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
43 lines
2.3 KiB
Markdown
43 lines
2.3 KiB
Markdown
# F2 — SAE monitor *(DESIGN-FIRST)*
|
|
|
|
## 1. Component
|
|
The interpretability monitor — a sparse-autoencoder watcher pointed at the **machinery (subagents)**,
|
|
**never the homonculus**. The outward-facing half of the central cut ("trust the center, verify outward").
|
|
Also the detector that fires the G1 stress loop.
|
|
|
|
## 2. Status / certainty
|
|
DESIGN-FIRST · ABSENT. Topology C5 (watches subagents, never Brain); detection mechanism C1/C2.
|
|
|
|
## 3. Language & location
|
|
TBD · new location e.g. `src/sae/`. ML interpretability (SAE over subagent activations).
|
|
|
|
## 4. Does / does-not
|
|
- **Does:** watch subagents (F3) + BERT/LoRA (they're AI systems); **detect agent outputs in opposition
|
|
to the top Eth-Int convictions** (A7), via the convictions' semantic-isomorphy tags → fire G1.
|
|
- **Does-not:** watch the Brain/scratchpad (forbidden by the central cut); correct cognition (it detects,
|
|
it doesn't edit — elimination breeds obfuscation, per self-doc B2).
|
|
|
|
## 5. Interface contract (proposed)
|
|
- `watch(subagent_activations) -> findings` (structure verification).
|
|
- `detect_opposition(outputs, top_convictions A7) -> opposition_signal` → handed to G1 (stress-loop).
|
|
- **Hard boundary:** no hook into the Brain (C4) — structurally impossible by construction.
|
|
|
|
## 6. Dependencies & stubs
|
|
A7 top convictions (with isomorphy tags) — *stub:* fixed conviction set. F3 subagents — *stub:* canned activations.
|
|
G1 consumes its signal — *stub:* print the signal.
|
|
|
|
## 7. Invariants / laws
|
|
- **L1 (C5):** SAE points at the machinery, **never the homonculus** (the Brain is the one thing nothing audits).
|
|
- **L2 (C4):** detection only — **no closed elimination loop** on cognition (judge the fruits at conduct, not the mind).
|
|
- **L3 (C2):** opposition = output semantically isomorphic to a *top conviction's antithesis* (A7 tags).
|
|
|
|
## 8. Build steps
|
|
1. Fix the no-Brain-hook boundary structurally. 2. Build subagent activation watching. 3. Build the
|
|
conviction-opposition detector (needs A7 isomorphy tags). 4. Wire the G1 stress emission.
|
|
|
|
## 9. Tests
|
|
Structural: no Brain hook exists. Opposition-detection: an output matching a top conviction's antithesis fires; aligned output doesn't.
|
|
|
|
## 10. Open items
|
|
- SAE training/target (C1). The **isomorphy match** (output ↔ conviction tag) mechanism (C2, shared A7/G1).
|