pki.sgit.ai / documents / synthetic-readers

The Screenshot Boundary Is The Instrument And The Patience Budget Must Come From Outside The Model

TypeDev brief Versionv0.33.61 Date20 August 2026 AuthorDinis Cruz (project lead) and collaborators LicenceCC BY 4.0 Sourceraw markdown · view on GitHub

Summary

The tabletop programme for testing pages with synthetic readers, and the three methodological corrections that decide whether it measures anything. Two agents: one renders the site and takes screenshots (an existing browser-automation service — a caller, not a build), one acts as a persona that receives pixels and nothing else and answers with spatial clicks. The screenshot boundary is the instrument, not a limitation: a persona that can read the source understands the page better than any human could. The page under test must be a fixed artefact authored before the run — a page generated during the run measures the model agreeing with itself — and the patience budget must be set exogenously, because patience generated by the same model that generates the confusion will always cohere with it. The honest limit: synthetic readers find defects and cannot report preferences, so they clear the levels and humans judge the variants. And the brief raises, for the third time, the rule against simulating named individuals — proposing the narrow, testable exception: an archetype composed from several people that none of them would recognise.

Key concepts

Key ideas

On this site

Became document 08 of the Map Your Case pack — the synthetic-reader programme — with the two simulation rules carried verbatim into the pack's principles.

Read the document

📄 Original document · v0.33.61 · 20 August 2026 · rendered from the raw markdown (the source of truth)