Colophon

This site is built with Hugo (v0.157.0 extended). Plain HTML and CSS on standard pages; the newsletter form uses minimal inline JavaScript for spam protection (timestamp plus honeypot) and submits to a TypeScript handler on Cloudflare Workers, with Cloudflare D1 for subscriber state and Resend for double-opt-in confirmation. Atkinson Hyperlegible Next is the body and heading face, served self-hosted as woff2. Static hosting on NearlyFreeSpeech.

Authorship

AI-assisted throughout, in both drafting and revision, under a documented styleguide that prohibits em-dashes, throat-clearing, marketing voice, and self-applied “honest” framings. Each piece passes a final styleguide gate before publish; quantitative claims are verified against source artifacts in the dev tree. Final responsibility for everything published here is mine.

Sources for the quantified counter

Selective Brier is a calibration metric where 0.0 is perfect and 1.0 is worst; it evaluates only the predictions the system committed to, excluding correct abstentions. Crowd-level forecasting accuracy sits around 0.15; superforecaster baseline is around 0.10. See the methodology writeup for the full framing.

Counts on the homepage are verified against source artifacts on the date stamped at the bottom of the counter. Sources:

• Skills authored. Counted as the number of real skill directories containing a SKILL.md that I wrote, across my global skills root (~/.claude/skills/) and my WordPress workspace root. Verified 2026-07-26: 85 are mine; 81 of those declare a test-contract block — the public claim is stated as “over 80 with test contracts” so it stays accurate as the corpus grows. Symlinks to third-party skills (Obsidian/canvas tooling and similar) and installed scaffolds written by others (SpecKit) sit alongside and are not counted. Employer-context skills are excluded from the count entirely. Two are public with their test-contracts intact at github.com/shawnmcb/claude-a11y-skills (released 2026-07-08), as partial public sourcing for this count.

• NetterTech Events test count. From the NetterTech Events PHPUnit configuration at the time of writing. Aggregate of unit, integration, and feature tests across the five-plugin suite (Events free core, Pro, Migrator, Rentals, Seating); 10,200+ test methods total, verified 2026-05-15 via grep of ^\s*public function test|@test across plugin test directories. Free core component contributes 4,994 test methods; the official WP plugin scan returns no issues.

• Veridi gaming countermeasures. Thirteen countermeasures documented in the production methodology (thirteenth added 2026-05-19; verified 2026-07-03): Confidence Laundering, Citogenesis, UNVERIFIABLE-by-Design, Preprint Pump-and-Dump, Selective Skepticism Exploitation, Tier Inflation, Framing Manipulation, Coordinated Legitimate Sourcing, Anchoring, Data Disappearance Exploitation, Institutional Capture, Substrate Self-Reference, Warm-up-then-Defect. Methodology files in the Veridi repository.

• Veridi adversarial test suite. Thirty-one claims targeting the documented vectors with a target detection rate of at least 85% (verified 2026-07-03 against the suite files). Includes substrate-self-reference and indirect-prompt-injection scenarios mapped to NIST AI RMF subcategories with MITRE ATLAS technique IDs.

• Pragma + Praxis combined benchmark. 93 of 95 combined claims pass under current methodology versions (Pragma v1.6: 55 / 55; Praxis v1.4: 38 / 40), per the focused regression dated 2026-05-26, up from 87 / 95 at the v1.0 baseline. The two residual Praxis partials are root-caused pathway-ranking cases with a methodology-refinement ticket open.

• Veridi calibration set composition. 100 claims: the original 95-claim validation set (Smoke Tests 1, Golden Test Sets A through C 70, Adversarial Test Suites v1 and v2 24) plus the 5-claim GTS-D Wave 1 targeted extension dated 2026-05-04. GTS-D adds boundary-category coverage on recent-ambiguity intent attribution, non-US legal status, manipulated-media provenance, hazard-versus-risk framing, and policy-label compression. Selective Brier on the 100-row set is 0.0253 (95% CI [0.0193, 0.0320]) on 89 committed verdicts; the original-95-only figure was 0.0251.

• Language coverage. The Veridi app supports English and French as user-facing languages; the verification pipeline itself runs in English. Non-English source coverage is research-stage across five language groups (Chinese / Mandarin, Spanish / Portuguese, Arabic, Japanese / Korean, Hindi / Urdu); in-language verification is not yet shipped.

• Design system tenure. 3+ years at Anaplan as the sole continuous contributor on the internal design system. Became frontend lead when the founding engineer moved on. Team composition rotated across the span; I was the constant.

• Accessibility track. Twenty years of WCAG-aligned work across NMDP (HIPAA-regulated), nonprofit, and enterprise contexts. From employment history; specific WCAG 2.2 AA compliance verified for current shipped work.

If a count looks stale, it probably is: the verified-on date in the counter footnote shows last update.