Project Structure

DELTA/
|-- delta/                    Core library
|   |-- graph.py              DeltaGraph + sparse COO multi-hop edge adjacency
|   |-- attention.py          Node, edge, dual parallel attention (return_weights)
|   |-- router.py             PostAttentionPruner + LearnedAttentionDropout
|   |-- memory.py             Variational bottleneck tiered memory (hot/warm/cold)
|   |-- partition.py          BFS seed-expansion partitioning O(N+E)
|   |-- reconciliation.py     Node-edge co-update (ReconciliationBridge)
|   |-- constructor.py        Transformer-based graph bootstrap
|   |-- brain.py              BrainEncoder + BrainConstructor (differentiable graph construction)
|   |-- model.py              Full DELTA model
|   |-- baselines.py          GraphGPS + GRIT implementations
|   |-- datasets.py           Dataset loading (load_lp_data for FB15k-237)
|   +-- utils.py              Helpers, synthetic data, benchmark generators
|
|-- experiments/              Phase-by-phase validation (63 phases)
|   |-- phase1-15             Core architecture validation
|   |-- phase16-21            Architectural fix benchmarks
|   |-- phase22-25            Scale & integration on GPU
|   |-- phase26-30            Roadmap validation + multi-seed
|   |-- phase31-37            H100 / Colab experiments
|   |-- phase38-40            Differentiable constructor, self-bootstrap, correct LP
|   |-- phase41-45            Multi-hop compositional reasoning + inference timing
|   |-- phase46-54            Attention temperature optimization + multi-seed validation
|   |-- phase55-58            Brain architecture + multi-seed density validation
|   +-- phase59-63            Scaling & depth management (N=2000, N=5000, subsampling ablation)
|
|-- notebooks/                Colab-ready infrastructure
|   +-- delta_colab_ready.py
|
|-- tests/                    Unit tests (44/44 passing)
|   |-- test_graph.py
|   |-- test_attention.py
|   |-- test_router.py
|   |-- test_memory.py
|   |-- test_utils.py
|   +-- test_baselines.py
|
|-- docs/                     Phase result documentation
|   |-- phase_55.md           Brain architecture port results
|   |-- phase_56.md           Constructor density ablation results
|   |-- phase_57.md           Brain temperature annealing results
|   |-- phase_58.md           Multi-seed density validation results
|   |-- phase_59.md           Depth scaling at N=2000 results
|   |-- phase_60.md           Residual gating results
|   |-- phase_61.md           DELTA vs DistMult controlled comparison
|   |-- phase_62.md           Scale to N=5000 results
|   |-- phase_63.md           E_adj subsampling ablation results
|   +-- PUBLICATION_ROADMAP.md  NeurIPS/ICLR publication strategy
|
|-- mkdocs-src/               Documentation source (this site)
|   |-- index.md              Home page
|   |-- architecture.md       Architecture + bootstrap + timeline + compat
|   |-- ARCHITECTURE_VISUAL.md  Interactive three-paradigm visual
|   |-- the-brain.md          Vision + capacity paradox + roadmap
|   |-- key-findings.md       44 key findings by stage
|   |-- validation-phases.md  Complete phase result tables
|   |-- status-and-roadmap.md Status + gaps + publication pathway
|   |-- research-methodology.md  AI assistance disclosure
|   |-- setup-and-running.md  Setup + commands + cloud GPU
|   +-- project-structure.md  This page
|
|-- requirements.txt
|-- mkdocs.yml
+-- README.md

---

Core Library (delta/)

File	Purpose
graph.py	DeltaGraph data structure, sparse COO multi-hop edge adjacency, edge adjacency caching, vectorized incidence matrix
attention.py	NodeAttention, EdgeAttention, DualParallelAttention, ReconciliationBridge — all with return_weights support
router.py	PostAttentionPruner (soft sigmoid gating), LearnedAttentionDropout, legacy ImportanceRouter
memory.py	Tiered memory (hot/warm/cold) with variational bottleneck and KL regularization
partition.py	BFS seed-expansion partitioning in O(N+E) with importance-aware seeding
reconciliation.py	Node-edge co-update (ReconciliationBridge)
constructor.py	Transformer-based graph bootstrap with per-layer edge projections
brain.py	BrainEncoder + BrainConstructor — differentiable Gumbel-sigmoid edge selection for self-constructed graphs
model.py	DELTAModel — full pipeline: constructor -> partition -> dual attention -> pruner -> reconciliation -> memory
baselines.py	GraphGPSModel (2022) and GRITModel (2023) implementations for comparison
datasets.py	Dataset loading utilities; load_lp_data() for FB15k-237 experiments
utils.py	Synthetic data generators, benchmark tasks, helper functions

Baselines (delta/baselines.py)

Model	Reference	Params (typical)
GraphGPS	Rampasek et al. (2022)	~228K
GRIT	Ma et al. (2023)	~197K
CompGCN	Vashishth et al. (2020)	Via utils.py
TransE	Bordes et al. (2013)	Via utils.py
RotatE	Sun et al. (2019)	Via utils.py
DistMult	Yang et al. (2015)	~47K