Yaroslav Vasylenko

Autonomous research engineer — I build evidence-led, deterministic systems for noisy, adversarial environments.

_{neuroscience  ·  AI  ·  financial systems  ·  independent  ·  Poltava, Ukraine 🇺🇦}

Every repo here is an active line of research — not a portfolio piece. Work ships only after climbing hypothesis → theory → fact, with both confirmed and rejected results kept on record.

I study the biological brain — not to simulate it, but to extract its operating principles and encode them into verifiable artifacts.

The brain is a prediction machine. Its core loop is perceive → compress → infer → act → verify → close.

I build the same loop into software: agents for neurons, CI for synaptic weight, adversarial verification for the dendritic vote that gates a spike.

Real experiments with verified results and public failures. Negative results are tagged, not deleted. Pre-registration is on, fail-closed reporting is on. Every claim has a SHA-256-anchored ledger entry; two runs of the same code produce the same hash.

1. Hypothesis  — state the prediction and the gate it must pass.
2. Theory  — minimal falsification harness: in-sample effect · out-of-sample time-split · sanity-check against literature ceilings.
3. Fact  — binary verdict. Phase-2 OOS below gate → REJECT. No maybe-later.

If Phase-2 OOS collapses more than 3× versus Phase-1 in-sample, that alone rejects. The temporal split is the load-bearing discipline.

Adversarial orchestration.  Creator → Critic → Auditor → Verifier. Each layer is an independent witness of its own error. witness ≠ actor, even inside one system.

Zero tech debt contract.  A task is not closed while any of these is true: lint/types/tests not green · TODO/FIXME/dead code left · docstrings or tests missing on changed code · CI red on the PR · physics/brand/invariant guards not passing.

V1  reversibility × blast_radius      # gate before every action            weight = 1.00
V2  signal / (signal + noise)         # compress without losing meaning     weight = 0.95
V3  numerical_verified(claim)         # quantitative → prove, or mark HYP    weight = 0.90
V4  adversarial_survived(hypothesis)  # PARCH-FALSIFY-001 before commit      weight = 0.85
V5  context_integrity(session)        # flag drift, never inherit silently   weight = 0.80
V6  genuine_execution(task)           # unhelpfulness is never "safe"        weight = 0.75

ΔV > 0   ∧   dΔV/dt ≠ 0     —     invariant YV1

Thirteen public repositories across four substrates — each an active line of research, with negative results kept on record.

📈 quant & physics

🔬 verification, integrity & falsification

🧠 neuroscience & mechanistic models

🌐 AI systems & cognitive architecture

This profile holds itself to the discipline it claims. The factual claims are gated, not asserted — tests/test_profile.py decomposes the page's integrity into five invariant classes, and each class defends exactly one promise:

python -m pytest        # run every class yourself

A green badge above is not a claim — it is a re-runnable measurement.

Owner neuron7xLab (user) · last 30d · generated 2026-06-22T03:11:15.164197+00:00 · source: gh_api + local CMAR scan