docs(M288): V1_004 3-knob dispatch recipe — post-M32d toolkit

[noahgift]

Summary

Recipe doc for the next-step V1_004 bench dispatches. Documents the 3-knob toolkit shipped via paiml/aprender M32d follow-ups:

• Knob 1: temperature/top_k/top_p (aprender#1842)
• Knob 2: repetition penalty (aprender#1844, supersedes #1843)
• Knob 3: streaming SSE (aprender#1835 contract; impl pending)

What this doc captures

• 3-knob table with tuning intuitions
• 4 sub-bench scripts (A: sampling alone, B: sampling+penalty, C: penalty alone, D: all 3 + small max_tokens)
• Outcome interpretation framework (4 axes)
• Companion-side prerequisite: env-var plumbing for 5 new knobs through bench script → ccpa-arena-bench → apr code → `QuantizedGenerateConfig`

Why now

M287 evidence locked in the greedy baseline: 6/20 fixtures all `driver_error turns_before_error=4`. M32d works (4× turn throughput vs pre-M32d 0) but the 30B-MoE is stuck in textual loops. The 3 knobs each address a different theory of WHY. This doc gives the operator a structured sequence to test each.

What this is NOT

• NOT a V1_004 discharge — sub-benches not yet dispatched
• NOT a binding sequence — operator may pick any subset
• NOT companion-side code — env-var plumbing is a separate follow-up

Mechanical recipe doc. M-counter NOT bumped.

Test plan

• `bash scripts/test-doc-drift.sh` (if reachable)
• CI gate + workspace-test (pre-existing inherited failure; admin bypass)

🤖 Generated with Claude Code