CR deviation — only the 32B rung was dropped; qwen7b was run as pre-registered

Recorded 2026-09-19. The pre-registration (PREREGISTRATION.md) and ladder.json name two weak local rungs: qwen2.5-coder:7b (rung 0) and qwen2.5-coder:32b (rung 1).

What was run vs. dropped

• Rung 0 qwen2.5-coder:7b — run as pre-registered. The main PC has an RTX 5070 Ti Laptop GPU (12 GB VRAM); qwen7b (~4.7 GB + context) runs entirely on the GPU, so the system-RAM pressure (15 GB total) is not a factor for generation. naive+gs, k=3, per the design.
• Rung 1 qwen2.5-coder:32b — DROPPED. 32B (~20 GB) exceeds the 12 GB VRAM and cannot run on this machine. This is explicitly permitted by ladder.json note_min_rungs: with 32B dropped, the rungs qwen7b → gpt-mid → {frontier} still span weak-local → mid → strong, and the ≥3-rung minimum holds.

Effect on the analysis

• The a-priori ladder order is unchanged (§8, no post-hoc re-ordering): qwen2.5-coder:7b → gpt-mid → gpt-frontier / gemini-frontier / claude-frontier.
• The monotone-trend test (PREREGISTRATION §7, Jonckheere-Terpstra / Page) runs over this ordered ladder. Dropping the 32B removes one interior point but does not change the direction being tested (weak → strong).

Note

An earlier attempt substituted llama3.2:3b for qwen7b out of over-caution about the 15 GB system RAM, before confirming the model runs on the 12 GB GPU. That was stopped and discarded before measurement; qwen7b is the model of record for rung 0. To also fill rung 1, re-run qwen2.5-coder:32b on a machine with ≥24 GB VRAM.