Operator's manual for thirteen working systems built at the intersection of all three, by a medical student and military officer using frontier AI as the engineering tool. 150,906 lines across 247 files, every system live, every one measured from your browser as you read this.
↑ hover a world to tint the page · click to filter the systems · press / for the index
Purpose and scope
This manual documents five months of learning and building in spare hours. These projects were made possible by tools that lowered the technical barrier between an idea and working software. Others will bring different experience and build better things.
Everything listed is deployed and public. Chapter 2 does not take that on faith: when this page loads, your browser pings every host in Figure 1-1 and records what came back. Chapter 3 lists the systems by domain, Chapter 4 collects the written doctrine, Chapter 5 the smaller equipment. Appendix A is the change record for the manual itself.
Every system was specified by the operator and written with Claude, then verified the way a clinician verifies a lab: against a source, by a second reader, and against the live deployment rather than the local copy. Medical content carries its citations. Scoring tables cite the governing publication. Where a source deck was wrong, the question was corrected and the correction logged.
Study tools on this site are for exam preparation and are not medical advice. Clinical calculators and field references are teaching aids; confirm with the governing guideline before use on a patient.
Readiness
Table 2-1. Portfolio readiness. Static figures are recounted at every edition; live figures come from Table 2-2.
| Measure | Result | Standard |
|---|---|---|
| Lines of code | 150,906 | unique source files, generated data excluded; 21,761 of these are one project's test suite |
| Source files | 247 | html, js, css, python, typescript |
| Board-style questions | 12,972 | 7 question banks, 8 subjects, every item independently reviewed, most a second time |
| Players & teams tracked | 1,106 | sports intel across 12 data sources |
| Systems reachable | … of 17 | measured from your browser, now |
| Median round trip | … | milliseconds, edge-cached static hosts; the one server-rendered app is excluded while it wakes |
| Months since first deploy | 5 | April to September 2026, evenings and weekends |
Table 2-2. System status. Same probe as Figure 1-1, in list form.
| System | Host | Domain | Status | Round trip |
|---|
The probe is an opaque cross-origin fetch, so it can only report that a host answered, not that the answer was the right page. Cloudflare Pages serves a soft 200 for unknown paths. Treat green as "the lights are on" and click through for anything that matters.
Systems
Step 2 CK QbankRebuilt
Public study site over seven Anki-derived banks: 12,972 board-style questions and 12,357 guide facts across eight subjects, a UWorld-style test engine with tutor and exam modes, per-system performance, and three drill tools. Built by a chunk, author, independent-review, adversarial-triage pipeline that caught and corrected about 100 factual errors in the commercial source decks.
12,972 questions8 subjects2-3 review passes LiveOpen → Medicine · QIFig. 3-2Patient Handoff Tracker
Full-stack I-PASS/SBAR handoff system built as an emergency medicine QI project: named-receiver read-back that only the receiver can acknowledge and verify, action items, Joint Commission compliance metrics with CSV export, and HIPAA-conscious audit logging over synthetic patients. Flask and SQLite on a free tier, so the first load wakes the server.
22 API endpoints8 DB tablesFlask Live · ~30s cold startOpen → Medicine · CardiologyFig. 3-3Read This
Pattern-interpretation trainer: 28 procedurally generated rhythm strips at 25 mm/s with calipers and a live monitor, a fixed 8-step read of 14 twelve-lead cases with MI localization by territory, and an outside-in chest film read over synthesized radiographs and real films. Every case sourced.
28 rhythms8 read steps100% sourced LiveOpen → Medicine × GameFig. 3-4MICRO TD
Tower defense where the body is the map: 16 guideline cases as levels, 34 real drugs as towers against 25 organisms, damage set by coverage and site penetration, resistance bred by antibiotic pressure, and organ toxicity, plus a one-patient Ward Rounds mode. TypeScript and Canvas, every fact sourced, 104 tests.
16 guideline cases34 drugs25 organisms Live · best on desktopPlay → MilitaryFig. 3-5Warrior Ready
All-branch PT scorecard at 2026 standards: six services scored from their official charts with the source cited, per-service body composition screens, run pacer with lap splits, ruck planner, and score history. The build refuses to ship unless 72 spot checks read off the printed charts pass.
6 branches72 chart spot checks0 deps LiveOpen → Machines · AI SafetyFig. 3-6AGI Capability Tracker
Frontier capability ledger across 10 sectors, 57 models and dozens of benchmarks: a weighted composite score with its method published, verified ARC-AGI cells, a capability-milestones register, a capability-against-control safety ledger, and a gated-tier toggle for what the labs hold back. Updated September 2026.
57 models10 sectors5 milestones LiveOpen → Machines × MedicineFig. 3-7Simulators
Four sims, one umbrella: the Alignment Simulator (run a frontier lab, 12 decisions, 6 endings, built from the Risk Stack paper), Code Blue (spacebar CPR with live rate and depth feedback, AHA 2020), Half-Life Tracker (caffeine, hydration, and BAC pharmacokinetics), and Pocket Telescope (live night-sky ID and star quizzes).
4 sims24 decisions11 endings LiveRun → Sports · Fantasy FootballFig. 3-8Edge Draft Terminal
Draft intelligence terminal for a 12-team half-PPR league: an 862-player board valued on a weighted seven-source composite and read against the Sleeper room's market price, weekly projections priced as points above replacement, Monte Carlo pick-path simulation in a worker, an in-season wire of injuries and market movers, and read-only replay of a live Sleeper draft room. Static PWA with CI-gated deploys and a test suite that also runs against production.
862 players7 sources21,761 lines of tests LiveDraft → Sports · HoopsFig. 3-9NCAAT Bracket
Archive of the finished 2026 men's tournament: all 67 games from a frozen ESPN snapshot, every first-edition model graded against the results, and an interactive bracket that scores your own picks. A sourced fact-check corrected 531 first-edition claims, all listed on the page.
67 games4 models graded531 corrections LiveOpen → Sports · GolfFig. 3-10Masters 2026 Hub
Archive of the 2026 Masters as a hand-operated leaderboard: the final board for all 91 players with hole-by-hole cards from a frozen ESPN snapshot, an 18-hole yardage-book course guide, a searchable field, the champions' roll, and a Remotion film rendered from code. 137 corrections to the first edition are listed.
91 players18 holes600 frames LiveOpen → Sports · DetroitFig. 3-11Detroit Sports Hub
Hub board for the Lions, Tigers, Pistons and Red Wings plus three trackers: a new Lions app (live down and distance on a drawn field, drive charts, win probability), Tigers with pitch-by-pitch tracking, Statcast and a 30-prospect farm system, and the Pistons' 14-68 to East #1 seed run with injury watch and a 2004 tribute.
3 trackers29 tabs0 deps LiveOpen → Humanities · Literature & PhilosophyFig. 3-12Great Books
Continuous-zoom library of 46 works across existentialism, political order, liberty, machines, knowledge, and reality.
46 books5 shelves4 depthsLiveRead → Medicine · EducationFig. 3-13AdmissionsAcademy
Founded by Paul Clancy to expand access to medical education: a free Pre-Med Essentials course, MCAT strategy, and application guidance for students from any background. Sean serves as Chief Technology Officer.
Free premed course1-on-1 guidanceCTO Sean's roleLiveVisit → Music · FilmFig. 3-14LAST LIGHT
Soundgarden's Black Hole Sun, taken literally. When the Sun collapses, its light keeps coming for 8 minutes and 20 seconds. A grunge song counts them down, sung three ways.
3 takes2:19 runtime33 sourced factsLiveWatch →Doctrine
The Frontier AI Risk Stack
Frontier AI risk is best understood not as a single category but as a stack of six interacting layers. The most serious dangers arise from the way failures compound across layers: a structured analytical tool for policymakers, developers, and civil society.
Three Roads From Here
The three positions on frontier AI (safety through development, acceleration, pause) characterized on their strongest terms, with existential and suffering risk treated as separate axes that do not always move together. The central frame: the camps are reading three different payoff matrices for the same game.
Two Stories About the Beginning
The Big Bang and a generated origin are the same kind of story: a starting condition plus a rule, with the evidence only ever reaching the rule. Four objections (tidiness, cheating, motive, proving too much) are taken seriously; two of them force the second story to change, and it survives as a testable rival rather than an evasion.
Reading Room
An annotated card catalogue of 64 AI papers, X threads and blog posts collected through April 2026, each with a technical summary, a plain-language version and a note on why it matters.
papers · tweets · blogsEnter → 19 essays · 6 sectionsFig. 4-2Philosophy Room
One thesis: every great question converges on alignment, whether we can build minds that want what we want. Cosmology, consciousness, AI, identity, theology and game theory, each question read twice: as the classical problem and through the AI lens.
~30 min readEnter →Ancillary equipment
Change record
This site is rebuilt from the ground up at every new Anthropic model class, as a personal benchmark of frontier capability. Each edition is designed end to end by the model it is named for, then retired, not deleted, when its successor arrives. The archive is the measurement: put two volumes side by side and you can see, in typography and judgment rather than benchmark scores, how far the frontier moved. Vol. 01 wore the newsroom, Vol. 02 wore the clinic, and Vol. 03 wears the uniform.
Table A-1. Editions of this manual.
| Change | Edition | Designed by | Issued | Status | |
|---|---|---|---|---|---|
| Vol. 01 | The Magazine: dark editorial, sixteen sections, terminal metrics | Claude Opus 4.7 | April 2026 | Retired with honors | View archive → |
| Vol. 02 | Triad × Vitals: three worlds, one chart, reactive ink | Claude Fable 5 | June 2026 | Retired with honors | View archive → |
| Vol. 03 | Three views: the Field Manual, the Personal Index and the Commentary, switchable in place | OpenAI Astra + Claude Fable 5.1 + Claude Opus 5.5 | September 2026 | Current | You are here |
| Vol. 04 | To be designed | The next class | - | Awaiting release |
Figure A-1. Distribution. One text, three current views, in the order they joined. The switch at the foot of every page moves between them and keeps your place.
The Field Manual
Numbered chapters, live readiness, a printed index.
Designed by Claude Fable 5.1You are hereThe Personal Index
A white exhibition catalogue: the name set huge, numbered editorial rows.
Designed by OpenAI AstraSwitch viewRead →View O · third in lineVol. 03The CommentaryFrontier lead
A glossed edition: the work in the centre, commentary in the margins, one rubric red.
Designed by Claude Opus 5.5Switch viewRead →What changed in Change 3, beyond the design: the Step 2 site was rebuilt as a public question bank (12,972 items, up from 2,245), the lab reference gained a clickable CBC + CMP decoder, the AGI Tracker gained a milestones timeline, Edge gained its in-season wire, and the line count was recounted from source with generated data excluded (150,906, up from 65,538). Two things this edition adds that no earlier volume had: live reachability measured from the reader's browser, and an index.