[claude] Research Triage: SSD Self-Distillation acknowledgment (#128) #165

Merged
claude merged 1 commits from claude/issue-128 into main 2026-04-07 02:07:55 +00:00
Member

Fixes #128

What this does

Creates docs/research-ssd-self-distillation-2026-04.md — a full research acknowledgment for the Apple SSD paper (arXiv:2604.01193).

Contents

  • Assessment: P0 relevance — +7.5pp on Qwen3-4B is exactly our operating scale
  • Fork/Lock theory: explains why temperature asymmetry matters and the fleet implication (we are leaving performance on the table with single-temperature configs today)
  • 4-phase implementation plan:
    • Phase 1: Temperature audit (zero cost, this week)
    • Phase 2: SSD pipeline on Qwen3-4B via Ollama + unsloth (Mac-feasible, 1–2 weeks)
    • Phase 3: Nightly self-improvement loop integrated with burn mode + RetainDB (#112)
    • Phase 4: Sovereign fine-tuning confirmation
  • Infrastructure assessment: documented RAM/storage/time requirements — all Mac-feasible
  • Risks and mitigations: domain transfer, capability regression, data scale
  • Acceptance criteria status: all 5 criteria addressed or tracked
  • Recommended follow-up issues: 3 scoped issues for execution
Fixes #128 ## What this does Creates `docs/research-ssd-self-distillation-2026-04.md` — a full research acknowledgment for the Apple SSD paper (arXiv:2604.01193). ## Contents - **Assessment**: P0 relevance — +7.5pp on Qwen3-4B is exactly our operating scale - **Fork/Lock theory**: explains why temperature asymmetry matters and the fleet implication (we are leaving performance on the table with single-temperature configs today) - **4-phase implementation plan**: - Phase 1: Temperature audit (zero cost, this week) - Phase 2: SSD pipeline on Qwen3-4B via Ollama + unsloth (Mac-feasible, 1–2 weeks) - Phase 3: Nightly self-improvement loop integrated with burn mode + RetainDB (#112) - Phase 4: Sovereign fine-tuning confirmation - **Infrastructure assessment**: documented RAM/storage/time requirements — all Mac-feasible - **Risks and mitigations**: domain transfer, capability regression, data scale - **Acceptance criteria status**: all 5 criteria addressed or tracked - **Recommended follow-up issues**: 3 scoped issues for execution
claude added 1 commit 2026-04-07 02:07:23 +00:00
Research triage for Apple's Simple Self-Distillation paper (arXiv:2604.01193).
Documents fleet-relevant findings: +7.5pp on Qwen3-4B (our exact scale),
zero-infrastructure fine-tuning, Fork/Lock theory implications for temperature
tuning, and a 4-phase implementation plan from immediate T_eval audit through
nightly self-improvement loop.

Refs #128
claude merged commit 59653ef409 into main 2026-04-07 02:07:55 +00:00
Sign in to join this conversation.