metadata
license: mit
language:
- en
tags:
- welfare
- forecasting
- time-series
- pytorch
- ethics
- phi-humanity
library_name: pytorch
pipeline_tag: time-series-forecasting
datasets:
- synthetic
metrics:
- mse
PhiForecasterGPU — Welfare Trajectory Forecaster
A CNN+LSTM+Attention model that forecasts Phi(humanity) welfare trajectories — predicting how 8 ethical constructs (care, compassion, joy, purpose, empathy, love, protection, truth) evolve over time.
Model Description
PhiForecasterGPU takes a 50-timestep window of 36 welfare signal features and predicts the next 10 timesteps for both the aggregate Phi score and all 8 individual constructs.
| Property | Value |
|---|---|
| Architecture | CNN1D (2-layer) → Stacked LSTM (2-layer) → Additive Attention → Dual heads |
| Parameters | ~1.3M |
| Input | 50 timesteps × 36 features |
| Output | 10-step Phi forecast + 10-step × 8 construct forecast + attention weights |
| Formula | Phi v2.1 (recovery-aware floors) |
| Training | 100 epochs, 44,800 sequences from 8 scenario types × 50 seeds |
| Best Val Loss | 0.000196 MSE |
Architecture
Input (50 × 36)
↓
CNN1D (2 conv layers, kernel=3, hidden=256)
↓
Stacked LSTM (2 layers, hidden=256)
↓
Additive Attention (query-key-value)
↓
┌───────────────┬────────────────────┐
│ Phi Head │ Construct Head │
│ (MLP → 10) │ (MLP → 10 × 8) │
└───────────────┴────────────────────┘
36 Input Features
- 8 raw constructs: c, kappa, j, p, eps, lam_L, lam_P, xi
- 8 volatility signals: rolling 20-step standard deviation per construct
- 8 momentum signals: 10-step price momentum per construct
- 5 synergy signals: geometric mean of construct pairs (care×love, compassion×protection, joy×purpose, empathy×truth, love×truth)
- 5 divergence signals: squared difference of construct pairs
- phi: aggregate welfare score
- dphi_dt: first derivative of phi
Graph-Enhanced Mode (43 features)
When trained with graph_features_enabled=True, the model accepts 43 features (36 base + 7 graph topology features from the detective knowledge graph):
| Feature | Description |
|---|---|
| graph_density | Edge density of the full graph |
| entity_pagerank | PageRank centrality of focal entity |
| entity_degree | Normalized degree of focal entity |
| entity_clustering | Clustering coefficient |
| community_size | Fraction of graph in same community |
| avg_neighbor_conf | Mean edge confidence to neighbors |
| hub_score | HITS hub score (normalized) |
The Phi(humanity) Formula
Phi is a rigorous ethical-affective objective function grounded in care ethics (hooks 2000), capability theory (Sen 1999), and Ubuntu philosophy:
Phi(humanity) = f(lam_L) · [prod(x_tilde_i ^ w_i)] · Psi_ubuntu · (1 - Psi_penalty)
- f(lam_L): Community solidarity multiplier
- x_tilde_i: Recovery-aware effective inputs (below-floor constructs get community-mediated recovery)
- w_i: Inverse-deprivation weights (Rawlsian maximin)
- Psi_ubuntu: Relational synergy term + curiosity coupling (love × truth)
- Psi_penalty: Structural distortion penalty for mismatched construct pairs
Training Scenarios
| Scenario | Description |
|---|---|
| stable_community | Baseline — all constructs ~0.5 |
| capitalism_suppresses_love | Love declines as purpose erodes |
| surveillance_state | Truth rises while love collapses |
| willful_ignorance | Love rises while truth collapses |
| recovery_arc | Collapse then community-led recovery |
| sudden_crisis | Compassion + protection crash mid-scenario |
| slow_decay | Gradual institutional erosion |
| random_walk | Correlated random walks with mean-reversion |
Usage
import torch
from huggingface_hub import hf_hub_download
# Download checkpoint
path = hf_hub_download("crichalchemist/phi-forecaster", "phi_forecaster_best.pt")
# Load model (see spaces/maninagarden/model.py for class definitions)
from model import PhiForecasterGPU
model = PhiForecasterGPU(input_size=36, hidden_size=256, n_layers=2, pred_len=10)
model.load_state_dict(torch.load(path, map_location="cpu", weights_only=True))
model.eval()
# Inference: (batch, seq_len=50, features=36) -> phi, constructs, attention
X = torch.randn(1, 50, 36)
phi_pred, construct_pred, attn = model(X)
# phi_pred: (1, 10, 1) — next 10 Phi values
# construct_pred: (1, 10, 8) — next 10 values for each construct
# attn: (1, 50) — attention weights over input window
Live Demo
Try it at crichalchemist/maninagarden — interactive scenario explorer, custom forecasting, and experiment comparison.
Citation
This model implements the Phi(humanity) welfare function formalized in the detective-llm project. Key theoretical foundations:
- hooks, b. (2000). All About Love: New Visions. William Morrow.
- Sen, A. (1999). Development as Freedom. Oxford University Press.
- Fricker, M. (2007). Epistemic Injustice. Oxford University Press.
- Metz, T. (2007). Toward an African moral theory. Journal of Political Philosophy.
Limitations
- Trained on synthetic data — real-world calibration pending
- Phi is a diagnostic tool, not an optimization target (Goodhart's Law)
- 8-construct taxonomy is Western-situated; requires adaptation for diverse philosophical frameworks
- Recovery-aware floors add theoretical richness but the model has not yet been retrained to fully leverage them