One answer, every time

Ask an agent the same money question twice and you can get two different numbers. Ask traj, and you get the same one every time. Here is why, and how to check it yourself.

Three cream robots with orange eyes wave from sand dunes under an orange sun, beside the words "One answer, every time. Deterministic by design."

An AI agent is good at the parts of a financial plan that need judgement. It asks what you earn, notices what you forgot, and explains a result in your own words. It is less good at being the calculator. A language model writes its answer by sampling, so the same agent, asked the same money question twice, can give you two different numbers.

Three identical robots on sand dunes each hold up a card with a different bar chart, and glance at each other's cards.
The same agent, the same question, three answers.

That is fine for a paragraph of advice. It is not fine for a retirement plan you mean to come back to next year. If the number moves, you need to know whether your life changed or the model did.

The agent judges, traj computes

traj splits the work in two. Your agent supplies the judgement: it asks you for the facts, decides what to run, and reads the result back to you. The arithmetic happens in traj, a command line that never calls a model. So every number your agent quotes is one traj returned, and traj returns the same one each time you ask.

Three identical robots stand in a row on sand dunes, each holding the same bar chart, joined by a thin orange line.
The same agents with traj: one question, one answer.

What keeps the answer the same

Not calling a model is the start. A program can still drift in quieter ways, so traj closes each of them:

  • No floating point for money. Amounts are whole numbers of the smallest unit, satang for baht. 1.005 baht becomes 101 satang, not the 100 that floating-point rounding gives. As the money package puts it, float noise never decides.
  • Exact integer growth. Interest compounds in whole-number arithmetic. When we compared two JavaScript engines, their built-in Math.pow, exp and log disagreed on 228 of 10,296 results, so the projection engine uses none of them.
  • No clock. Every date a command acts on is one you pass. “Today” is a date you give, so a command you run tomorrow answers the same question it answered today (ADR 0018, ADR 0029).
  • One output format. Your agent asks for JSON on every call, and the format is never guessed from the terminal, so a pipe and a screen cannot disagree (ADR 0010).

Check it yourself

Here is a 30-year projection. A salary of 40,000 THB a month grows with 3% inflation, and all of it is saved into one cash account earning 4.5% a year, from 31 March 2026. No tax is applied, so the figures are before tax.

Save this as base.json. Amounts are in satang, and rates are in basis points:

{
  "schemaVersion": 4,
  "currency": "THB",
  "economy": { "inflation": 300 },
  "accounts": [{ "id": "cash", "kind": "asset", "rate": 450 }],
  "flows": [
    {
      "id": "salary",
      "kind": "income",
      "earned": true,
      "amount": 4000000,
      "frequency": "monthly",
      "start": "2026-01-01",
      "growth": "inflation"
    }
  ],
  "cashflow": {
    "surplus": [{ "account": "cash" }],
    "contribute": [],
    "withdraw": []
  }
}

Then run it in a throwaway ledger, so your own data is not touched. This needs traj 0.4.0 and jq. On macOS, use shasum -a 256 in place of sha256sum:

export TRAJ_HOME="$(mktemp -d)"
traj init
traj scenario apply base --file base.json
traj projection run --scenario base \
  --as-of 2026-03-31 --horizon 2056-03-30 --json \
  | jq -c .tables | sha256sum

We ran it five times with traj 0.4.0, and again in a second fresh ledger. Every run printed the same hash:

dd9bc2f0dd33caed788ad28b92742748ed95c3f683e332af577a6f466fb807ad

Every run also gave the same closing net worth on 30 March 2056: 43,504,723.86 THB, or 17,924,821.75 THB in real terms after the plan’s 3% inflation. Run it on your machine and you should see the same hash, down to the last satang. If you don’t, that is a bug, and we want to hear about it.

A robot drops blocks into a small cream machine, which prints a card with a bar chart; a row of identical cards lies on the sand beside it.
Same inputs in, the same card out, every time.

What deterministic does not mean

Deterministic is not the same as right. traj computes exactly what your plan says. If an assumption is wrong, such as a 4.5% return you will not get, the answer will be wrong in exactly the same way every time. That is the reason to keep the judgement with you and your agent, and the arithmetic in a tool whose work you can check.

traj works in Thai baht today. It is free and open source, and it keeps your ledger on your own machine.

Try it

Install traj and give your agent the setup skill. The commands are on the installation page. Then ask your agent a question about your money twice, and compare the answers.

Found a wrong fact or a number you cannot reproduce? Tell us, and we will correct the post.