A vertical actuarial workbench. The reserving and mortality modules compute live in your browser - edit the triangle or the pricing inputs, press the module's compute button, and every figure, factor and error bar recomputes. The market module shows an offline-backtested forecast with its measured error; its 80% interval undercovered the actual path in the recorded backtest (34% coverage) and is not validated for prediction. All math reproduces from public data; every estimate ships with its uncertainty.
Cumulative paid losses, RAA reference dataset (Reinsurance Association of America, Historical Loss Development Study 1991, p.91; the dataset of Mack 1994), $000. Edit any observed cell and recompute: factors, projected development (orange), ultimates, IBNR and Mack standard errors all update in your browser. The same math, run in Python, matches the CAS chainladder-python package (Mack convention) at every displayed decimal.
SSA period life table, 2023 (2026 Trustees Report) - US general population, not insured lives. Net premiums only: no expense loading, no margin, no underwriting classes. Computed life expectancies reproduce SSA's published values (male e₀ 75.79, female e₀ 81.06). The published table ends at age 119 with qx 0.9266 (not 1.0), so a small survival tail beyond 119 is truncated; ages here are capped at 110. Choose inputs and press Compute premiums.
Survival from age 65 by sex (static reference view). Female annuity costs more: lower mortality, longer expected payment stream. Interest rate is an explicit assumption, not a forecast.
Kronos-mini (4.1M params, MIT code+weights) ran locally on CPU - $0, no API. These are recorded outputs of an offline rolling backtest on SPY daily candles (6 out-of-sample windows, 30-day horizons, 12 sampled paths each, in two separate unseeded runs of 3 windows); the model needs PyTorch and cannot run inside this page, so this module is evidence, not a live calculator. It is a market-price forecaster for the asset side only - not a validated economic scenario generator for regulatory capital.
Measured, not asserted: the 10-90% band captured the actual path 34% of days overall - per window 93%, 10%, 20% (run A) and 10%, 47%, 23% (run B), two separate unseeded executions joined for display (a prior 8-path run gave 28%). Three weaknesses at once: the model is overconfident, per-window results vary widely, and 12 sampled paths per window make each band a weak empirical estimate. Any capital, reserving or pricing use must recalibrate interval widths first. Sampling used no fixed seed; recorded paths are archived with the build.
Paste the source treaty text and a summary. Every dollar amount and percentage in the summary must appear in the source, and the checker lists the treaty terms the source states that the summary leaves out. The preloaded source is a verbatim 1,614-character excerpt of the Travelers Companies 10-K for fiscal year 2025, "Corporate Catastrophe Excess-of-Loss Reinsurance Treaty" (treaty period January 1 to December 31, 2026): SEC EDGAR, accession 0000086312-26-000065, characters 212,200 to 213,814 of the page text with tags removed and whitespace collapsed. The excerpt ends mid-word where the window ends. Example 1 is the verbatim output of a small local model (Qwen3 1.7B) from Nov 2 on a window of this text. Examples 2 and 3 are fixtures we wrote to test the checker, not results from any customer or model. Nothing here generates a summary, and it is not actuarial or financial advice.
What this does not do. The original goal, an AI actuary combining MiroFish (multi-agent simulation) and Kronos (price forecasting), is not finished. No usable model extraction runs here, and the Kronos tests in the market section did not beat a no-change baseline. This checker only verifies a summary that someone else produced, on five claim types, with a bounded set of phrasings.