mirror of
https://github.com/allaunthefox/Research-Stack.git
synced 2026-07-31 03:05:21 +00:00
67 lines
2.3 KiB
Markdown
67 lines
2.3 KiB
Markdown
# Hutter Prize & Signal Policy: Testing Specification
|
|
|
|
This document consolidates the original prompts and logic used to formalize the Research Stack's compression validation and dynamic routing.
|
|
|
|
**Last Updated:** 2026-04-27
|
|
**Status:** Testing specification reference
|
|
|
|
## 1. Hutter Prize: Minimal Validation Path
|
|
To get usable data for the pipeline without synthetic artifacts:
|
|
|
|
### Setup
|
|
1. **Clone Dataset**: `git clone https://huggingface.co/datasets/leloy/enwik8`
|
|
2. **Pull Shard**: `git lfs pull --include="test-00000-of-00002.parquet"`
|
|
3. **Extraction**:
|
|
```python
|
|
import pyarrow.parquet as pq
|
|
table = pq.read_table("test-00000-of-00002.parquet")
|
|
rows = table["input_ids"]
|
|
windows = rows[:1000] # each window is 128 bytes
|
|
data = [list(w) for w in windows]
|
|
```
|
|
|
|
### Sanity Metrics
|
|
- `cells_per_window` ≈ 64 (pack bytes → 4-bit Cell)
|
|
- `distinct_cells` << `total_cells`
|
|
- `distinct_signatures` << `distinct_cells`
|
|
- `repeated_patches` > 0
|
|
- `admissible_ratio` ~ 1.0
|
|
|
|
---
|
|
|
|
## 2. Signal Policy Logic
|
|
Make signal data first-class as policy input without letting it touch truth.
|
|
|
|
### Dynamic Transition Law
|
|
$$S_{t+1} = apply(route(sig(S_t), signalBand_t, priority_t))$$
|
|
|
|
### Core Implementation (Lean 4)
|
|
```lean
|
|
namespace ExtensionScaffold.Compression.SignalPolicy
|
|
|
|
inductive SignalBand where
|
|
| quiet | active | stressed | extreme
|
|
|
|
def SignalBand.weight : SignalBand -> Float
|
|
| .quiet => 0.0 | .active => 1.0 | .stressed => 2.0 | .extreme => 3.0
|
|
|
|
def policyOfBand : SignalBand -> SignalPolicy
|
|
| .quiet => { exploreBias := 0.0, tunnelBias := 0.0, promoteBias := 1.0, gossipBias := 0.0 }
|
|
| .active => { exploreBias := 0.5, tunnelBias := 0.5, promoteBias := 0.5, gossipBias := 0.5 }
|
|
| .stressed => { exploreBias := 1.0, tunnelBias := 1.0, promoteBias := -0.5, gossipBias := 1.0 }
|
|
| .extreme => { exploreBias := 1.5, tunnelBias := 1.5, promoteBias := -1.0, gossipBias := 0.5 }
|
|
|
|
-- Resulting Logic:
|
|
-- Signal becomes a band -> Band changes policy -> Policy changes routing/promotion/budget.
|
|
-- State and Admissibility stay untouched.
|
|
```
|
|
|
|
## 3. Comparative Test Plan
|
|
Run baseline routing on Hutter shard stats vs. signal-biased routing. Compare:
|
|
1. **Route Replacement Rate**
|
|
2. **Promotion Count**
|
|
3. **Tunnel Count**
|
|
4. **Admissibility Stability**
|
|
|
|
---
|
|
*Portable Reference - Gemini CLI*
|