Audit the declarations a computational result carries, and check they are still true. Producer freshness, counter coherence, provenance pins, partial runs. Read-only by construction.
Causal Inference for Multiomics
Repository for the 3rd paper of my PhD.
Source-available IX-Function causal-transfer harness: tests cross-domain causal structure reuse through pre-outcome prediction, reality-delta scoring, uncertainty preservation, falsification, negative controls, model review, donor handoffs, and bounded Wave 6 review gates. No AGI claim.
R package for negative control analysis.
De novo cosmetic peptide design from the skin ECM degradome, scored against composition-preserving nulls
Evaluation integrity layer for RAG/LLM golden sets: content-addressed suites that refuse unsound comparisons, negative controls that must go red, human-calibrated item analysis, and a release gate that refuses thresholds it cannot resolve.
Empirical AI-in-Education study of wrong-answer-conditioned BM25 retrieval on frozen SciQ, with shuffled lexical controls, question-block uncertainty, and a transparent misconception-aware RAG prototype.
Knowledge or sparsity? A controlled test of gene set masks in an interpretable single-cell VAE.
The same sequence filter scores AUC 0.684 against naive controls but only 0.569 against composition-preserving scrambles, and recall at a matched control rate falls from 27.4% to 6.3%. The choice of control decides how good the filter looks.
Make research claims earn their way out of a pipeline: claims registry, fail-closed release gates, blind locks, negative controls
Reproducible analysis and matched-negative-control framework for testing molecular specificity in electrophysiology-to-transcript prediction using Allen Patch-seq data.
A tiered reference set of 797 candidate non-interacting drug pairs for benchmarking drug–drug interaction prediction, with FAERS co-exposure denominators, statistical-power tiers, and a reproducibility harness.
Small checks that prove they can fail. A guardrail nobody has seen fire is theater — so every tool ships negative controls and a --selftest that runs them.
Property-matched decoys cut a DHFR screen's property-only AUC from 0.913 to 0.841 - not to 0.5. Matching reduces decoy bias without removing it.
Evidence-centric AI governance tools: PASS / ACT / CT-safe evaluation harness, specs, and examples (public subset).
What a benchmark scores with nothing in it: content-free members against each task's own majority baseline, plus a judged arm and a judge-disagreement control. Preregistered. First sweep done, 1 of 20 tasks flagged outright.
To associate your repository with the negative-controls topic, visit your repo's landing page and select "manage topics."