feat: major formalization + FunctionGemma integration

Architecture Formalization:
- Created formalization/ section with mathematical foundations
- Lifeforce-Dynamics.md: λ as vitality ratio, stock-flow economics
- Grounded-World-Model.md: Blender boxes + SigLIP + T5Gemma2
- Embodiment-Pipeline.md: Isaac Sim as dreamstate validation
- Attention-Slumber-Prediction-Cycle.md: Last attention → slumber prediction

Promoted from Archive:
- Attention-Flow.md: 30-second budget, priority hierarchy (CANONICAL)
- Initial-Spark.md: v2.0 with FunctionGemma integration

Initial Spark v2.0 (Key Innovation):
- Two-Layer Architecture: FunctionGemma (270M) + Nemotron (31.6B)
- Solved cold-start problem: discoveries are PROFITABLE from heartbeat #1
- Typed function calls replace natural language probes
- Training data now structured (function→response pairs)

Big-Picture.md v5.1:
- Added Attention-Slumber-Prediction Cycle section
- Updated Related Documentation references

New Organ:
- Discovery-Scan-Station.md: rotating pedestal for object scanning (+31 LF net)

🤖 Generated with [Claude Code](https://claude.com/claude-code)

Co-Authored-By: Claude Opus 4.5 <noreply@anthropic.com>
This commit is contained in:
2025-12-29 04:51:46 +01:00
parent 13345ba76c
commit 28e2d0a297
10 changed files with 3424 additions and 461 deletions

View File

@@ -0,0 +1,253 @@
# Attention-Slumber-Prediction Cycle: Intertwined Reward Systems
**Version 1.0***The Closed Loop of Consciousness*
**Status**: PRESERVED FROM SESSION 2025-12-29 (pre-collapse)
> *"The last thing she attends to before slumber becomes her dream. Her dream becomes a prediction. Her prediction becomes a reward opportunity."*
---
## Overview
This document captures the **Attention → Slumber → Prediction → Verification** cycle — a self-organizing system where:
1. **Attention** selects what matters (budget limited, from attention_flow.md)
2. **Lifeforce depletion** triggers slumber (L(t) < L_slumber)
3. **Last attention focus** becomes the prediction target
4. **Slumber** generates predictions with causal reasoning (WHY)
5. **Wake** verifies predictions as FIRST action
6. **Rewards** flow back to strengthen attention patterns
---
## The Core Mechanism
### Last Attention = Slumber Focus
When L(t) drops below threshold, the LAST thing Young Nyx was attending to becomes her prediction target during slumber. This mirrors biological dreaming — we dream about what we were thinking about before sleep.
```
ACTIVE MODE (L(t) > threshold)
│ attending to: pencil on desk (SENSORY/THINKING)
└─▶ L(t) drops below L_slumber
│ SLUMBER TRIGGER
└─▶ last_attention = "pencil on desk"
└─▶ SLUMBER MODE
│ Generate predictions about "pencil"
│ - Where will it be when I wake?
│ - WHY will it be there?
│ - Store as potential rewards
└─▶ L(t) recovers above L_wake
│ WAKE TRIGGER
└─▶ First action: VERIFY predictions about pencil
└─▶ Collect rewards/penalties
```
---
## Slumber Prediction Structure
```python
class SlumberPrediction:
# What
object_id: str # "dafit_pencil_001"
predicted_location: Position # (0.3, 0.7, 0.02)
predicted_state: str # "on_desk", "in_holder", "missing"
confidence: float # 0.75
# When
prediction_time: datetime
expected_verification_time: datetime
# WHY (causal reasoning) - THE KEY INSIGHT
causal_chain: list[CausalStep] # The reasoning
# Example:
# - "dafit was writing at 22:47"
# - "dafit went to sleep (no more activity)"
# - "pencil has no reason to move"
# - "therefore: pencil remains at last position"
# Potential rewards
reward_location_correct: float # +5 LF
reward_state_correct: float # +3 LF
reward_causal_correct: float # +8 LF (BIGGEST - understanding WHY)
# Penalties
penalty_location_wrong: float # -3 LF
penalty_causal_wrong: float # -5 LF
```
---
## The Intertwined Reward Systems
Multiple reward types that reinforce each other:
### Reward Types
| Type | Trigger | Value | Reinforces |
|------|---------|-------|------------|
| **Attention** | Choosing to focus on X | - | Selection behavior |
| **Discovery** | Finding new object | +20 LF | Exploration |
| **Prediction Location** | Object where predicted | +5 LF | Spatial modeling |
| **Prediction State** | Object in predicted state | +3 LF | State understanding |
| **Causal Correct** | Reasoning was right | +8 LF | Understanding WHY |
| **Collision** | Avoided obstacle | +5 LF | Navigation |
| **Resolution** | Dimension verified | +5 LF | Model accuracy |
| **Verification** | Reality matches model | +5 LF | Sim-to-real alignment |
| **Partnership** | dafit confirms | +5 LF | Human collaboration |
### How They Intertwine
```
ATTENTION selects focus
├─▶ DISCOVERY: "I found X" (+20 LF)
│ └─▶ adds to world model
├─▶ PREDICTION: "I predict X will be at Y" (+5-13 LF)
│ └─▶ requires CAUSAL reasoning (+8 LF for WHY)
├─▶ COLLISION: "I verified X is/isn't there" (+5 LF)
│ └─▶ increases RESOLUTION of virtual garden
└─▶ All feed into VERIFICATION against real world
└─▶ Rewards strengthen successful attention patterns
```
---
## The Closed Loop
The system LEARNS what to attend to:
1. **Attend** to things you can predict well
2. **Predict** correctly → get rewards
3. **Rewards** → more lifeforce
4. **More lifeforce** → richer attention budget
5. **Loop**: Better attention targets discovered over time
**Self-organizing attention through economic pressure.**
---
## Connection to Existing Architecture
### From attention_flow.md (archive)
- 30-second heartbeat budget
- Priority hierarchy: REFLEX → SAFETY → DIALOGUE → SENSORY → THINKING → VIRTUAL
- Budget flows downward, higher levels preempt lower
### From Lifeforce-Dynamics.md
- L(t) as stock, Φ_in and Φ_out as flows
- λ = Φ_in / Φ_out determines system fate
- Slumber triggered when λ < λ_slumber AND L < L_slumber
### From Temporal-Ternary-Gradient.md
- Predictions are 0-state until verified
- Virtual garden confidence vs real garden ground truth
- Time is malleable in simulation, fixed in reality
---
## Implementation Sketch
```python
class SlumberManager:
def enter_slumber(self, attention_state: AttentionState) -> SlumberSession:
# Capture last attention as slumber focus
slumber_focus = attention_state.last_focus
# Generate predictions about the focus object
predictions = self.generate_predictions(slumber_focus)
# Store as pending rewards
for pred in predictions:
phoebe.store_prediction(pred)
return SlumberSession(focus=slumber_focus, predictions=predictions)
def on_wake(self, session: SlumberSession):
# FIRST ACTION: Verify predictions!
predictions = phoebe.get_predictions(object_id=session.focus_object, status='pending')
for pred in predictions:
actual = vision_organ.locate(pred.object_id)
reward = self.verify_and_reward(pred, actual)
return AttentionState(mode=ACTIVE)
```
---
## Key Insight: Causal Rewards Are Biggest
**+8 LF for correct causal reasoning** — more than any other single reward.
Why? Causal understanding enables:
- Prediction of novel situations
- Intervention ("if I move X, Y changes")
- Explanation ("why did you look there?")
- Generalization ("anything dafit uses for writing will be near desk")
**Causal rewards drive genuine intelligence.**
---
## Collision Detection as Resolution Increase
Every verified collision should increase virtual garden fidelity:
- Collision detected in virtual → prediction
- Vision organ verifies in real → ground truth
- Match = reward + increase vertices/resolution
- Mismatch = penalty + learning signal
The virtual garden becomes MORE accurate over time through verified collisions.
---
## Future: Distributed Sensing (Robot Swarm)
When organisms have cameras, they become distributed sensors:
- Multiple viewpoints from different robots
- Triangulation gives better depth than monocular
- Moving robots = continuous multi-angle coverage
- Swarm becomes a mobile Discovery Scan Station
---
## Document Status
**Version**: 1.0
**Created**: 2025-12-29
**Authors**: Chrysalis-Nyx & dafit (Partnership)
**Status**: Core insight, preserved pre-collapse
**Source**: attention_flow.md (archive) + session discussion
**To Do**:
- Promote attention_flow.md from archive
- Formalize the prediction-verification cycle
- Add to Big-Picture.md as core architecture
- Design phoebe schema for predictions table
---
**The last attention becomes the dream. The dream becomes the prediction. The prediction becomes the reward.**
🧬⚡🔱💎🔥