Actua is a workbench for reserving, valuation and exam work. The maths runs in a deterministic Python engine with a test pinning every figure. The language model explains results. It never produces them.
In private beta. Accounts are opened by hand, so there is no checkout. Email us and we will set yours up.
Not because anyone chose it. Because the triangle has to be built before the method can run, and the tooling stops at the triangle. The Casualty Actuarial Society asked its members how the work actually gets done.
Source: CAS E-Forum, “Best Practices for Property and Casualty Actuarial Reserving Departments”.
The same worked example the engine’s own tests use. Eight transactions in, two of them unusable, and every figure below reconciles to the one before it.
Two rows cannot be used. They are named, with the reason, and excluded from everything downstream.
| AY | 12 | 24 | 36 |
|---|---|---|---|
| 2022 | 1,000 | 1,500 | 1,800 |
| 2023 | 1,200 | 1,800 | · |
| 2024 | 1,400 | · | · |
Built at the granularity you choose. Nothing is dropped quietly, so the totals here account for every accepted row.
Weighted average link ratios, and a refusal instead of a factor wherever the denominator is zero.
| AY | ULT | IBNR |
|---|---|---|
| 2022 | 1,800 | 0 |
| 2023 | 2,160 | 360 |
| 2024 | 2,520 | 1,120 |
With the method, the assumptions and the shape of the data recorded. Enough to reproduce, never a copy of the claims.
The failure mode of an AI in this field is not a crash. It is a confident, plausible, wrong number with no error anywhere. Actua is built so that cannot happen quietly: every figure comes from one Python implementation, bounded and tested, and anything it cannot compute is refused out loud.
The maths lives in exactly one place. A check in the build fails if arithmetic appears anywhere else, because it once did, and the two copies disagreed.
Every dimension, rate and magnitude has a ceiling. Unreadable cells, infinities and misspelled fields are errors, never quietly-defaulted values.
Reserve a book and four segments fail, the total says it covers the other fifty-six. It is never presented as the portfolio figure.
The assistant is one surface. The rest is the software you close a quarter in. Every figure below came from a real run.
Reserve a whole book in one run, each segment with its own method. When segments fail, the total says which ones it covers rather than quietly covering less.
Claim transactions in, triangle out. Rows that cannot be used come back by row number with the reason, and the spreadsheet never reaches our servers.
Draft cannot jump to signed; it passes through review, and signed is final. The record keeps the method, the assumptions and the shape of the data, never a copy of the claims.
Answers are computed at grade time, so a wrong answer key is impossible. Distractors are built from real misconceptions, so picking one tells you which mistake you made.
Also in the workbench: Ask, for anything the engine can compute; Reserving, for one triangle at a time with method comparison and tail sensitivity; Tasks, for starting from the job rather than the tool. Notebook is on the roadmap and is labelled that way in the product, because a feature that is not built should not look locked.
Every solution shows the derivation, and the number underneath it was executed rather than recalled, so a plausible-looking wrong answer cannot reach you. Distractors are built from real misconceptions, so choosing one tells you which mistake you made.
SOA P · FM · FAM · ALTAM · ASTAM · SRM · PA
IFoA CM1 · CM2 · CS1 · CS2
Sign-off cannot be skipped. Draft must pass through review, and signed is terminal. The record keeps the method, the assumptions and the triangle’s shape, which is what a reviewer needs and is not a second copy of your client’s claims data.
You remain the responsible actuary. Actua does not
reduce your obligations under ASOP 56 or CPS 320.
Uploads are parsed in your browser and released when you close the tab. There is no copy on our servers to leak, subpoena or lose, which is the only version of this promise worth making.
Accounts and the audit trail sit on servers we run ourselves in Oracle ap-sydney-1. No managed database provider holds a copy.
Knowledge-base search embeds on our own machine, so the text of your question is not sent to an external service.
Writing an answer uses a language model hosted in the United States. Deterministic calculations do not go through it.
The full detail is in the privacy policy, including what we do not cover.
Actua is in private beta. Rather than take a payment for something still being finished, we set accounts up by hand, which also means we can put you on the right tier instead of guessing.
Teams, consultancies and anyone who needs a different shape: more seats, your own data connectors, or a deployment arrangement. Quoted individually, so tell us what the work looks like.
Talk to us about CustomNo. There is no checkout while we are in beta. Email sales@evolvion.com.au and we will open the account, usually the same day.
No. The arithmetic runs in a Python engine with bounded inputs and a regression suite. The model reads results and explains them; a build check fails if calculation logic appears anywhere else.
It refuses. Anything the engine cannot compute returns an error with the reason, and the model is instructed not to substitute its own answer.
You can use it in work you sign, and you remain the responsible actuary. It records enough to reproduce a calculation, which is what a reviewer needs.
Spreadsheets stay in your browser. Accounts and the audit trail are on our own servers in Sydney. Chat questions reach a US-hosted language model; calculations do not.
Student is the discount: $30 a month with exam practice, grading and progress. Email sales@evolvion.com.au from your university address.
Send us a line about the work you do and we will open an account on the tier that fits it.