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:
253
architecture/formalization/Attention-Slumber-Prediction-Cycle.md
Normal file
253
architecture/formalization/Attention-Slumber-Prediction-Cycle.md
Normal 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.**
|
||||
|
||||
🧬⚡🔱💎🔥
|
||||
|
||||
Reference in New Issue
Block a user