Tech Demos

Syndrome-mitigation QEC (VEC)

Independent re-implementation of arXiv:2610.12400 (Google Quantum AI, found via @zlatko_minev): reweight shots by MWPM-vs-exact-ML decoder disagreement sectors to cancel residual logical errors in post-processing. Pre-registered d=3/5 experiment at p=0.05, seeds 0/1/2.

• VEC cuts residual logical bias 32–152× below plain MWPM and 15–101× below postselection on every seed; ⟨Z_L⟩ returns to 1.000±0.001 (ideal +1)
• PARTIAL: overhead ratio η<4 holds at d=5 (2.4) and d=7 (1.73, near the paper's quoted weights) but not d=3 (5.4–5.7) — matching the paper's distance trend
• Exact d=5/7 ML via transfer-matrix contraction verified against brute force; breakpoints: empty Y sector at p=0.01, threshold collapse, drift, 0.31%-error syndrome-only noise learning

cd apps/syndrome-mitigation-qec
uv venv .venv && source .venv/bin/activate
uv pip install -r requirements.txt
./run_all.sh

Correction Training 2D toy

Pedagogical re-implementation of sunovivid/correction-training (arXiv:2610.14136): a stop-gradient corrector iteratively refines the denoiser's prediction at each sampling step. Pre-registered matched-NFE K sweep on 8 Gaussians, seeds 0/1/2.

• K=0 baseline beats every K∈{1,2,4,8} on MMD² on all three seeds (0.84 vs 3.99 ×10³ at best K, seed 0) — the paper's ImageNet gain inverts on a near-optimal 2D denoiser with no exposure-bias gap to close
• Breakpoints: ordering holds to K=64 and across NFE 6–240; the one flip is 4 well-separated modes, where K=2 wins. Trains in ~5 s/seed on CPU
• MIT upstream; no checkpoints downloaded. Prepared (unexecuted) Colab for real SiT-REPA ImageNet sampling — GPU spend needs owner approval

cd apps/correction-training-2d
uv pip install --index-url https://download.pytorch.org/whl/cpu torch
uv pip install -r requirements.txt
./run_all.sh

LoMa local feature matcher

Demo of davnords/LoMa (arXiv:2604.04931): LightGlue-style matching with LoMa-B plus rotation-invariant LoMa-R. CPU clear-experiment with pre-registered latency, determinism and rotation breakpoints.

• LoMa-B match latency about 8.3–8.7 s mean per upstream pair on 4 CPU threads, under the 30 s budget
• Correspondence counts drift by a few matches across seeds 0/1/2 (strict identity fails). LoMa-R beats LoMa-B on inlier ratio at 45–90°, but not at every angle ≥ 45° (fails at 135°)
• MIT code; matcher inherits Apache-2.0 from LightGlue. Weights auto-download; not committed

cd apps/loma-matcher
uv pip install -r requirements.txt
uv pip install "lomatch @ git+https://github.com/davnords/LoMa.git"
./run_all.sh

llama.cpp Decision Models

Web UI for Decision Models in llama.cpp: send a state plus typed questions (choice / score / noul) to POST /v1/systemone and see option probabilities as bars.

• Mock mode is on by default (keyword heuristics, clearly labelled), so it runs without a GPU or server
• Point it at a real server, e.g. Julia-1 (144M, Apache-2.0) via llama serve

cd apps/llamacpp-decision-models
bun install && bun run dev

EmbeddingGemma 2

Semantic search with google/embeddinggemma-2: text to 768-d cosine similarity over a tiny local corpus, with a FastAPI server and a Vite UI that ranks matches with scores.

• Loads the model text-only through Sentence Transformers (vision and audio encoders dropped)
• If the model can't download, it falls back to a hash embedder that the UI labels MOCK

cd apps/embeddinggemma-2
pip install -r requirements.txt
python server/app.py

# always safe: no GPU, no weights
$ python3 scripts/dry_run.py

# ~1.9 MB Synthetic Causal Dynamics checkpoint
$ python3 scripts/run_light_inference.py --download

# or open in Colab
notebooks/jepa_anything_demo.ipynb

JEPA-Anything

Kit for Gen-Verse/JEPA-Anything (arXiv:2609.20800), domain-agnostic predictive world models via Orthogonal Predictive Factorization. Ships a Colab notebook plus local scripts, not a training stack.

• Dry run is the guaranteed path; it does not claim to reproduce paper benchmarks
• Optional smoke inference with the lightest domain checkpoint from Hugging Face

$ python3 scripts/smoke_test.py
{
  "task": "image-vlm",
  "status": "graceful-skip",
  "cuda_available": false,
  "PIXELUMM_CKPT": { "set": false }
}
# expected without a GPU and weights

PixelUMM

Kit for nv-tlabs PixelUMM, encoder-free unified image/video understanding and generation in pixel space (Qwen3-8B backbone). Targets the image-understanding task (--task image-vlm).

• Code is Apache-2.0; the checkpoint is under the NVIDIA One-Way Noncommercial License
• Full runs need a high-end NVIDIA GPU, CUDA 13 and the weights; guardrails are left on by default

Active borylation

Demo of Minot et al., Nature Computational Science (2026) (ETH/Roche): active learning and geometric GNNs for Ir-catalysed C–H borylation. Shows the released EquiformerV2 checkpoint's per-atom regioselectivity on CPU, and a retrospective BALD vs random acquisition run.

• EquiformerV2: top-1 site matches the product for 3/3 upstream examples (all three are also in its training set)
• BALD vs random (XGBoost ensemble, 3 seeds): BALD needed more molecules to reach the target MCC, so the hypothesis failed
• This app is GPL-3.0 (from minotm/active-drug-discovery); data CC-BY-4.0

Bone suppression for chest X-rays

Demo of Qure.ai's qureaiorg/bone-suppression (arXiv:2609.24937): two TorchScript models split a frontal chest X-ray into bone, soft tissue and lung component. It runs on CPU and includes a pre-registered stability test and breakpoint sweep.

• Reproduces the authors' published outputs. CPU latency was about 10 s per 1024 px image for both models, under the 15 s budget
• Stability under σ=0.01 noise: mean bone SSIM 0.942–0.945, just below the 0.95 bar. The decomposition breaks on inverted, JPEG q≤10, 64 px or high-contrast-clipped input
• This app is CC BY-NC-SA 4.0 (weights' terms). Not a medical device and not for diagnosis

cd apps/bone-suppression-cxr
pip install -r requirements.txt
./run_all.sh