# Conditions: the cross-generator specificity 2x2

## Design

A 2x2 factorial, identical to the xAI clean 2x2 in the companion receipt, rerun
on Gemini Flash:

| Factor | Levels |
|:-------|:-------|
| Specificity | present / absent |
| Quality demands | present / absent |

Four cells: SPEC_QUAL, SPEC_ONLY, QUAL_ONLY, BARE. 10 runs per cell, 40 outputs.

- Generator: gemini-3-flash-preview
- Length control: NONE. Instead of capping output length, scores are normalized
  to markers per 1,000 words (density). This is the design choice that resolves
  the length confound; see the README.
- Task: the same Northvane Analytics strategic-analysis prompt used in the xAI
  kit, so the only differences from that kit are the generator and the
  length-control method.

## Scoring

Each output is scored on six specificity markers (company, scenario, market,
algorithm, competitors, specific recommendations) plus word count, from which
density per 1,000 words is computed. The specificity effect is the
specificity-present cells (SPEC_QUAL + SPEC_ONLY) against the specificity-absent
cells (QUAL_ONLY + BARE), on the density measure.

## Why Gemini Pro is absent

Gemini Pro was run on the same design but truncated its outputs to roughly 60
words, which makes the density measure unstable. That arm is inconclusive and is
not reported as a result. Cross-generator here means xAI and Gemini Flash.