---
name: research-integrity-contract
description: Run AI-agent research (market studies, datasets, competitive scans) under an integrity contract that makes the output citable: verbatim-or-empty extraction, source URL per row, frozen statistic definitions and an independent recount before publishing. Use when the user asks agents to collect data, build a dataset, scan a market or produce research others will cite.
---

# Research integrity contract

Agent research is fast; this contract is what makes it publishable. It was
built running a real multi-agent market study (79 job postings, six parallel
agents, one capture day) where the independent audit caught two errors that
would otherwise have shipped.

## The contract (apply to every row the agents produce)

1. **Verbatim or empty.** Text fields are copied literally from the source.
   Anything not visible stays empty. Never estimated, never paraphrased into
   what the model expects the source to say. A summary cannot be checked
   against the source; verbatim text can. An honest gap can be filled later;
   an invented datum poisons every statistic that touches it.
2. **URL per row.** Every row carries the address where the data was seen.
   Mirrors, estimates and anomalies get labeled on the row itself. Per-row
   provenance is what separates a citable dataset from an opinion with a table.
3. **Nothing invented.** No people, contacts or entities that the sources do
   not show. Out-of-scope rows are kept and flagged, not deleted, so the
   cleaning stays inspectable.

## The pipeline shape

Split the ground into small coherent segments (by geography, language or
category) and run one agent per segment in parallel, all delivering into the
same schema. Parallelism compresses weeks into a day; the segment split keeps
each agent's context small and its extractions clean.

## Freeze the statistics

Every figure you will ever quote gets a written definition and is computed
once against the dataset, with the result stored in a file. From then on
everyone cites the file; nobody recounts from memory. Without frozen
definitions the same question yields different numbers depending on who
counts and when. Write the exact wording rules too ("38 of 79 demand
measurable KPIs" is not "38 demand ROI").

## The independent recount

Before publishing, someone or something that did not build the pipeline
recounts every headline figure from the raw data, with publication blocked
until all reproduce. Budget half a day: it is the cheapest insurance in the
project. Real catches from the reference run: a compliance false positive
(the string "AI Act" matching "AI activities") and a headline count whose
definition lived only in someone's head.

## Instrument before you run

Log tokens, agent time and human review minutes from the first minute. The
reference study skipped this once and as a result could not publish its cost
per row; the lesson is now a rule. Disclose sampling bias plainly (what your
capture method over- and under-represents).

## Source and method

The full case with the pipeline, failures and replication guide:
[y8y.ai](https://y8y.ai/es/casos/investigacion-de-mercado-con-agentes-de-ia/).
Author: Sebastián Ocampo ([LinkedIn](https://www.linkedin.com/in/seb-ocampo/)).