08 — Synthetic readers
Summary
The v0.33.61 screenshot-boundary brief operationalised. Two agents: a renderer that is a caller of the estate's existing browser-automation service, configured never to pass text or structure on; and a reader that receives pixels and nothing else, is never told the page's purpose, and clicks spatially. The page under test is a fixed artefact authored before the run — a page generated during the run measures the model agreeing with itself — and the patience budget is set exogenously so abandonment is a measured event rather than a coherent story. The honest limit is stated as the method's credibility: synthetic readers find defects and cannot report preferences. Rule one puts the simulation marker in the filename, the headers and beside every quote, because export is where markers die; rule two's narrow exception is the archetype, with a test checkable by somebody other than the author.
Key concepts
- The screenshot boundary — the instrument, not a limitation — and somebody will try to improve it away
- The fixed comprehension question — what would you do now, and what did that page tell you — in those words, every time
- Portrait versus archetype — the property list is the archetype; the individual is only where properties were sampled
- The service boundary — comprehension questions only — and it is a property of the product, not of its terms of service
Key ideas
- A synthetic run should be difficult to quote misleadingly even by somebody trying.
- The runs worth publishing most are the abandoned ones, because those changed the design.
- The calibration record — published, including the misses — is what separates a product from a plausible-opinion generator.
- First budgets are stated to be argued with: six screens, ten minutes, eight clicks — and round two's come from round one's spend data.