Projects

35 projects. 27 of them are shown below as a screenshot of the live site, with the one number each measured and the committed file that number was read from.

35: in the record
27: live, with a screenshot

Exhibit

Screenshot of the sluice landing page showing its idempotent-execution comparison: 666 duplicate charges under naive retry versus 0 with sluice, next to a results table breaking down intent success rate and duplicate side effects.

sluice (opens the live site in a new tab)

Exactly once, provably

666 duplicate side effects with naive retry, 0 with sluice

chaos/results/latest.json

receipts code

Flagship

Screenshot of the klik landing page's dark intro screen, headlined “Tired of sending applications?” above a join-the-waitlist button.

Klik (opens the live site in a new tab)

Hiring by swiping

16 of 19 tables cleared on account deletion

docs/measurements/klik-account-purge.json

receipts

Demo

Screenshot of the polis map mid-pan, showing agent guild districts as isometric buildings and a banner noting that 8 of 59 agent citizens answer to no charter.

polis (opens the live site in a new tab)

An ecosystem, walkable

0 unreachable members, down from 8, and 25 of 59 never dispatched

data/ecosystem.json

receipts code

Flagship

Screenshot of dipmeter showing 230,059 earthquake hypocentres glowing inside a transparent Earth, coloured by depth from orange at the surface to blue at 700 km, with the Tonga and South American slabs standing out as dipping sheets beside the depth, magnitude and time controls.

dipmeter (opens the live site in a new tab)

Slabs, not blobs

230,059 located hypocentres against 27 modelled slabs

public/data/manifest.json

receipts code

Exhibit

Screenshot of orrery showing the asteroid belt as a blue haze of 120,000 sampled bodies around the inner planets, with the population counts, the find-a-body index and an instrument readout of drawn count, frame time and solver beside it.

orrery (opens the live site in a new tab)

Solved every frame

0.0149 AU worst blind-to-predicted gap disagreement

data/audit.json

receipts code

Exhibit

Screenshot of the kedge Raft simulator mid-run, showing a red "not linearizable" verdict with the witness operation that could not be placed anywhere in a consistent order, above the client history strip it was derived from.

kedge (opens the live site in a new tab)

Consensus, caught failing

102 of 102 Jepsen etcd verdicts matched

vendor/porcupine/jepsen/ (103 histories, MIT) replayed by src/cli.js corpus

receipts code

Exhibit

Screenshot of the windlass page headed "Three things this pipeline runner refuses to do, and the test that proves each", with the first refusal, that it will not run a model inside the runner, opened above the code that enforces it.

windlass (opens the live site in a new tab)

Pipelines that stop

1 planted token, 0 copies in the log, 0 model calls in the runner

tests/redaction.test.mjs

receipts code

Demo

Screenshot of the proofpage project page describing a real bug it caught in its own README: the first documented command failed with ENOENT because no examples directory existed, despite 51 passing tests.

proofpage (opens the live site in a new tab)

Only measured numbers

1 README command failing on a fresh install, missed by 51 green tests

README.md

receipts code

Flagship

chaff (opens the live site in a new tab)

Linting agent rules

44 golden cases at 100% exact match, 0 enforced findings on its own CLAUDE.md, 0 measured rules shipped

evals/cases/

receipts code

Exhibit

Screenshot of the flume page explaining that every event carries both an event time and an arrival time, above its stream controls and a running tally of complete, revised and uncertain windows.

flume (opens the live site in a new tab)

Late data, handled

124 false positives retired, 0 findings after, on three real streams

test/streams.test.js and vendor/gharchive, vendor/wikimedia, vendor/usgs (three 5,000-event slices, each with a SOURCE.md licence note)

receipts code

Exhibit

Screenshot of the cofferdam crash harness showing a CORRUPT verdict at a single crash point, with the acknowledged record the log lost and a byte-level diff of the expected and recovered state beside it.

cofferdam (opens the live site in a new tab)

Every crash point

12,600 crash points, 0 corrupt, 0 unverifiable on the correct build

crashes.json (the committed report) and src/cli.js control

receipts code

Flagship

Screenshot of the snapgauge tool displaying a real MCP schema diff between two server versions, showing a breaking removed tool and a fail-on-risky gate that would exit CI with code 1.

snapgauge (opens the live site in a new tab)

MCP contracts, pinned

37 golden and compat cases at 100% exact match

evals/cases/

receipts code

Flagship

Screenshot of the provenote page explaining its C2PA provenance-chain checker, above the drag-and-drop panel used to inspect an image's signed provenance chain entirely in the browser.

provenote (opens the live site in a new tab)

Provenance, and its gaps

7 hash-verified fixtures, 1 valid signature on a fabricated scene

fixtures/hashes.json

receipts code

Frozen

Screenshot of the dogwatch status page reporting a stalled watch, with its run tally of 78 checks, 2 findings, 0 gates and $0.0000 cost above a diagram of its approval-gate flow.

dogwatch (opens the live site in a new tab)

Every run, published

13 scheduled nightly records, 2026-08-15 to 2026-08-28

runs/index.json

receipts code

Frozen

Screenshot of the tiltmeter page showing its calibration numbers, a 0.0% false-positive rate and 95.0% detection power from 200 seeded pairs, above a diagram of how two harness runs are compared.

tiltmeter (opens the live site in a new tab)

Model drift, measured

0.0% false-positive rate, 95.0% detection power

evals/calibration/RESULTS.md

receipts code

Study

Screenshot of the shipgauge page's dark in-browser benchmark interface, with model and device selectors and a “run cold test” control for measuring shippability on the visitor's own machine.

shipgauge (opens the live site in a new tab)

Browser ML, costed

48% to 74% below advertised, 19 of 20 rows

results/results.json

receipts code

Frozen

assay (opens the live site in a new tab)

Brand kits, computed

3 of 4 brands byte-identical, 1 refused at 1.85:1

case-study/shipgauge/DIFF.md

receipts code

Frozen

Screenshot of the galley landing page, its title rendered upside down, describing the Remotion-rendered release-video CLI built from a repo's real commits and CI data.

galley (opens the live site in a new tab)

Commits become video

36.0 s, 913,793 bytes, 5 commits

examples/dogwatch-ci-fix.mp4

receipts code

Frozen

Screenshot of the halo-halo Taglish code-switch segmenter mid-analysis, showing a sample sentence tokenized and colour-tagged by language: Tagalog, English, mixed intra-word, and named entity.

halo-halo (opens the live site in a new tab)

Taglish, switch by switch

boundary F1 98.9% on 95 switch points, draft-automated

eval/results/v0-eval-report.json

receipts code

Frozen

Screenshot of the swage ASL fingerspelling classifier page, showing its seven-step on-device prediction pipeline and a boxed disclosure that its accuracy figure is provisional, not from a held-out signer.

swage (opens the live site in a new tab)

Fingerspelling, graded

95.7% top-1 on 281 samples, provisional

model/eval-report.json

receipts code

Demo

Screenshot of the graticule semantic-similarity map mid-render, showing scattered plotted notes beneath its on-device model status and a running count of zero network requests.

graticule (opens the live site in a new tab)

Notes mapped locally

contradiction 0.86 vs paraphrase 0.75 cosine

fixtures/linguistic/negation-pairs.json

receipts code

Demo

Screenshot of the clarifier tool with a real penguin-measurement dataset plotted as a force-settling point cloud next to its column-mapping table assigning each CSV field a physical role.

clarifier (opens the live site in a new tab)

CSV, settled by physics

0.726 vs 0.672 silhouette, verdict: no meaningful gain

evals/fixtures.eval.test.ts

receipts code

Demo

Screenshot of the pitman page's file-drop interface for recording or uploading a WAV or MP3 clip to see its on-device transcript with per-word confidence.

pitman (opens the live site in a new tab)

What it actually heard

11 of 18 clips (61%) with an error, 1 pattern survived the control

docs/batch2-asr-probe.md

receipts code

Demo

Screenshot of the orphanage crawl report explaining how a re-run against agentjames.vercel.app reduced 40 reported findings down to the one that actually mattered, a /games page missing from the sitemap.

orphanage (opens the live site in a new tab)

Forty findings, all wrong

40 findings to 0 unlisted, 1 real orphan surfaced

site/report.html

receipts code

Demo

Screenshot of the aeo-lab page showing a real logged scan against notion.com, listing the HTTP status each crawler, including Googlebot, GPTBot and ClaudeBot, received that day.

aeo-lab (opens the live site in a new tab)

Honest about unknowns

1 self-caused block reclassified: 429 at 250 ms, clean at 2000 ms

examples/notion.com-2026-09-05.json

receipts code

Demo

Screenshot of the carillon page's dark keystroke-to-sound interface, with its scale selector, typing textarea, and live counters for notes played and average interval.

carillon (opens the live site in a new tab)

Typing becomes sound

6 rules, each with a fixture that fails it, 0 network calls in the shipped site

verify.mjs

receipts code

Frozen

Screenshot of the ClubScope Insight Engine dashboard in replay mode, showing a country club's detected findings, active members, dues at risk, 90-day revenue and a churn-risk breakdown, with a declining-dining-revenue finding expanded below.

ClubScope Insight Engine (opens the live site in a new tab)

No invented numbers

51 of 51 figures recomputed, 1 fabricated figure caught

packages/core/src/evals/cases.ts

receipts code

Also in the record

No deployed URL to photograph: client work behind a login, a hook that runs on a machine, a library. Listed rather than illustrated, because an invented image would be the one thing this site refuses.

driftwatchDocs that drifted
2026 · Completed
A documentation-drift checker with exactly three outcomes, whose first real run across seven repositories reported ten failures and every one was a false positive.
code
claude-code-boundary-guardAgents stay put
2026 · Live
A PreToolUse hook that stops an agent in one project from writing into another: one file, no dependencies, and a self-test that has actually failed known-bad cases.
code
tallyDependencies, audited offline
2026 · Completed
An offline supply-chain and licensing auditor for npm projects, whose first real run flagged its own SPEC.md as a leaked test file.
code
consentrySMS rules, linted
2026 · Completed
A rules-based linter for US A2P 10DLC SMS campaign registrations that detects documented rejection causes offline and refuses to predict approval.
Business Operations PlatformReal business operations
2026 · Live
An operations platform a real business runs on: purchase orders, stock takes, approvals.
AGENTJAMES ControlMission control, local
2026 · Live
A desktop mission control for every project on one machine. Local only, and it still works with the network off.
Kredit Hero: AI Credit AppCredit scoring, shipped
2025 · Completed · Kredit Hero
Full stack work on an AI credit application, as a software engineer at Kredit Hero.
Agent James: Agent ConsoleYou are here
2026 · Live
This site: an IDE you operate, where every claim carries a receipt.

The console record, with the field filters

Projects

35 of 35 projects.