SPB Git

spb/localvm-research Public License

Running LLMs larger than memory on a consumer Mac — falsification-driven research: margin-gated deferred refinement, out-of-core verification on Apple Silicon. TR-01 published.

Python 63.2% JavaScript 23.5% CSS 11.8% Shell 0.9% Makefile 0.5%

History of research/LOG.md · clear filter

  1. Publications: TR-01 technical report + figure generator + site tab
    …
    - docs/publications/TR-01: abstract, substrate, expG/expD + three
  refutations, architecture, 1.7B continuum, 32B out-of-core, honest
  prior-art attribution (per novelty_check), limitations, repro
- tools/make_pub_figures.py: 4 SVG figures regenerated from results/
  (validated palette, collision-checked)
- Web: /publications tab, publication cards, /raw asset route, image
  link rewriting in markdown, figure styling

Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
    Simon-Pierre Boucher committed 3 h ago (Aug 12, 2026) · 1 file changed +15
  2. candidate_01: W=64 lever NEGATIVE — rollback rate bounds sweep amortization
    …
    Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
    Simon-Pierre Boucher committed 4 h ago (Aug 12, 2026) · 1 file changed +13
  3. Web v2: major visual upgrade, hamburger mobile menu, comments, SVG logo + Phase 11 novelty check
    …
    - Redesign: blur sticky header, hero gradient, elevated cards, footer grid
- Mobile: slide-down hamburger menu, full responsive pass
- Comments: persistent JSON storage (data/, excluded from deploys),
  honeypot + per-IP cooldown, public discussion page
- Logo: thesis-as-image (large model, small resident set) used as favicon
- research/novelty_check.md: components anticipated (QSpec, Lever, CATS,
  UGSD, SpecExec, mlx-flash); composition not found; 4 narrow surviving
  claims (N1-N4); attribution obligations recorded

Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
    Simon-Pierre Boucher committed 4 h ago (Aug 12, 2026) · 1 file changed +31
  4. candidate_01 32B scale run: charter 16.A/B demonstrated in prototype form
    …
    q8-32B (does not fit beside base on 48 GB) streamed at ~11.6 GB/s
sequential; +0.28 nats over pure-q4 (only fitting alternative);
3.72 GB/token logical (9.4x under checkpoint); 1.69 tok/s with
measured headroom.

Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
    Simon-Pierre Boucher committed 4 h ago (Aug 12, 2026) · 1 file changed +22
  5. candidate_01 v2 results: q8-equivalent quality with q4 resident (bf16 judge)
    …
    verify-all -0.1297 vs pure-q8 -0.1328 at 237 MB/token (9x under
checkpoint); tunable tau continuum 47%/83%/100% of quality gap.
Charter 16.D/E criteria met at 1.7B scale.

Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
    Simon-Pierre Boucher committed 4 h ago (Aug 12, 2026) · 1 file changed +19
  6. expG@8B: favorable scaling; candidate_01 first run: bytes pass, metric insight
    …
    8B: esc@99% falls to 22.6% at 4-bit. Candidate: 154-231 MB/token logical
(9-14x under checkpoint); verify-all at hardware ceiling (q8 self-
agreement 1.56%/token prefill/decode flips); margin-mode seq-fidelity
fails as registered -> quality-level metrics now mandatory.

Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
    Simon-Pierre Boucher committed 5 h ago (Aug 12, 2026) · 1 file changed +23
  7. expA+expB: NEGATIVE results — no pageable/cacheable dynamic sparsity on dense SwiGLU
    …
    expA kill: 77% of 64-neuron blocks for 95% energy; expB kill: 99.9%
union at 128 tokens, Jaccard ~ random null. Fine-grained escalation
routes all closed; evidence selects C1->C2 merge (batch-amortized
margin-gated refinement).

Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
    Simon-Pierre Boucher committed 5 h ago (Aug 12, 2026) · 1 file changed +26
  8. expF: NEGATIVE result — quantization damage diffuse across depth, no layer leverage
    …
    1.5x sensitivity spread (kill <2x); top-29% of layers repair only 24%
of lost agreement; masking + cooperative repair effects documented.

Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
    Simon-Pierre Boucher committed 5 h ago (Aug 12, 2026) · 1 file changed +19
  9. expD: results — progressive coding validated, two-tier policy 97.6% @ 35% esc
    …
    Monotone convergence, gating composes across stages, static-parity
caveat recorded; bytes-per-escalation identified as binding unknown.

Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
    Simon-Pierre Boucher committed 5 h ago (Aug 12, 2026) · 1 file changed +20
  10. expG: full results — margin gating promoted (AUROC 0.85-0.98), affine 2-bit base dead
    …
    48 trajectories x 4 bit-widths on Qwen3-1.7B; esc@99%: 60/37/5% at
3/4/8-bit; ParetoQ cliff confirmed at naive 2-bit (3% agreement).

Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
    Simon-Pierre Boucher committed 5 h ago (Aug 12, 2026) · 1 file changed +22
  11. Phase 4: candidate ranking + web platform deployed to www.localvm.dev
    …
    4 candidates selected (Progressive-Residual Runtime preferred; Gate-Zero
falsifier battery runs first). Platform live on M3U96b via PM2+ngrok;
sync-content.sh filter-order fix (8 GiB testfile exclusion).

Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
    Simon-Pierre Boucher committed 5 h ago (Aug 12, 2026) · 1 file changed +36
  12. Phase 3: research gaps — 24 falsifiable approaches (G01–G24)
    …
    Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
    Simon-Pierre Boucher committed 6 h ago (Aug 12, 2026) · 1 file changed +16
  13. Phase 2: state-of-the-art map (research/state_of_the_art.md)
    …
    ~35 technique groups in six families, charter §5 fields, overlap
analysis, settled-vs-open ledger, expH substrate numbers integrated.

Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
    Simon-Pierre Boucher committed 6 h ago (Aug 12, 2026) · 1 file changed +23
  14. expH: full run results + analysis — SSD ceiling ~13.1 GB/s iostat-validated
    …
    Cold random reads 67 MB/s (4KiB QD1) to 13.8 GB/s (1MiB QD8, at ceiling);
<5% degradation under saturated Metal matmul; cache-contaminated cells
flagged invalid; hypothesis confirmed. Storage is not the binding
constraint at >=256KiB QD>=4 access patterns.

Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
    Simon-Pierre Boucher committed 6 h ago (Aug 12, 2026) · 1 file changed +20
  15. Phase 1: literature sweep complete — 5 theme notes, ~300-source bibliography
    …
    research/notes/{quantization,sparsity_pruning,out_of_core_memory_systems,
decomposition_progressive,speculation_error_stability}.md + merged
bibliography.md + LOG entry with converged findings.

Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
    Simon-Pierre Boucher committed 6 h ago (Aug 12, 2026) · 1 file changed +36
  16. Bootstrap: charter-compliant skeleton, header enforcement, hardware manifest, experiment scaffolds
    …
    - Repository layout per CLAUDE.md §3
- tools/check_headers.py enforcing §0.1 (58 files pass)
- tools/new_experiment.py scaffolding with seven-field hypothesis block (§10)
- benchmarks/hardware_manifest.py: macOS-native fingerprint (M5 Max / 48GB verified)
- Eight micro-experiment scaffolds (§9, exp A–H)

Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
    Simon-Pierre Boucher committed 6 h ago (Aug 12, 2026) · 1 file changed +29