Upload README.md with huggingface_hub

README.md

@@ -4,21 +4,32 @@ tags:
   - time-series
   - foundation-model
   - temporal-causal
+  - causal-judgment
 license: mit
 ---
 
 # TCPFN — Temporal Causal Prior-Fitted Networks
 
-Zero-shot causal inference foundation model for temporal and industrial time series.
+A causal reasoning foundation model — predicts effects, judges trustworthiness, operates zero-shot.
+
+## 5 Novel Contributions
+1. **Temporal Token Design** — first PFN for temporal panel data
+2. **Causal Judgment Head** — null detection, regime classification, identifiability, confounding, mediation
+3. **Causal Regime Prior** — direct, confounded, mediated, feedback, IV, front-door
+4. **Self-Calibration** — auto-detects natural experiments in sensor data
+5. **End-to-End System** — discovery + estimation + judgment + RCA, all zero-shot
 
 ## Capabilities
-- **Causal Discovery**: AUROC 0.903 on Tennessee Eastman (zero-shot)
-- **Effect Estimation**: Temporal CATE trajectories
+- **Causal Discovery**: pairwise interventional CATE with judgment-aware edge scoring
+- **Effect Estimation**: temporal CATE trajectories with distributional output
+- **Causal Judgment**: null-effect detection (NullF1 up to 1.00), regime classification (4 classes, RegimeAcc up to 0.80)
 - **Root Cause Analysis**: 8-method ensemble (AERCA, ESD, ProRCA, GCM noise, ICC, Shapley, counterfactual, chain tracing)
 
-## Models
-- `models/temporal/final.pt` — Main model (257 MB, 200K steps)
-- `models/discovery/final.pt` — Discovery model (0.9 MB, 100K steps)
+## Model
+- `models/temporal/final.pt` — One model for everything (200K steps, curriculum-trained)
+- Causal Judgment Head: 5 trained outputs (null prob, confounding, identifiability, mediation, regime)
+- Training: mixed_prior (40% CausalTimePrior + 30% base + 30% CausalFM)
+- Curriculum: Phase 1 CATE-only → Phase 2 +Null → Phase 3 Full judgment
 
 ## Usage
 ```python
@@ -28,12 +39,16 @@ analyzer = TemporalCausalAnalyzer(
     temporal_model="models/temporal/final.pt",
 )
 report = analyzer.run("sensor_data.csv")
-print(report.edges)
-print(report.recommendations)
+print(report.edges)          # causal graph
+print(report.judgments)      # null prob, regime, identifiability per edge
+print(report.root_causes)   # ranked root causes
 ```
 
-## Benchmark Results
-See `benchmarks/discovery_results.json` for full results on 12 datasets.
+## Training Metrics (step 200K)
+- EffectLoss: 2.98 | JudgmentLoss: 2.72 (bounded)
+- NullDetectAcc: 1.00 | NullF1: 1.00 | NullAUROC: 1.00 | NullBrier: 0.02
+- RegimeAcc: 0.70 | RegimeMacroF1: 0.47
+- EffectRMSE: 0.077
 
 ## Paper
-Stalupula et al., "Temporal Causal Prior-Fitted Networks: Zero-Shot Causal Inference for Industrial Time Series"
+Stalupula et al., "Temporal Causal Prior-Fitted Networks: Zero-Shot Causal Inference and Judgment for Industrial Time Series"