions · formulae — rendered with proper math (KaTeX).
Run output and diagnostics — model-specific charts/tables (ATA factors and CDF for chain-ladder, p5/p95 ranges for bootstrap, etc.). Null/empty fields are hidden; objects and arrays are summarised, not dumped as raw JSON.
A scoped chat about that specific model's result.

The board pack¶

The Board Pack hub node has a ⤢ that opens the printable report:

An executive summary in plain English, written for the signing actuary and the board. It leads with the figure that matters, says how far the methods agree, gives the uncertainty as a range in words, and ends with what the figures rest on and what to check before relying on them — models that couldn't be applied, approximations not fit for sign-off, synthetic data. It uses no software names or statistical shorthand. Scelo writes it from the run results alone; with an AI provider connected, the model rewords that draft for flow but must keep its figures.
Estimates (a forest plot or a metrics list), a trajectory overlay, and a per-model breakdown.
Every attached model that produced no figure, listed with its reason under not applicable to this data or failed — so the count of model runs is never mistaken for the number of models attached.
A download pdf button (uses the system print dialog).

You can also open it from the toolbar: report · pdf.

Convening the swarm¶

From the result side panel you can send a result to the multi-agent swarm — the focused result or, with nothing focused, the dominant run:

Convene council — choose the number of agents (12 → 192) and whether to include the society pulse, then run.
The council deliberates (this uses the swarm server and the local LLM, so it takes time — a few seconds per agent).
When it finishes, a synthesis card shows how much of the council trusts, distrusts or is uncertain about the result. A model result isn't a community the swarm's forecast can simulate, so the council judges it as stated; distrust means it found a flaw, uncertain that the result alone isn't enough to judge.
Click Open in swarm to jump into the full swarm view for that run.

The swarm is its own server

Council and simulation features talk to the swarm server, which Scelo IDE bundles and starts with the app on a loopback port (3010 by default — see Running the swarm). If a council reports "swarm server unreachable", open the swarm view for the server's status, its log and a restart swarm server button. A large 192-agent council on a local model can take many minutes — a smaller subset (12–48) completes much faster.

Action	What it does
rerun & regenerate	Re-run all models and regenerate the narrative
re-layout	Snap nodes back to the default circle
edit models	Back to Tools
export · code	Export the whole run as a script
report · pdf	Open the printable board pack