Design nanoGPT memorization experiments: AdamW vs Muon with clip_xmax monitoring

[charlesmartin14]

Scope

Adds the independent folder baseline/experiments/nanogpt_memorization/ with an experimental design and tested measurement components. Existing baseline source and configurations are unchanged.

This is not yet an end-to-end training implementation. No training or actual WeightWatcher analysis was run in this preparation session. The remaining controlled data/training adapter and acceptance gates are explicit in TRAINER_CONTRACT.md.

Design

• Inherit the exact source-backed one-head nanoGPT recipe at commit 3749c36334382a20e48bfe2473c1dc4a1470a830, with recipe/model/optimizer/spectral blob checks.
• Compare AdamW with standard Muon + auxiliary AdamW, not silently with MuonClip.
• Separate clean controls, natural/random sequence repetition, finite-universe canary exposure, synthetic key–value associations, and an optional retention fork.
• Distinguish whole-sequence greedy extraction, token match, teacher-forced NLL/accuracy, exact canary rank/exposure, and fixed-boundary prefix compression.
• Require fix_fingers='clip_xmax', ERG=True, randomize=True through the existing CPU-clone/RNG-isolated WeightWatcher monitor. Preserve per-layer raw/clipped alpha and fit/correlation evidence.
• Five paired seeds; fixed token budgets; validation-only selection; no inference of a globally optimal recipe or a causal spectral memorization detector.

Implemented here

• Source-pinned study.json and plan/config resolver (does not start jobs).
• Behavioral metrics, exhaustive finite candidate ranking, random-token probes, and exact presentation schedules.
• Strict wrapper around the existing WeightWatcher hook.
• Full protocol, integration contract, and honest execution ledger.

Validation

• python -m pytest -q baseline/experiments/nanogpt_memorization/tests: 32 passed, CPU; rerun also passed.
• All 10 uploaded files verified against local Git blob hashes.
• Tests use a controlled transition model, not nanoGPT training-performance evidence.
• Actual pinned-model optimizer-step, end-to-end sampler, checkpoint/resume, and target-hardware WeightWatcher tests remain outstanding.

Remaining before training

Implement natural-span reservation/decontamination, association/rule generators, paired full-record intervention sampler, training/evaluation/checkpoint integration, and the explicit preflight gates. Then begin with the four-run pilot rather than automatically launching the full 30-run campaign.