Core concepts
Checked numbers
Before an answer reaches you, each figure in it is checked against the tool results of the session. A figure that cannot be traced goes back for correction, and is cut if it still fails.
How it works
For every figure, the model declares where it came from:
- observed — a value a tool returned;
- derived — computed from observed values, with the formula written out;
- proposed — a target or an assumption, not a measurement;
- cited — taken from a source named on the same line;
- count — how many of something.
The check verifies each figure by its role: an observed value must match a tool result, a derived value must follow from its operands, a cited value must name its source. The declarations are removed before the answer is shown.
Which numbers are checked
Decimals, percentages, currency amounts and numbers in table cells are measurements and are checked. Dates, years, codes, ordinals and plain integers in a sentence are not, except integer prices of instruments quoted at 1,000 or more. A rounded figure must be the tool value correctly rounded to the digits it shows, and within 0.5% of it: 30.2052% may be written 30.21% or 30.2% but not 30.20%, and 0.82467 may be written 0.825 but not 0.82.
Backtest reports
Since 0.1.16, what a backtest wrote about itself counts as evidence for the report about it: every value in the run card, the metrics, the risk X-ray, the Monte Carlo validation and the rebalance summary. A per-bar table counts only for the rows the agent actually read. A file the model wrote itself never counts, and a value from one backtest does not support a figure attributed to another.
What you see
- While it runs: Checking the figures in this answer (round N)…
- A figure that fails is sent back with a specific correction, at most twice.
- A figure that still fails is replaced by
(omitted※), and a footnote says how many were cut. The rest of the answer is released. - The whole answer is replaced by a short refusal only when a price question had no fetched price, when the answer described a different instrument than the one fetched, or when cutting the failing figures still does not pass.
Getting your figures through
- Ask the agent to fetch the data before it quotes numbers.
- When comparing backtests, let it name the run each figure comes from.
- When you supply a number yourself, say where it comes from ("per the 2025 annual report").
- Assumptions such as a discount rate are shown as proposals, not as measurements.
Limits
The check shows that a figure matches the data of the session. It does not show that the data is right or that the analysis is sound. Plain integers in prose, units written as words, and a correct value quoted under the wrong name are not caught.