Headline — paired generation samples

Placeholder. There are no samples yet — they populate only after the first rung (S1.0) ships, which it has not. Nothing below is a model output; it is the template that real paired samples will fill. Do not read the condition labels as results.

A “paired sample” will show the same prompt run two ways — once normally, once under a steering intervention — so the difference is visible side by side. Planned layout:

[Prompt] (category)

  Baseline:  <model output>
  C1 suppress transport@L0:    <output>
  C2 suppress routing_heavy@L12: <output>
  C3 suppress hybrid all:        <output>
  C4 random-class trap:          <output>
  C5 all-heads trap:             <output>

For each principled condition, paired samples illustrate:

• What the model produces by default
• What it produces under the intervention
• Class-discrimination across conditions

If S1.0 SHIP: 5–10 paired samples, one per category × condition, plus Δlogit heatmap.