Aim 4.5: Random direction control for category rankings

From EXPERIMENT_QUEUE.md — Planned (run next)

Project 200K FineWeb-Edu + 200K LMSYS onto 10 random unit vectors in Qwen3-32B layer 32 (5120-D).

For each direction: take top/bottom 200 docs, classify with same taxonomy (genre, author stance).

Also project 3,514 category docs (from category projection experiment) onto same 10 directions, compare median rankings.

Key question: do the 4 BH-surviving findings (Genre + Author Stance × 2 corpora) appear on random directions too?

• If yes → axis is not special, category structure is generic high-D geometry
• If no → axis captures genuinely specific discourse mode structure

Compute: ~90 min model inference (speculators+vLLM on H200) + ~30-60 min Claude classification (~4K docs) + analysis. Pod: thomas-rebuttals (4x H200). Estimated cost: ~$5-15 Claude API.