mirror of
https://github.com/SHOGGOTH-SECTOR/sica-fondt.git
synced 2026-07-31 16:16:26 +00:00
20e99f434c
Wave 1 remainder + Wave 2 design-first specs, completing one build plan per sub-organ: - B1 tarot emulator, B2 SOUL/Big-3, B3 Celtic Cross - C2 4+4 metacog, C3 inference cycle, C4 Brain - D1 Ada border, D2 the medium, D3 mini-rag - E1 3 GnuCOBOL invariant stores, E2 VARIANT stores, E3 RAG family - F1 MoRAG, F2 SAE monitor, F3 subagents - G1 stress-loop contract, G2 governance, G3 defense model Each: contract + stubs for independent build; unfitted constants marked C1. https://claude.ai/code/session_015hmgREHNsxYCuim33yUF2c
1.9 KiB
1.9 KiB
F1 — MoRAG (mixture-of-RAG) (DESIGN-FIRST)
1. Component
The context assembler in the periphery: BERT classifies/routes the incoming situation, LoRAs specialise, and it injects the assembled context on its way to the inference loop (crossing Ada).
2. Status / certainty
DESIGN-FIRST · ABSENT (no code). Role C4; implementation C1.
3. Language & location
TBD · new location e.g. src/morag/. ML toolchain (BERT classifier + LoRA adapters).
4. Does / does-not
- Does: classify/route (BERT), specialise (LoRA), assemble + inject context at cycle step 2.
- Does-not: police (D1 — and note modulators like LoRA cross Ada too, so a poisoned adapter meets Ada first); decide (C2/C4); hold the memory it draws from (E3).
5. Interface contract (proposed)
assemble(input, room_ns) -> injected_context(BERT route → select LoRA(s) → pull E3 RAG → pack).- Output crosses Ada (D1) before reaching the Brain; modulators (LoRA deltas) also pass D1's BBB.
6. Dependencies & stubs
E3 RAG (declarative source) — stub: fixed context. D1 (crossing) — stub: allow-all. Brain C4 consumes injection.
7. Invariants / laws
- L1 (C5): MoRAG injects, Ada polices — assembly is the organ's job, admission is Ada's.
- L2 (C5): modulators (steering vectors / LoRA) reach the Brain only by clearing Ada's BBB.
- L3 (C4): it is a mixture — BERT routes among specialised LoRAs.
8. Build steps
- Define the assemble contract. 2. Stand up a BERT router (classify situation → LoRA selection).
- Wire LoRA adapters + E3 retrieval. 4. Route output through D1 into step-2 enrich (C3).
9. Tests
Routing test (situation → expected LoRA); injection crosses a D1 stub; modulator blocked when provenance bad.
10. Open items
- Model/adapter choices + hosting (C1). LoRA adapter set (C1). Copyleft posture for ML deps.