-
Sun, Nov 1, 2026 7
-
Figures (sim, capacity sweep, two-node), paper results v0.1 with two-node evidence, LOG
…
Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com>
-
Gate E: authenticated TCP transport, expert worker, remote executor with local fallback (localhost bit-identical), distributed bench, power sampler, LOG
…
Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com>
-
SSD tier: parallel demand reads + expert-contiguous TITAN layout (cold 16-slot 8-10 -> 36 tok/s); sim v2 results; probe fix; LOG/DECISIONS
…
Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com>
-
Qwen3-30B parity on M3U96a (bit-identical, 121.7 vs 121.5 tok/s), make_tables.py, cold-NVMe sweep script, LOG
…
Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com>
-
Gate C: predictors (prev_rank, transition depth d, probe), trace simulator, async prefetcher, lazy installs, miss-cost breakdown, v2 traces, LOG/DECISIONS
…
Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com>
-
Gate B/C: sync-free fast path, bench_decode, capacity sweep, routing traces, link/net probes, LOG
…
Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com>
-
Gate A/B: scaffold, mandate, inventory tooling, prior art (59 refs), registry/store/slot-table runtime, numerical tests bit-identical to mlx-lm
…
Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com>
-