Terrano09/AVID-EPJ-Baseline-Models-and-Data

AVID Baseline Architecture, Weighted-Bellman R2D2 World Model & Long-Horizon CSS Datasets (AVID_EPJ_Patches)

This repository hosts:

1. Long-Horizon Collected Quantum Syndrome & Non-Stationary Drift Datasets (datasets/):
   • datasets/avid_long_horizon_d3_250k_rounds.npz: 250,000 fast-loop syndrome rounds (50,000 rounds/regime across R1–R5).
   • datasets/avid_r2d2_replay_buffer_trajectories.npz: 600 non-stationary recurrent replay trajectories ($T=16$ windows $\times W=256$ rounds = 2,457,600 fast-loop rounds).
   • datasets/syndrome_packets_v1.bin & datasets/dataset_manifest.json.
2. Exact $2^m$-State Dynamic Programming Joint Syndrome Likelihood (AVID-Joint-DP-MAP) with Pair-Aware Logical-Triangle LLR Projection,
3. PyTorch Weighted-Bellman R2D2 (R2D2-Graph-RSSM) + Spatio-Temporal Causal Transformer + 5 Rank-4 Regime LoRA Adapters (.pt checkpoints in checkpoints/),
4. Complete End-to-End Mathematical Modeling Suite (docs/ and mathematical_model/, Chapters 00–08),
5. Quantized Signed 16-bit Q7.8 LLR Weight Banks (checkpoints/q78_weight_banks.npz) and Frozen Contracts (architecture/).

Shared-Input Logical Error Rate (LER) Benchmark Across 5 Non-Stationary Regimes (50,000 Rounds/Regime)

Noise Regime	B0 Static BP	B1 Marginal L-BFGS	B2 Observable PID	B3 Covariance	AVID Joint-DP MAP	Weighted-Bellman R2D2	B4 Oracle BP
R1_stationary_baseline	0.0486	0.0486	0.0486	0.0486	0.0486	0.0486	0.0486
R2_tls_hotspot_q4_q5	0.0564	0.0564	0.0564	0.0564	0.0588	0.0564	0.0564
R3_zz_crosstalk_burst	0.1984	0.1984	0.1984	0.1984	0.1143	0.1143	0.1143
R4_boundary_drift	0.0534	0.0534	0.0534	0.0534	0.0534	0.0534	0.0534
R5_combined_stress	0.0970	0.0970	0.0970	0.0970	0.0970	0.0544	0.0517

Included Model Checkpoints (checkpoints/)

• checkpoints/weighted_bellman_r2d2_backbone.pt: Trained PyTorch WeightedBellmanR2D2WorldModel (R2D2-Graph-RSSM) state dict.
• checkpoints/spatiotemporal_transformer_backbone.pt: Trained PyTorch SmallSpatioTemporalTransformer state dict.
• checkpoints/lora_adapter_R1..R5.pt: 5 Rank-4 Regime LoRA Adapter checkpoints.
• checkpoints/full_world_model_bundle.pt: Complete PyTorch checkpoint bundle + training curves.
• checkpoints/baseline_vs_avid_benchmark.json: Complete parameter recovery, Q7.8 weights, and LER metrics.
• checkpoints/q78_weight_banks.npz: Hardware-ready signed 16-bit Q7.8 register banks for weight_map_reg.sv.

Mathematical Modeling Folder (docs/ & mathematical_model/)

• 00_Mathematical_Modeling_Index.md through 08_Weighted_Bellman_R2D2_and_World_Model_Math.md