Tools

Every actuarial team reconciles, and the afternoon goes on hunting for the reason. Drop in two files and it pairs the columns, joins on the key and splits the difference by whatever you group on. Then twelve detectors look for the ordinary causes (leading zeros, amounts in pence, a missing filter, duplicate rows, gross against net), each applied to your rows and measured before you see it. A model can propose more under the same test, so on the worked example what you read is “multiplying these 57 annuity incomes by 12 explains £136,727 of the difference, and here they are”. Accept the ones you believe and it draws the bridge, with whatever is left standing in its own colour. Nothing is uploaded: a model sees a profile of your columns, never your rows.

It proves that calculation on examples it was never shown, and writes up every assumption it made. Everything that computes runs in your browser, the Python in Pyodide and the spreadsheet in HyperFormula; the only thing that leaves the machine is the prompt.

Backsolve works out a calculation that already happened; this builds one that has not been written yet. Two models each write the specification; you settle where they differ, or take the best guesses. Tests are written from it before any code exists, in four kinds only, and the code, written without sight of them, runs against them in your browser, then over every policy in your file. An Excel model built from the specification with no AI is checked against the Python policy by policy. At Accept you try the finished calculation on figures of your own, worked out again as you type by the code and by the Excel model, see a map of how its parts link with each figure in its box, and accept or reject acceptance tests made from the specification, adding cases of your own. You leave with the specification, code, tests, workbook, the signed-off acceptance tests, the map and a summary structured to TAS 100.

A bulk calculation in Excel is an array program in disguise: a formula filled down is one calculation over every policy, a formula filled right is a step through time. Accelerator reads the workbook in the browser, finds the policies (a sheet that works one policy at a time, the policy picked by a cell, or a row per policy), finds the lines through time and the answer, and carries every formula into Python by fixed rules, with Excel’s own rounding, INT, MOD and lookups. The Python runs here on every policy under Pyodide, and the same translation runs as a WebGPU kernel on the visitor’s own GPU, each checked against Excel’s answer policy by policy. The same run takes about a minute in Python alone, against the quarter of a second above on the GPU. What-if tables are refused too, naming the cell. The download is the Python with its tests, the inputs as CSV, a reconciliation workbook, the mapping from each Excel range to its line of code, an audit of the workbook, and a summary set out as TAS 100 asks.

A harness the agent cannot touch fits each model, scores it against a cost function you can edit, and keeps a change only if the score fell. Each problem has one factor carrying no signal, to show whether the agent buys parameters that earn nothing, and one it is forbidden to use, which the harness refuses. The holdout is scored on every fit but shown only at the end, the one number nobody optimised against. A control with no AI, greedy forward selection, ties with the agent on mortality: the value is the reasoning and the write-up, not the last decimal of deviance.

A first look at a new table usually means a week of one-way tables, a GLM and a chain of emails. Here the columns decide what runs: a yes-or-no event with exposure gets rates, actual against expected, a Poisson GLM, a trend and survival curves; an amount gets distribution fits and large values; a triangle gets the chain ladder with Mack. Nothing is built for one kind of data, so mortality, lapses, claims and take-up all go through the same fourteen techniques. The Python that made every figure goes into an audit folder with a log of every choice, and a script there reruns the analysis and checks every figure matches.

The usual voice sales agent is a prompt describing the product and the rules, and hope. This inverts it. Each turn the model is handed one sentence and returns a filled-in form (an intent, an objection, a vulnerability flag, some numbers) and everything else is code: it never chooses the next node, never sees the quote, and is not told whether the sale is going well. Sixteen obligations are tied to the nodes whose words discharge them, and a fuzzer has run two thousand calls without reaching a sale with one outstanding. Any sign that the caller is not in a position to decide stops the sale. The quote is a real annuity calculation: a Makeham survival curve, escalation, a guarantee period, a joint life and health priced as an age rating. On the cheapest model the AI hears and answers a whole call for about a penny, and anyone can name the firm, reword the questions or add their own before they ring.

Load a Word document, say what it is for and who reads it, and click the standards to hold it to: TAS 100 and TAS 200, the Actuaries’ Code and the APSs, the PRA’s Solvency UK reports, the FCA’s Consumer Duty, product governance and pricing rules, Lloyd’s, IFRS 17 and good practice, eighteen packs checked against their sources, or your own, typed, loaded from a file or drafted by the AI. A requirement several standards share is one check. The AI gives every check a verdict with the words that show it, and each quote is looked for in the document; a Gap Score counts what is missing by fixed points, and a verdict you overrule changes it at no cost. The AI then writes changes for the gaps, leaving any fact only the author has in square brackets, and they are shown as Word shows tracked changes, to accept, reject or reword. The download is your own file with the accepted changes tracked and each gap a comment, a clean copy, the checklist as a workbook and the review as a page.

Drop the files of a monthly process into one zone per month: a claims extract, the assumptions, the triangles, the reserving model and the board summary, say, for two or more month ends. Each file is paired with its partners in the other months by its name with the dates taken out and by its sheets, and the process is drawn as a map from the files themselves: a formula reading another workbook proves a link, and says which month’s file it reached, so a link left pointing at last month’s file shows at once. Buttons over the map pick out one kind of problem at a time, and Trace Back lights everything that feeds a figure. Each month is compared with the one before, cell by cell, with formulas read relative to their own cells so a range grown to take in new rows is not a change. Forty checks, each of which can be switched off, given a severity or threshold and reworded, look for stale and broken links, typed-over formulas, totals that stop short, checks no longer nil, history restated, duplicates, dropped records and assumptions changed with nothing to say why, and you can add checks of your own in words. A bridge recalculates the files together, swapping one month’s file for the next at a time, so every step of the headline’s movement is a file and what the files do not explain is its own step. The download is an Excel workpaper with the bridge in live formulas, a Word review and the map.

Each page opens with a short animation, then hands you the same numbers as charts you can drag: move the retention and watch the cession change, or watch a one-way rating table overstate a young-driver loading. Explainers from present value, mortality laws and GLMs to profit testing, IFRS 17, Solvency UK and the ORSA, plus The Placement, a game of client orders from a single layer up to retrocession. No model is called: every figure is computed in the browser, so the animation and the charts never disagree.

The branches run from learning from data and linear models (ridge as credibility, the lasso) through classification, smooth curves and choosing a model, to additive models, trees, and inference and averaging. Every level is dressed in a pricing decision, such as a lapse model, an age curve an underwriter will sign or a validator counting parameters, and scored against unseen policies. Every number is computed in the browser from the same book the levels use, so a figure and a level never disagree. No model is called and nothing is uploaded.

The first act builds a network from sums, weights and bends, through the downhill step and the chain rule to overfitting and a network that reads digits. The second is a toy language: a table of counts, softmax, surprise in bits and temperature. The third is attention, up to a one-layer transformer, and the fourth is serving a model: rounding its weights, the memory each conversation takes, and a minute of traffic priced per token. Beside them run three optional tracks: logic, from single gates to an adder and Shannon’s tie between circuits and algebra; robots that the network drives; and seventeen maths workshops, each saying where a language model uses its maths. Every figure comes from the game’s own engine, which works out each slope by backpropagation. No model is called and nothing is uploaded.

A deterministic engine keeps the books, the story reacts to what you choose, and wrong calls plant problems that surface later. A model writes a new company, cast and story one year at a time, or play one of two hand-written campaigns with no network calls at all.

Thirty feeds are read three times a day by a scheduled job and committed as one list, so a visit fetches nothing and the page is instant: the trade press, the FCA, the PRA and the Bank of England where they touch insurance or pensions, and searches by subject, from bulk annuities and the CMI to Ogden and AI in claims. The list keeps no more than eight hundred stories, usually about a week, and says the dates it covers. The same story from four outlets becomes one story with four tellers, and the most-told stories sit at the top. Ask a question and a model reads the newest four hundred items on the tab you are on and answers with the items it drew on, each checked against the list. Copy any view as a Markdown briefing for the Monday note.