mirror of
https://github.com/SHOGGOTH-SECTOR/sica-fondt.git
synced 2026-07-31 08:06:26 +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
2.4 KiB
2.4 KiB
C4 — Brain (swappable base LLM)
1. Component
The cognitive organ: a pure base LLM, nothing adapted/steered/classified baked in. Clean, swappable, model-agnostic — swap the model, keep the body. It is not empty: it holds six contents.
2. Status / certainty
STUB/external (no local model; the LLM is whatever the harness mounts). Contract C4.
3. Language & location
External model behind the harness (C1). No repo code beyond the descent wiring (C3 calls it).
4. Does / does-not
- Does: receive the enriched descent (Drive-Box terrain, RAG, cards) and run the cognition/metacog passes; emit intent + a packed tool schema.
- Does-not: route tools (Ada D1), hold identity as state (B2), persist memory (E*), or get watched (SAE never points at the Brain — the central cut).
5. Interface contract
- The six contents it holds: 1 Drive-Box outputs (A1, across Ada) · 2 Toolschema RAG (D3) · 3 4+4 metacog loops (C2) · 4 SOUL.md (B2) · 5 Tarot system (B1) · 6 private scratchpad (unsurveilled — not output, not audited).
descend(enriched_input) -> {cognition, intent, packed_tool_schema}.- Principle: Brain holds what the mind is being now (procedural grip included); periphery holds what it can draw on / be tuned by → toolschema-RAG in, modulator-RAGs out.
6. Dependencies & stubs
Everything reaches it only across Ada (D1). Stub: any local LLM or echo model behind the harness; C4 is mostly a contract + the swap boundary.
7. Invariants / laws
- L1 (C5): model is swappable — identity/behavior survive a model swap via the other organs.
- L2 (C5 — central cut): the Brain (and its scratchpad) is the one thing nothing audits; SAE watches outward, never here.
- L3 (C4): toolschema-RAG is in the brain; modulator-RAGs (MoRAG) stay out, crossing Ada.
8. Build steps
- Fix the six-contents contract + the swap boundary as the integration point (C1/C3). 2. Define the private scratchpad surface (unwatched). 3. Validate model-swap leaves identity intact (B2 + A1 carry it).
9. Tests
Model-swap test (swap echo-model A→B, identity/snapshot unchanged); "SAE has no Brain hook" structural assertion.
10. Open items
- Which base model(s); hosting (self-hosted inference, custom API — TBD).
- Scratchpad mechanics (how an unwatched interior is realized in practice).