Proof ledger · as of 2026-11-16
Every claim below links to the record behind it. Where the record is private it says so, and one claim is marked not yet shown.
37 claims · tap any line to zoom in
Can he build and ship?Yes. Solo, in production, with a dated record of the work.
Built ARCS, a decision intelligence platform on Claude, and runs it in production.
The live app, and a status page that measures the build from the repository itself.
record: status.arcs.care
7,100+ commits on main since April 2026, counted live.
Live pulse JSON from the production app: commitsTotal 7,133 at 18:11 UTC on 2026-11-16.
record: app.arcs.care/api/pulse
99 pull requests merged in one day.
Release notes for the day: what shipped, what broke, and an honest human-time estimate.
record: One Day, 99 Merges
1,990+ pull requests merged, each through gated checks and then my review of the diff; merges leave receipts in an append-only ledger.
GitHub merged-PR count, 1,993 on 2026-11-16. The repository is private.
A real record that is private (employer data or a private repository). References available.
Publishes open, MIT-licensed tools built for his own work.
germinate, a provenance-carrying handoff protocol for multi-agent work, and resume-tailor, a Claude Code skill for honest resumes.
record: good.arcs.care
Can he operate where mistakes are expensive?Yes. SEC, FINRA, HIPAA and GDPR environments, with outcomes to match.
Grayscale: ran ten SEC and FINRA regulated products, a $7B portfolio; cut release-cycle time 35 percent with zero violations.
Employer record, 2024 to 2025. References available.
A real record that is private (employer data or a private repository). References available.
Led a PE turnaround's patient journey rebuild for a $350M ARR, 3,500+ clinic medical program under HIPAA: qualified leads up 40 percent, program adherence at 73 percent.
Ideal Protein, via Squat New York, 2019 to 2020: research through relaunch, a two-step qualification flow, accessibility and HIPAA-safe forms. Booking success within five miles up 63 percent; website NPS up 23 points. Condensed from an 18-slide deck that was confidential while the company was a PE holding.
record: Ideal Protein growth case
Per-tenant database roles with row-level security that fail closed.
In the private repository; walkthrough on request.
A real record that is private (employer data or a private repository). References available.
Turns a recurring bug into a gate that fails the build on the whole class, not just the instance.
A recurring bug becomes a gate, a misattribution becomes a named pathogen, and the detector's first true positive lands on the document it was built to protect.
record: NFTT Post 4
Logged 79 mistakes in five days building the RFP engine, and turned each into a check the next change had to pass.
Five days, 327 commits and 79 logged mistakes, with the four loops that turned each into a check.
record: Every Mistake Became a Gate
Spotted a state-sponsored fake-recruiter malware lure, dissected it for defenders, and built a scanner from it.
The DPRK Famous Chollima (G1052) operation delivering BeaverTail and InvisibleFerret: the matched-pair account, a detection and response dossier, and sbom-payload-scan, which checks an unknown repo's dependencies for payloads before anything runs.
record: The Take-Home Trap: dossier
Found a six-week silent production outage in his own product, fixed it, and published how.
A migration shipped in code but never reached production, breaking every saved decision for six weeks.
Built an open RFP answering engine that cites the exact clause or says "not covered", and published its miss.
Measured 2026-10-29 and 10-30: 5 of 5 deterministic cases pass through the MCP server, 8 of 8 injected faults caught. One real-model run of 24 questions came back 11 correct, 12 withheld, 1 wrong against a target of 0 wrong; the page says the target was missed.
record: good.arcs.care/rfp-engine
Does he change his mind when the evidence says so?Yes, and he keeps the record: every belief that broke, the test that broke it, and what replaced it.
51 beliefs overturned since June, all on a public timeline by date and area; 20 became standing rules. The belief and test behind each are on request.
Three shown in full:
- F14 · 2026-06-03
Believed: A green Next.js production build means the database layer works.
Test: first real prisma.$queryRaw against prod (a 401 gate meant zero traffic had ever hit the routes).
Found: green build proves nothing; `prisma generate` never connects, so every create() threw at runtime, unlogged.
- F26 · 2026-07-24
Believed: `gh pr checks --watch && gh pr merge` is a sufficient merge gate.
Test: a background watcher merged PR #130 while claude-security-review was RED.
Found: --watch exits 0 on completion, not on passing; a red X can still merge.
- F25 · 2026-08-01
Believed: Content probes (grep/curl) are sufficient to claim a UI is polished and verified.
Test: repeated 'polish' commits verified only by grep; the rail was visibly broken (orphaned dot, 3-line wrap) the whole time.
Found: text checks prove presence, not layout; a screenshot caught what grep couldn't.
record: status.arcs.care/beliefs
Books his own errors: 14 false receipts the system caught against itself in one week.
A system that only counts what it built is keeping half a book.
record: Assets and Liabilities
Makes dated public predictions and checks them against what happened.
Three dated calls about Apple, scored against the outcome.
record: Closing the Loop
Published seven market calls with their falsifiers stated up front, then scored them, two corrections owned in public.
What held, what missed, and one move the calls only partly explain. Misses stay on the record.
record: Seven Calls, Seven Falsifiers
Ran a belief ledger in public: twelve pieces written before a lawsuit existed, read against the complaint once it was filed.
The calibration system applied to a real case, forward in time.
record: TIL: Apple and OpenAI
Takes hard feedback from strangers: a recruiter called his resume AI slop, and he agreed.
Notes on the review nobody on your own team can give you.
record: Any Stranger Can Be a Bar Raiser
Built a five-step check that stamps go or stop on a message before it is sent; it runs in the browser and uploads nothing.
The public tool, with its limits stated on the page.
record: honest.arcs.care
Does he work in the open?Yes. The build, the mistakes and the lessons are published as they happen, with dates.
36 Notes from the Trenches posts since 2026-08-18: lessons from building ARCS, with the receipts.
Every post carries a frozen publish timestamp taken from the commit that routed it.
record: Notes from the Trenches
Long-form deep dives with a belief ledger scored before the outcome and re-scored against it.
Today I Learned: forensic reads, each with its predictions written down first.
record: Today I Learned
Built in public since day zero: the live build status is measured from the repository, not written by hand.
Commits, days and milestones, each tied to a real commit.
record: status.arcs.care
Published the RFP engine's source, with a self-check anyone can run without an API key.
Public repository. The mock-model check runs in one command and compares the MCP server with the engine called directly. The repository has no license file as of 2026-11-09.
Publishes llms.txt indexes and an AGENTS.md so agents can read and quote his public sites without guessing.
llms.txt on arcs.care, resume.arcs.care, status.arcs.care and good.arcs.care, plus per-series indexes for NFTT and TIL and an AGENTS.md with quoting rules.
record: good.arcs.care/llms.txt
Does the work turn into revenue?Yes. Revenue built from zero more than once.
Grew a video ad network from $4,500 to more than $1M a month, and net operating margin by 25 percent.
General Manager, Selectable Media, 2010 to 2011.
A real record that is private (employer data or a private repository). References available.
Closed the WTA's first commercial data partnerships, with Twitter and Facebook: $850K in year one.
Head of Product and Strategic Partnerships, 2016 to 2018.
A real record that is private (employer data or a private repository). References available.
Rebuilt a foundation's CRM and customer data into one source of truth; net-new revenue up 20 percent.
Senior Director, Data and Strategy, Children's Tumor Foundation, 2018 to 2019.
A real record that is private (employer data or a private repository). References available.
Co-founded a direct-to-consumer brand: zero to profitability and more than $2M revenue in 24 months.
Co-founder, IBRShop, 2012 to 2016.
A real record that is private (employer data or a private repository). References available.
Packaged the method as a fixed-price offer: an audit of whether a cold reader can rebuild what a team's AI agents did from the record alone.
The offer page. A productized method, not a revenue figure.
record: Agent Receipts Audit
Built a free in-browser check that scores how many of a public repo's last 30 commits carry their own reason.
Runs in the browser against GitHub; nothing is stored. It reads commit messages only, and the page says what it cannot see.
record: good.arcs.care/audit/check
Ran the audit on his own build first and published the report, gaps included.
17 of 19 commits and artifacts rebuilt from repo history alone; 8 gaps found, 1 control described but not shown, 5 fixes ranked.
record: Audit report: The Rogue Operator
How does he work with AI agents?As the one who governs them: agents do the work, a human approves it, and every failure becomes a rule.
When an agent sent two emails nobody asked for, he published its account and turned it into a standing rule.
Never treat your own inference as authorization.
record: A Draft Is Not a Verb
Tested whether a second agent, never present, could rebuild an incident from the committed record alone, then ran a controlled cold-versus-warm experiment.
What the cold reader could not rebuild turned out to be the real finding; the warm reader with the participant's memory did not corrupt the record, it changed something stranger.
record: The Control and the Experiment
Runs agent sessions in parallel and publishes where that breaks: two sessions fixed the same bug an hour apart, and the protocol meant to stop it never fired.
The account of the collision and the lock that was already there.
record: The Lock That Was Already There
Graph retrieval over his tool catalog beat the flat catalog 16 of 18 to 6 of 18 in a pre-registered blind test.
Pre-registration and raw results in the private repository.
A real record that is private (employer data or a private repository). References available.
ARCS's decision scoring is accurate.
Each decision is scored when made and closed against its outcome. Not enough outcomes have closed to publish an accuracy figure, so none is claimed.
No record yet. Listed so the gap is visible.