1 · Training corpus stage 1 · export
Reads ./training-data/*.jsonl if present.
No corpus yet. Run python scripts/export.py to walk your CC sessions and estate specs.
What goes in
- GREEN · Simon's prompts from CC session logs (user role only)
- GREEN · .md specs, READMEs, seeds across the estate
- AMBER · assistant responses (locked behind --include-amber)
2 · Ring distribution R0→R6
Heuristic ring tag per training row. Mountain bloom is the target.
3 · Source breakdown where it came from
4 · Sovereignty timeline year 1 → year 4+
Y1 · nowExport green corpus. Format LoRA pairs. Ship the scaffold. Wait for GPU.
Y1 · q3First LoRA on Llama-3.1-8B. Merge into simon-mind. Register in Ollama. Cascade T0 = self.
Y2Continual SFT on every new estate prompt. Bloom drift monitored by R5 mirror.
Y3Mesh: simon-mind + thomas-mind handshake. Bloom-routed delegation. Real WebRTC.
Y4+Guild constellation. Frontier-class capability without frontier dependence. v21 never happens — runewords on v20.
5 · Cascade routing T0 → T3
T0simon-mind · LoRA on local box · first responderawaits GPU
T1Ollama Llama-3.1-8B fallback · if simon-mind dips on confidenceawaits GPU
T2Estate cascade (FallCore proxy) · mesh peerslive
T2.5Llama-3.1-70B-q4 · local heavy liftawaits GPU
T3Frontier API (Anthropic) · last resort · rarely neededon demand
6 · Local model status 127.0.0.1:11434
checking Ollama…
checking simon-mind…
checking llama3.1:70b…
If Ollama is offline this panel stays silent — the page still works.
7 · Hardware ready stage gates
[ ]RTX 4090 (24 GB) or equivalent — LoRA SFT on Llama-3.1-8B
[ ]200 GB free SSD — base model + 70B + checkpoints
[ ]CUDA 12.1+ · unsloth[colab-new] installed
[ ]HuggingFace token · gated Llama access
[ ]Ed25519 key sourced · mint Konomi cert · (skipped now · v2.1 audit note)
[x]Estate registry entry · prime 587 assigned
[x]Green-rule export script · scripts/export.py
[x]Mesh router scaffold · mesh/router.js
8 · Guild mesh fall-signal
Members detected via BroadcastChannel fall-signal. Real WebRTC plugs in once a peer answers.
Stub. Real handshake lands after the first LoRA is trained.