TradingCalc
DETERMINISTIC AI·TRADING MATH·LLM MATH ERRORS

Deterministic AI for trading math

Deterministic AI is a design in which a language model decides what to do and explains the result, while every number comes from a tool that gives the same output for the same input, from a formula anyone can read. In trading, the agent calls a calculator for a position size, an option Greek or a VaR figure instead of writing the figure itself.

97%
of 630 model answers within 1% of the exact value (October 2026)
12 of 17
wrong answers stated with a confidence of 80% or more
46
canonical vectors a result can be checked against
01WHAT IT MEANS

A model on its own, and a model with a deterministic tool

Asked twiceMay answer differentlySame inputs, same output
MethodNot visible in the answerA published formula and test vectors
Checking itRedo the math yourselfVerify a signature, rerun a vector
When it cannot answerTends to answer anywayReturns a flag, for example identifiable: false on a flat implied-volatility solve

The model is still the one that reads the request, chooses the tool and fills in its inputs. Deterministic here describes the calculation, not the conversation.

02LLM MATH ERRORS, MEASURED

Mostly right, and the wrong answers sound the same

Every month we give current models fresh trading-math questions and score each answer against an exact reference. In October 2026, 6 models answered 630 questions and 97% were within 1% of the exact value. Statistics over a series of returns and the direction of inverse (coin-margined) contracts account for most of them. The task types with the most wrong answers:

VaR from 40 daily returns6 of 30 wrong, 5 confidently
Annualized Sharpe from 30 daily returns3 of 30 wrong, 3 confidently
Call delta at the implied volatility of a quoted price2 of 30 wrong, 1 confidently
Bybit inverse PnL with taker fees2 of 30 wrong

Read in full, 16 of the 17 wrong answers are a slip in hand arithmetic or a swapped side, the rest an unfinished numerical solve, and none came from an unclear question.

The size of the gap matters less than its shape: 12 of the 17 wrong answers came with a stated confidence of 80% or more, so nothing in the reply separates them from the right ones. Researchers at MIT CSAIL describe the same pattern from the training side: standard reinforcement learning rewards a lucky guess as much as a careful answer, and adding a term that penalizes the gap between stated confidence and accuracy cut calibration error by up to 90 percent on a 7-billion-parameter model.

Results by model and task, with every raw reply Why AI agents need exact calculations
03CHECKED, NOT TRUSTED

Five ways to check a number without taking our word

Published formulasEvery tool states the formula it computes, conventions included: inverse contracts, coin-settled options, pip sizes, fees.
Methodology
Canonical test vectors46 exact input and output pairs for the core calculators, rerun live on every visit to the verification page.
Live verification
Independent cross-checksOption pricing and risk statistics are compared with QuantLib and Python numerical libraries, not with a second copy of our own code.
How it is checked
Signed resultsEach MCP tool result carries a signature over its exact text, verifiable offline with a published key.
Signed responses
A daily public recordOnce a day the verification state is signed and its hash written to Solana, so history cannot be rewritten quietly.
Daily record
04ONE CALL

The same question always gets the same number

“What is a 30 day BTC 90000 call worth at 60% vol, spot 85000?” The agent turns it into a tool call, and the tool returns 3,824.36 USD today, tomorrow and on any machine. The Black-Scholes formula behind it is on the methodology page.

workflow.run_black_scholes
{
  "spot": 85000,
  "strike": 90000,
  "daysToExpiry": 30,
  "volatilityPct": 60,
  "riskFreeRatePct": 0
}
// result
{
  "callPriceUsd": 3824.36
}
05WHERE IT STOPS

What a deterministic tool does not do

  • It does not check that the model read the right inputs from the request: the right number for the wrong input is still wrong.
  • It does not choose itself: the model decides when to call a tool.
  • A tool that reads live market data is as good as that source on that day.
  • It covers options, forex, risk and statistics, prediction markets, on-chain safety, crypto spot and perpetual futures, not everything a trader might ask.
06QUESTIONS

Deterministic AI and LLM math errors, answered

What is deterministic AI?

A design in which a language model decides what to do and explains the result, while every number comes from a tool that returns the same output for the same input from a published formula. In trading math that means an agent calls a calculator for a position size, an option Greek or a VaR figure instead of writing the figure itself.

Are language models bad at math?

Not on a single, clearly stated question. In our October 2026 test, 97% of 630 answers were within 1% of the exact value. The remaining answers are the issue: 12 of the 17 wrong ones came with a stated confidence of 80% or more, so a reader cannot tell them from the right ones.

Where do LLM math errors show up in trading?

In October 2026 statistics over a series of returns and the direction of inverse (coin-margined) contracts account for most of them. No model made an error on: Black-Scholes price, delta, gamma and vega; implied volatility.

Does a deterministic tool make an AI agent fully reliable?

No. A tool returns the right number for the inputs it is given. It cannot know whether the model read the right inputs from the request, and a tool that reads live market data depends on that source. What it removes is one failure: a wrong figure that looks right.

How can I check a TradingCalc number without trusting us?

Every MCP result carries a signature that can be verified offline with a published public key, the formulas and test vectors are public, and the verification state is recorded once a day on Solana. The steps are in the documentation linked on this page.

07USE IT

Give your agent numbers it can check

The tools are reachable from any MCP client or with a plain HTTP call, and the free tier needs no signup.

For developers API and MCP reference MCP server