jakelawrence.xyz
Blog

Everything, newest first

Posts, essays, research, systems, games, and tools in one long scroll. For the curated reading list, see Writing.

The feed ยท everything, newest first
RESEARCH1 week ago7 min

Stress on 12,000 Ukrainian words was a lookup, except for 246 arguments

Ukrainian never marks which syllable is stressed in print, so a learner's word list has to. I added the stress to all 12,000 words on my open Ukrainian frequency list: most came from an open dictio...

ukrainianlanguage-learningopen-dataword-stress
RESEARCH1 week ago

Civic Crypto

Stablecoins are the one part of crypto with real public-sector consequences right now, and a research lab's best first move is to write down predictions about Illinois' first-of-its-kind crypto tax, with probabilities, before it takes effect on January 1, 2027. The second program of a public research lab, with its three studies drafted and none registered yet, which the page says.

StablecoinsPreregistration
RESEARCH1 week ago

Agent Visualization

One synthetic four-agent run goes wrong in a planted way: a source check is skipped, an uncited claim slips through, and the critic accepts it with every span green. Eight visualization techniques from the literature each show a different slice of that failure, live, and two cannot see it at all.

Multi-Agent SystemsObservability
TOOLS2 weeks ago6 min

Free Claude credits for solo builders go through Amazon

Anthropic's startup program now takes anyone, but its credits still want an institutional investor, and its credits only spend on its own API. For one person with no funding, the most reliable free...

ai-creditsclaude
ENGINEERING2 weeks ago6 min

The second CourtCollect is four percent the size of the first

In January I spent six days building CourtCollect, a court-to-collections platform that reached 174,238 lines. When a Texas court needed the work in May, what shipped was a 6,851-line review consol...

build-logcourtcollect
RESEARCH2 weeks ago

AI Acceptable Use

A plain-language AI policy for small governments and nonprofits, where every rule shows how to follow it with a spreadsheet and where this site enforces the same rule in code. Answer six questions and it trims itself to the rules you need.

AI PolicyAcceptable UseGovernance
RESEARCH2 weeks ago

SOC 2 Readiness

A SOC 2 self-assessment of a platform run by one person, graded twice: for design, the Type I question, and for operation, the Type II question. The controls that drop a grade between the two are mostly jobs that were built and then switched off.

SOC 2Trust Services CriteriaSecurity
RESEARCH2 weeks ago

Case Study Library

Every merged pull request in this repository's own history, audited, tagged, and knowledge-graphed, using the same content-as-code method the site turns on everything else, turned on the site itself. On top of it, hand-written case studies about how the site gets built, each cited into that corpus and receipt-checked in the browser.

Pull RequestsKnowledge GraphClassification
RESEARCH3 weeks ago6 min

My best number was a word list read back to me

A month of work on the Ukrainian frequency list ended with a DOI, a live lemmatizer, a correction sent to the authors of the profile I lean on, and a headline result that dropped nine points once I...

build-logukrainianopen-data
RESEARCHSep 3, 2026

Italian as Black

Nine hundred and two of my own games say I have been making the same unexamined decision since 2010, and losing about half a point a time for it. Play the position and the record answers back.

ChessSelf-MeasurementHabit
ESSAYJul 27, 20266 min

The napkin math is public now

Moonshot put a three-trillion-parameter model on Hugging Face, and the most interesting number in the release is not a benchmark. When dozens of providers race to serve the same public artifact, th...

essayopen-weightskimi-k3
ESSAYJul 27, 20267 min

Seeing like a model

Anthropic cut Claude Code's system prompt by 80 percent for the Claude 5 models and told everyone to simplify. So I audited the 66,792 bytes my own agents auto-load, atom by atom. The doctrine's cu...

essaycontext-engineeringclaude-code
RESEARCHJul 27, 2026

Ninety-Seven Percent Off

A gray market resells frontier AI access to Chinese users at up to 97.8% off official pricing, and the discount is not a discount. It is a measure of the fraud: the share of the bill that stolen cards, charged-back spend, and abused free trials force someone else to pay.

Invisible InfrastructureClassificationGray Market
RESEARCHJul 27, 2026

AlphaFold

AlphaFold made the whole protein universe legible, and it ships a per-residue confidence that tells you exactly where to trust it. Rotate a real prediction of p53 and watch its certain core give way to tails the model honestly cannot place.

AlphaFoldProtein StructureConfidence
RESEARCHJul 24, 2026

ICE Claims

A methodology inventory for claims about U.S. immigration enforcement. Not an argument: thirteen talking points from both directions mapped to the sources that would settle them, the four counting decisions that flip verdicts, and a status board for the official series that have gone dark. The best current data on ICE is adversarial (litigation-driven FOIA, independent clearinghouses), not official.

Open DataImmigration EnforcementLegibility
Loadingโ€ฆ