DELTA: Dual Edge-Linked Transformer Architecture¶
A research implementation of the DELTA architecture — building toward The Brain: a system that dynamically constructs its own relational graphs and reasons over them, without pre-defined topology or transformer scaffolding.
Core Thesis¶
Reality is a graph. Language is a lossy compression of reality into sequences. Transformers pay a quadratic tax to rediscover structure. DELTA operates on relational structure directly — and now constructs that structure from scratch.
The three-paradigm gap: GNN edges are passive scalar wires; DELTA edges are first-class computational citizens that attend to each other. That edge-to-edge attention produces the Phase 28 +24% noise robustness gap — and Phase 39's self-bootstrapped DELTA proves the system can build its own graph without any transformer scaffold.
Current evidence base: 63 completed experiment phases, 44 unit tests, competitive link prediction on FB15k-237, multi-hop compositional reasoning dominance, validated differentiable graph construction via BrainEncoder (Phases 55–58), and scaling evaluation with subsampling ablation at N=5000 (Phases 59–63).
Quick Stats¶
| Metric | Value | Phase |
|---|---|---|
| Experiment phases | 63 complete | 1–63 |
| Unit tests | 44/44 passing | — |
| Best LP MRR (DELTA-Full, temp-tuned) | 0.4905 | 52 |
| Brain LP MRR (self-constructed graph) | 0.4818 with H@10 0.8076 | 57 |
| Multi-hop champion | DELTA-Matched 3p MRR 0.742 +/- 0.009 | 45 (3-seed) |
| Depth scaling | 5p MRR 0.790 vs GraphGPS 0.690 | 44 |
| Self-bootstrap breakthrough | 0.757 +/- 0.041 — 157% of FixedChain | 39 |
| Inference speed | Per-query 0.8–0.9x GraphGPS (faster) | 45 |
| Noise robustness | +24% over vanilla GNN at 80% corruption | 28 |
Documentation¶
Start Here¶
- DELTA, Explained for a Software Engineer — The 30-second version, the core argument, layered evidence, and why this matters beyond the immediate result
Architecture¶
- Architecture Overview — Components, self-bootstrap, development timeline, backward compatibility
- Visual Explainer — Interactive three-paradigm comparison (Transformer -> GNN -> DELTA)
- The Brain: End Goal — Long-term vision, capacity paradox, roadmap horizons
Results¶
- Key Findings — 44 findings organized by research stage
- Validation Phases — All phase result tables (Phases 1–63)
Planning¶
- Status & Roadmap — What's validated, open gaps, roadmap, publication pathway
- Research Methodology — AI assistance disclosure
Setup¶
- Getting Started — Installation, experiment commands, cloud GPU setup
- Project Structure — Repository directory layout
Requirements¶
- Python 3.10+
- PyTorch 2.0+ (GPU:
pip install torch --index-url https://download.pytorch.org/whl/cu124) - CPU sufficient for Phases 1-24; GPU (6 GB+ VRAM) recommended for Phase 25+
DELTA architecture — conceived March 25, 2026. 63 experiment phases, 6 architectural fixes, 44 unit tests. KG scaling evaluation complete (Phases 59–63). Pivoting to sparse attention and sequence domains. See The Brain for the long-term vision.