Profile Numbers Certified From Records

Aaron Kushner

Fifty years of accumulated judgment, the curiosity of a kid, a lifetime of scientifically-literate futurism — and the most documented human-AI collaboration on record.

Citizen Scientist Series · 2026-07-03 · aaron.m.kushner@gmail.com · akushnerphd.me

Every résumé now claims "AI-native." This one doesn't claim it — it produces the chain of custody. The numbers below are not self-reported; they are derived from git history and machine-generated session logs, and the underlying records are available for inspection. That is not an incidental flourish: the author's core research program is provenance over assertion, and this profile applies the standard to its own subject.

The receipts

MeasureValueProvenance
Full-time human-AI collaboration sinceMarch 30, 2026 (~14 weeks)First commit, portfolio repository
Commits (primary repository alone)910 across 78 distinct active days — ~82% of calendar daysgit history
Recorded AI collaboration turns~53,000 assistant turns (21,338 logged Apr 6–May 6; 31,792 logged Jun 1–Jul 3; mid-May under-counted due to log rotation)Machine-generated session ledgers
Model output generated~59 million tokens in June–July alone (≈ $35K/month at API list rates, sustained)Per-turn token ledger from raw session records
Session archive on disk0.57 GB of raw conversation records (surviving 30-day window)Session transcript corpus
Research artifacts published118 provenance-tagged papers (position papers, field reports, specs, pre-registered experiments)Artifact metadata census
Active project repositories48, spanning ed-tech, provenance infrastructure, publishing, CAD, biomedical devicesPortfolio directory census

Methodology: all figures computed 2026-07-03 from timestamps in version control and machine-written logs — no human estimation. Turn counts are a floor: session transcripts auto-purge after ~30 days, so windows outside surviving logs are bounded by git activity rather than counted. Dollar figures are list-price throughput equivalents (a work-volume measure), not billing claims. The full audit trail — ledgers, hashes, session archives — is available on request.

The arc

The preparation started early: a childhood and adolescence immersed in the futurism of scientifically literate creatives — the science fiction canon read not as escapism but as engineering forecast. Then a career accumulating the things AI cannot supply: judgment, domain range, a proven record of lateral, non-linear problem-solving across engineering, research, and business. When GPT-4 made AI a tool instead of a novelty, the watching stopped and the immersion began. Since March 2026 that immersion has been effectively full-time — at an intensity the session ledgers above put, by simple arithmetic of hours, in the extreme tail of anyone, of any age, working with these tools.

The result is a working demonstration of a thesis: the highest-leverage AI users may not be the youngest — they are the ones with the most accumulated wisdom to multiply. One person, three months, roughly $500 in marginal tooling costs: a working K-12 integrity product (TeacherAware — sanctioned writing workspace, tamper-evident provenance chain, escalating privacy-preserving proctor), a live author-certification platform (AuthorAware), a citizen-scientist publishing exchange with adversarial multi-model peer review, a specced adaptive-education engine (open-ed), and 118 published research artifacts documenting all of it.

The Research Harness

Underneath the portfolio is the thing that makes its pace possible — and the thing that cannot be faked: a research harness built around Claude Code through months of daily use, hitting every pain point the tool has, and closing each one with versioned behavioral infrastructure. Persistent cross-session memory with a deliberated-position ledger. Multi-instance orchestration — persistent orchestrator seats, scoped worker spawning, crash recovery with full conversational continuity. Behavioral enforcement gates that make verification, provenance, and commit discipline structural rather than aspirational. Tamper-evident certification of research artifacts, reviewed adversarially by competitor models with no stake in the outcome. In current form: 13 enforcement hooks, 27 skills, 149 scripts, 190 memory files — every line of it version-controlled, timestamped, and self-documenting.

The harness is its own credential. It cannot be built quickly, cheaply, or from the outside — the only path to it is long, painstaking work inside the tool. And it is powerful enough that its author's position is deliberately conservative: before any public release, it should be discussed with Anthropic first — red-teamed, its force-multiplication weighed against its misuse surface. The prototype exists. The conversation is the ask.

Why this profile exists

Agentic AI is at its command-line moment: world-changing capability, unusable by most of humanity. Somebody builds the Mac. The author has spent fourteen weeks building the prototype of that shell — the behavioral layers, the provenance spine, the products that ride on it — and documenting the entire process at a fidelity no one else has attempted. This profile accompanies a set of proposals to Anthropic: a K-12 partnership ready to demo, an education-foundation concept, and a responsible-disclosure conversation about the harness itself. The endpoints differ; the credential is the same: look at the record.