Projects

A coordinate map from 44 signaling modules over 154 normal cell types (single-cell RNA, no cancer data); the malignant cells of 25 cancers converge onto one normal-cell region — the placental cytotrophoblast, a normal invasive state — contracting program space 6.3×. Within matched tumor–normal pairs, malignancy deepens the state ~6× (colorectal 43/46, paired P<10⁻¹²; kidney+liver 77/86; absent in high-baseline lung).

44 signaling modules treated as a cell's perceptual repertoire; 9-PC eigenspace from 154 Human Protein Atlas cell types; capacity-floor predictor; validity scorecard with three null controls. The toolkit behind the cancer-convergence finding above.

Memory consolidation, read through 44 perceptome modules across 7 datasets with 26 pre-registered predictions, resolves into a three-layer architecture — perception (senses the trigger), sentinel (senses proteostatic load), construction (builds the machinery). Region-specific (hippocampus builds new synapses; amygdala swaps receptors against a 1.8× baseline reserve). TLR9 knockout selectively removes the ER-chaperone arm (−60–75%) and impairs memory — direct causal evidence. 28/32 published perturbations predicted, 0 contradictions.

157 genes across 14 oscillatory pathways classified by temporal role only. Rise-phase genes map to oncogenes, recovery-phase to tumor suppressors (OR = 27.5, p = 3.6 × 10⁻⁹). Predicted inversions in growth-inhibitory pathways (p53, TGF-β) confirmed.

Algorithmic extraction of 128 NFLs from 159 KEGG signaling networks. NFL genes show 59-fold CGC enrichment over non-NFL same-pathway genes (p = 9 × 10⁻⁴⁴). Irreversible Authority metric predicts cancer gene fraction with Spearman ρ = 0.83.

Multi-phase pre-registered dissection of whether a perpetually-circling triad of coupled networks is functional, not decorative — on a Stuart–Landau triad and a trained GRU triad. A trainable triad self-organizes a coupling-maintained 120° rotating limit cycle (5/5 seeds, self-sustaining with the loss off); it is NOT a memory store (the coupling that holds the cycle suppresses capacity — founding hypothesis refuted); it IS a phase regulator carrying a Hopf-universal temporal code (normalized iPRC λ-invariant to CV ~3×10⁻⁵), emergent in the trained GRU (shape-match 0.999). The deepest law: an oscillation's existence is cheap (conservation / loss / blind flow) but its directed rotation — broken time-reversal symmetry — is the one irreducibly-active ingredient; on a dissipative map, smooth oscillation requires a complex eigen-pair.

935 ablation experiments from 161 ML publications. The DFE of architectural modifications matches biological DFEs; architectural diversification follows logistic dynamics with punctuated equilibria; 14 traits independently invented 3–5 times in parallel with biological convergences.

1,584 ablations × 8 checkpoints × Pythia 410M. The DFE shape evolves from delta-peak-with-outliers to heavy-tailed Student's t (df ≈ 2.3). 39% of eventually-critical heads emerge from noise; only 4% born-critical. L8H9 single-head phase-transition specialization between training steps 4k and 8k.

Cross-model pairwise head-ablation interactions: 78% of significant top-30 pairs in Pythia 410M show synthetic-lethal-like ε > 0 in loss space. The epistasis transition co-locates with the DFE crystallization at training steps 512–1,000, rejecting lottery-ticket emergence.

Biological counterpart to the ML epistasis study above: does the synthetic-lethal/redundancy regime found in Pythia 410M heads have a developmental analog in cell differentiation? Tests the prediction on Schiebinger 2019 reprogramming scRNA-seq, with Norman 2019 perturb-seq as a true-perturbation comparator. Same statistical apparatus across substrates.

412,730 recordings across three continents (PTB-XL, Chapman-Shaoxing, CODE-15). β is diagnostic of cardiac conduction anatomy: CLBBB vs CRBBB AUC ≈ 0.98 across all three populations. Aging flattens the spectrum; disease steepens it. Honest null on mortality.

Three-level variance decomposition on GTEx v8 + Tabula Muris Senis + Calico rat + macaque atlas. Tissue identity is preserved across 40 years of aging; the signature is systemic noise, not selective accumulation. Caloric restriction acts as a noise filter, not a structure restorer.

Peak-trough asymmetry of EEG oscillations in 215 (LEMON) + 608 (Dortmund Vital Study) adults, with 208-subject 5-year longitudinal follow-up. β-band asymmetry decreases with age; effect exceeds classical alpha slowing. Theta spatial entropy predicts memory independent of age.

Russian-language adaptation of the DAT (Olson et al., 2021). 21,159 submissions; Cronbach's α = 0.899; split-half reliability 0.696; no measurable practice effect. Live instrument at mool32.github.io/dat-ru.

Russian-language interactive experiment ("Ты не случаен") in the BrainyBara line of educational / playful experiments. Live at mool32.github.io/you-are-not-random.

A 4-player repeated non-cooperative game with incomplete information, stochastic shocks, oil-price dynamics, nuclear-escalation logic, and economic phase transitions. 500 Monte Carlo simulations; approximate Nash equilibrium via best-response dynamics. Built as content / teaching material on game theory, not as a forecast — interpretation discipline named explicitly in the README. Live demo at mool32.github.io/game_theory_models.

Telegram bot delivering short supportive messages four times daily; community-contributed phrases with moderation. Built during a difficult period in Israel ("matsav tov" — "good state" — is the inversion of the wartime-news framing "matsav"). The first public version of an ongoing family of small social-support bots; other variants are kept private for now.

(Gini, eff_N) framework applied jointly to mouse cell-type compositions (Tabula Muris Senis, Calico CR atlas) and to Pythia-410M head-importance during training. Same quadrant; CR rescues 64% of the Gini drift in rat bone marrow. A substrate-independent compositional signature, narrower than full DFE universality.

Three studies on whether the wisdom of crowds survives consultation of a small set of LLMs. GPT-4o, Claude, Gemini show pairwise error correlation r = 0.78 on 568 Metaculus questions; LLMs inherited human biases rather than introducing novel ones. A methodological observation about collective AI behavior at scale.

18 metrics tested across 5 independent EEG datasets, 126 subjects total. No metric replicates across datasets; the strongest within-subject signal (DERCo wPLI alpha CV) reverses direction on ZuCo. A universal zero-calibration EEG comprehension detector does not exist with current metrics.

Can an autoencoder's fitting process sense a biological axis it was never shown? Pre-registered (A∧B∧C∧D, leave-donors-out) on immune aging (GSE233321, 167 donors). The latent never beats matched linear PCA at recovering age (0/2 across seeds × dims); training erodes the signal; a synthetic positive control localises the exact regime where the method would win — high linear rank, low intrinsic dimension — which single-cell data does not occupy. The unsupervised latent independently re-derives the perceptome eigenspace (CCA p=0.001) but not its aging axis.

Canonical document covering project lifecycle (scaffolding → pilot → full run under locked preregistration), sign-convention locks, repository structure conventions, README templates, code discipline, manuscript discipline, AI-friendly publishing (llms.txt, robots.txt, JSON-LD, served BibTeX), and reference exemplars. 17 sections; the epistasis pair (ML side + biology side) is the worked example throughout.

A metric that moves between conditions is not yet a finding — it may be dropout, library size, a batch effect, or a factorial confound. Eight gates (mathematical independence under simulated dropout, factorial QC parity, n_genes matching, raw-data visibility, stratified controls, cross-dataset replication, plus two judgment gates the tool elicits) probe a black-box `metric(data) -> float` and stop at the first blocking failure. Built from three of my own failures, each of which survived weeks of work before a short QC check killed it. Ships as a pip package, a Claude Code skill, and an MCP server; 35 tests. Zenodo [10.5281/zenodo.21195679](https://doi.org/10.5281/zenodo.21195679).

Pre-registered retrodictive-calibration pilot: given a paper's abstract stripped of technical content, can an LLM predict its multi-year scientific impact substantially better than citation-velocity baseline? Primary metric: Spearman ρ between LLM score and composite Impact, with a +0.05 margin over baseline required for go/no-go. Pre-registration v1.0 hash-locked 2026-05-01 (SHA-256 in HASH.md); stop condition binary and pre-committed. Pilot 1A: NLP corpus + Opus 4.7 + Pythia 6.9B (contamination control). Stage-gate to Pilot 1B contingent on 1A passing. First concrete artifact applying the framework end-to-end.