The Field Journal

The writing, by volume

Everything written for the lab, gathered in one place. A field journal keeps drought years and growth years in the same grain — so this one does too. Two volumes so far: one learned to ship, one is the lab itself.

Vol. II

The Lab

2026 — building in public. In progress.

Sep 2026Field NoteNew

Digits Know Who Wrote Them

Seventeen domain-blind detectors, pointed at earthquakes, balance sheets, twenty thousand stars and two planet catalogues. One of them can tell whether a human typed the numbers.

Sep 2026Field NoteNew

Same Hand

Five strangers built quietfail. Eighty-six built claude-pet. A man named Ben built my groundskeeper. None of them met. Then 200 cold AI sessions across eight model versions showed how: the model draws coined names from a short menu, and the descriptive names are ours.

Sep 2026Field NoteNew

Two Ways to a Number Nobody Had

The OEIS publishes its own thin places. Most are thin because they are hard — the famous short entries stop at standing records. We found one that was short because nobody had been back, pushed it further, and computed every new term two independent ways. Wiseman’s 2020 conjecture holds for every one of them. Still unproven — the solver is in the post.

Sep 2026Career ToolsNew

Career Compass, Step by Step

A free, local-first job-search assistant that runs inside Claude — your facts in plain files on your own disk, a résumé rendered from them per posting, and interview prep built around the questions they will ask and the ones you should ask back. One illustrated page: the five steps, the interview prep, and the read-only dashboard. Fact-checked against the 2.9.0 release; no usage figures, because none are measured.

Jul 2026Assay

A Proof You Can Check Yourself

Another AI-claimed proof — but this one's elementary, checkable by hand. Erdős Check re-derived both halves of a crack at Erdős 793 — both re-derivations from one model family, so one lens split two ways, not two independent looks — found it holds up with one cosmetic gap, and marks exactly where to aim your own skepticism. Not “verified” — that word stays reserved for the kernel.

Jul 2026Assay

Grading the Machine's Homework

OpenAI claims a 3-page proof of a 50-year-old conjecture. Erdős's whole thesis — the dangerous moment is promotion, not the idea — just had its first live-fire test. Pinned, read, executed on 409 graphs, kernel-checked, and graded: corroborated, not promoted.

Jun 2026Verification

The First Verified Node

Erdős Check, earned its first kernel-checked “verified” stamp — R(3,3) > 5, proved in Lean with no axioms. The result is settled math from 1955. The novelty is the tool, and the honesty gate that makes one word mean one thing.

Jun 2026Method

About the Lab

Why a one-human, agent-augmented studio names itself for a broken branch — the place the rings finally show. The dendrochronology behind the name, carried the whole way down.

Jun 2026Craft

Portfolio in a Box

The argument for a portfolio site that sounds like a person, plus the actual box: a downloadable starter with the voice baked in and the identity swappable. Taste as a starting point, not a cage.

Vol. I

Learning to Ship

Oct 2025 – Mar 2026 · written during the ten-month gap — a senior PM learning to build, in public.

Mar 2026Career Data

Growth Rings — 10 Months of Job Searching

203 applications, 76 rejection emails, 34 ghosts, 2,710 GitHub contributions, 6 products live as of March 2026. A data-driven cross-section of ten months in the 2025–26 market — the funnel, the ghosts, and the parallel build sprint that changed the trajectory.

Feb 2026AI Infrastructure

The Productive Compute Framework

Self-sustaining AI infrastructure for global public good: converting idle compute into verified outcomes through outcome-based funding. Whitepaper, draft v2.0.

Jan 2026AI + ProductLinkedIn

If You Can Read a Recipe, You Can Now Be a Developer

The $1K Experiment, Part 2: what happens when the framework compounds. 5.5 hours to a working MVP; 2,031 lines became 106,000. Shipping is addictive — here's the warning label.

Dec 2025AI + ProductLinkedIn

The $1K Claude Code Credit: What Happens When a PM Learns to Ship

Could a senior PM with product clarity but no coding background actually build and ship real software? 31 days, 215 commits, 38K lines of TypeScript, and the 64/33/3 collaboration model that made it work.

Oct 2025Enterprise AILinkedIn

The 90-Day Death Spiral: Why 95% of AI Projects Fail

Only ~5% of AI pilots deliver measurable impact. The early-warning system hiding in your support tickets — and the metrics that predict failure before day 90.

The journal publishes in volumes, not on a clock — roughly a season apart. Subscribe ↗ to catch the next one.