360

Source document

[2026-03-17 19:08] xenonfun: ``` --- Session Summary Architecture validated - Mercury-delay-line field transport with oscillator transduction - Depth is the primary scaling axis (not K) - Retrieval is distance-invariant (DC@16 = DC@64 = DC@128) Key findings 1. Adam required — RotAdamW breaks retrieval (DC: 1.6% vs 90%) 2. Curriculum required — pure LM gives zero retrieval even with Adam 3. 2% curriculum sufficient — 59% DC; 20% gives 97.5%. BPB unaffected. 4. Retrieval partially persists — 54% retained permanently after curriculum removal 5. Depth scaling smooth — D=6→D=32: BPB 3.00→2.53, DC 72%→91% Optimization - antenna_fast.py: fused K=2 + bf16 + compiled → 1.37-1.53× speedup - seq=1024 → 843K tok/s (30% improvement over seq=256) - Dispatch analysis: antenna has 8× less work per dispatch than matched transformer - Hardware mapping doc: FPGA/Cerebras/Groq natural fit, cloud GPU testing path defined Locked default recipe Adam + K=2 + deep stack + 5% curriculum + channel split ``` [2026-03-17 19:13] xenonfun: ``` ⏺ Benchmark suite running. 5 models × ~5-10min each = ~30-45min total. Sequential, single GPU. Models being benchmarked: 1. Antenna default (K=2 D=24 d=768, 5% curriculum) — 1.69M params 2. Antenna thin-deep (K=2 D=48 d=384, 5% curriculum) — 1.60M params 3. Antenna no field (field disabled) — 1.69M params 4. Causal conv (d=128 L=12 k=5) — param-matched 5. MLP baseline (d=256 L=6) — param-matched Each gets: val BPB, DC@{4-128}, MT@{1-8}, generation metrics. This produces the clean publishable comparison table. ``` [2026-03-17 19:31] xenonfun: (files: image.png) [2026-03-17 19:59] xenonfun: claude down/degraded : ```⎿  529 {"type":"error","error":{"type":"overloaded_error","message":"Overloaded. https://docs.claude.com/en/api/errors"},"request_id":"req_011CZ9FA2Sn6LNEvKK9CEfPf"} ```