Phase 5: Run RunPod pilot ablation matrix (5k/25k) for paper-fidelity paths

[georgepullen]

Purpose

Execute controlled RunPod pilot ablation matrix to choose default fidelity path before scaling.

Mandatory Reading (blocking)

First comment must summarize:

• reports/NL_IMPLEMENTATION_ORACLE.md section 6.4 summary matrix
• reports/ablations.md
• docs/zeroshot_eval.md
• docs/continual_eval.md

Scope

Run matrix with fixed seeds/data/tokenizer:

• baseline legacy path
• delta_rule
• dmgd_mlp
• muon_ns
• CMS projection off/on
• stopgrad vs differentiable online writes (if available)

Required Eval Bundle per run

• zeroshot
• niah
• continual
• passkey
• pg19

Deliverables

• Run manifest table with command, checkpoint, log, eval paths.
• Comparative summary with recommended default.

Acceptance Criteria

• 5k and 25k variants completed for all enabled matrix entries.
• No untriaged failed runs.
• Recommendation justified by metrics and stability evidence.
• First issue comment contains mandatory reading summary.