---
id: KB-MA-020
url: https://app.codecontract.io/help/marta/when-the-answer-starts-from-bad-data
idioma: en
categoria: marta
audiencia: usuario
nivel: intermedio
actualizado: 2026-08-13
tambienEn: [es]
relacionados: [KB-MA-004, KB-DI-005, KB-MA-012, KB-MA-019]
citadoPor: [KB-MA-018]
---

# When the answer starts from bad data

_An answer can be perfectly reasoned and false, because the value it started from was misread._

**Responde a:** marta gave me the wrong date · the answer does not match the document · the assistant gives incorrect data · why does it quote an amount that is wrong

You ask when a certificate expires and get a date. The date is wrong, but not because it was invented: it is the one on record, and that one was stored wrong the day the document was read. The answer is faithful to what is there; what is there is what fails.

## Two failures that look alike and are fixed differently

| Symptom | Where it comes from | How it is fixed |
| --- | --- | --- |
| It says something that exists nowhere | It filled a gap | Ask for the source and check |
| It says a stored value, but a wrong one | It was misread from the document | Fix the value on its record, not the answer |
| It says an old value | There is a newer version unrecorded | Register the new version |
| It cannot find something that exists | Out of reach or badly named | A different problem with a different fix |

> [!IMPORTANT]
> The second row is the most deceptive, for one specific reason: **if the value is stored wrong, the answer will be coherent, confident and wrong — and asking again gives the same result**. Rephrasing the question does not expose it; only opening the document does. So when a figure or a date will have consequences, the check is not rereading the answer: it is looking at the paper it came from.

## How to catch it early

1. **Ask where it comes from before using it** — A value with a document behind it is checked in ten seconds.
2. **Distrust anything that sounds too neat** — First-of-the-month dates, round amounts: the readings that fail most.
3. **And cross-check against something you know** — If the supplier invoices monthly, an annual figure jars by itself.

> [!WARNING]
> What to do on finding one, and almost nobody does: **fix it at source, not only in the email you were writing**. Correct the figure in the message to the client while the stored value stays wrong and it will surface again tomorrow — in a report, in an expiry alert, or in the answer given to a colleague. A misread value does not bother you once: it bothers you every time anyone asks, always with the same confidence.

## What reduces these cases at the root

**En corto**

- Check the values on important documents when uploading, not when using them.
- Extract only what you will actually consult: fewer values, fewer inherited errors.
- And treat the first use of a value as the review that never happened.

> [!NOTE]
> This is not specific to the assistant: any report, alert or dashboard using that value gives the same result. The difference is that an answer written in plain language sounds more convincing, and so gets checked less.

**Can I ask it to check the document?**

It can cite where the value comes from; checking the paper says so is yours.

**What if the document is wrong, not the reading?**

Then it is the issuer's problem: ask for a correction and record it.

**Is it worth reviewing old values?**

Those in use, yes; start with the ones driving an expiry date.

## Ejemplos

**A company sends a client an expiry date that turned out to be misread.**

- Fixes the value on the document record, not only in the email

→ The renewal alert lands when it should and nobody else receives the wrong date.

**The answer rests on a misread figure.**

- Corrects the figure in its document

→ The next answer is already right.

**The error is corrected in the answer, not at source.**

- Corrects it where the figure originated

→ The error does not return tomorrow.

**An odd figure is spotted and ignored.**

- Checks the original document

→ The error closes on detection.

**The same bad figure affects several answers.**

- Corrects the source and asks again

→ Every answer is fixed at once.

**A decision is taken on an answer built from a bad figure.**

- Checks the source before deciding

→ The decision rests on what is true.