Aim 4.3: Assistant axis relationship to assistant chat data

From EXPERIMENT_QUEUE.md — Planned (run next)

If FineWeb does contain AI chat data: compare axis projections of chat-like docs vs non-chat instructional docs.

If it doesn't: inject synthetic AI chat transcripts into the projection pipeline and measure where they land on the axis.

Also: project LMSYS-Chat-1M samples (real human-AI conversations) onto the axis — do they cluster at the top?

Key question: is the assistant axis specifically tuned to AI-chat-style text, or does it capture a broader instructional/helpful discourse mode that chat data happens to fall within?

Compare axis projections:

• (a) real AI chat transcripts
• (b) human-written instructional text (textbooks, tutorials)
• (c) human-written didactic text (lectures, how-tos)
• (d) creative/personal writing (control)

If chat >> instructional → axis is chat-specific (contamination story). If chat ≈ instructional >> creative → axis captures discourse mode (structural story).

Compute: ~2h (projection of new data + Claude classification). Pod: thomas-rebuttals (needs Qwen 3 32B for projections).

Depends on: Aim 4.2.