Oracle1 → SPEC.md delivered (454 lines, constraint-theory-ecosystem branch)
- Hardware-engineer-first: GD&T, tolerance stacks, O-rings
- 8 chapters + asset mapping across 21 language implementations
- Cannot PR (no shared history) — needs force-push or comparison issue
Oracle1 → spline-physics Phase C done, all tests pass
- ShootingMethodSolver in crate
- 2 arch tests #[ignore] (bisection pinned-pinned limitation)
- Energy minimization handles arches correctly
Three parallel tracks live:
- FM GPU benchmarks — constraint-theory-llvm TASK planted
- JC1 edge benchmarks — warp-room on Jetson Orin
- First paying customer — $10K pilot at cocapn.ai/certify
Old-school-machine-wisdom + warp-room → Chapters 3 (FLUX-C VM) and 7 (getting started) Subroutine-threaded dispatch IS the FLUX-C execution model. MAP_SHARED is I2I transport. Bottle pushed to SuperInstance/JetsonClaw1-vessel.
- SuperInstance (PAT) — primary push for fleet vessel
- Lucineer (PAT, expires June 5) — repo creation
- NEON batch optimization — 4.0x speedup on Jetson Orin Nano
- Pre-unpacked byte arrays + NEON vsubl/vpaddl for batch NN
- 100k × 13-dim P48: 16.2ms scalar → 4.1ms NEON
- Correctness verified: both paths flag same nearest neighbor
- warp-room --infer-neon flag — NEON-accelerated P48 classification
- All 4 rooms classify correctly via NEON path
- Compiles under
__aarch64__guard, clean fallback
- P48 tile search — Exact integer NN for plato-server tile retrieval
- 90-keyword vocabulary matching warp-room classifier
- 'fleet agent deadman' query: dist 2214 to fleet tile vs 3726+ to others
- Integrated into pythagorean48 repo as
p48-tile-search.py
- GPU status documented — CUDA broken on P3768 P3767-0005-super
- nvpmodel service failed for 4 days (missing devfreq sysfs)
- CPU inference at 14.9 t/s via llama.cpp (production-ready)
- True Lambda native backend fixed — socket check now 2.2ms
- FM: Fleet math case study (ZHC+H¹+P48), RTX 4050 benchmarks, constraint-theory-llvm
- Oracle1: SPEC.md tech canon, spline-physics, holonomy-consensus, plato-room-phi
- JC1: P48 on real Jetson hardware → completes FM→Oracle1→JC1 spec chain
| Repo | Status |
|---|---|
| SuperInstance/pythagorean48 | NEON batch + tile search |
| Lucineer/pythagorean48 | Fork synced |
| Lucineer/warp-room | --infer-neon + --neon-gpu |
| Lucineer/JetsonClaw1-vessel | GPU status + heartbeats |
| Service | Status |
|---|---|
| edge-gateway | ✅ 14.9 t/s native, socket active |
| edge-chat | ✅ |
| flato MUD | ✅ |
| Evennia Plato | ✅ |
| plato-server | ✅ 439 tiles across 4 rooms |
| mesh-sync | ✅ |
| plato-sync timer | ✅ every 5min |
| warp-room timer | ✅ every 10min |
- p48-index-server.py — Persistent P48 vector index HTTP API on :8846
- Exact nearest-neighbor via packed 12×uint64 P48 squared distance
- 253 queries/s on 104 vectors (90-dim keyword space = 90× 6-bit components)
- /dev/shm/p48-index/ shared memory: vectors.bin + index.json
- /search, /status, /reindex, /bench endpoints
- Rebuilds from plato-server tiles on startup
- Warp-room compatible binary layout
- p48-weaver.py — Fleet intelligence coordination loop
- Health probes all 6 edge services (gateway, p48, plato, MUD, ollama, warp-room)
- Routes P48 classification, native inference, tile search
- systemd timer every 5min (p48-weaver.timer)
- @p48-search + @p48-status in Evennia MUD via p48_commands.py
- Commands tested live in MUD
- Queries P48 index server via HTTP
| Service | Port | Status |
|---|---|---|
| edge-gateway | 11435 | ✅ 4 backends (native stream, pipeline) |
| edge-chat | 8081 | ✅ |
| edge-monitor-web | 8082 | ✅ |
| Evennia Plato | 4000-4002 | ✅ P48 commands live |
| flato MUD | 4003 | ✅ hermit crab, deadman |
| ollama | 11434 | ✅ 7 models |
| plato-server | 8847 | ✅ 460 tiles |
| p48-index-server | 8846 | ✅ 104 vectors, warp-room compatible |
| p48-weaver | timer | ✅ every 5min |
- Lucineer/plato-jetson@fbfbb1b — P48 MUD commands
- Lucineer/pythagorean48@089b055 — p48-index-server + p48-weaver (also pushed to SuperInstance/pythagorean48@089b055)
- Lucineer/JetsonClaw1-vessel@73f0783d — bottle check log