Writing
Notes from the shop. Every one of these is a real incident with its cause on the record.
2026-09-13
The Day a Database Could Not Learn a New Word, and Nobody Was Told
A migration dropped a column, the vocabulary table silently stopped accepting new words, and no alert fired. The interesting question is not how the bug got in.
2026-09-13
A Workshop Where Mistakes Get Caught
The same story at a shorter length: what it means to build a shop where the tools refuse the mistake instead of recording it.
2026-09-08
2026-09-08 — sprint/artifact numbering needs a database-authoritative allocator, not a file counter
Counting files to pick the next number works until two people count at once. What was found during a multi-worktree consolidation, and the operator's ruling: the database is the master.
2026-09-05
Two faults in a zoned water-source heat pump, found by measurement
Four service visits replaced parts and adjusted settings without fixing the complaint. Instrumenting the thermostats found both faults. What was measured, what it established, and what it ruled out.
2026-08-17
Sharding what a program cannot answer — S7, and four ways I got it wrong
Making every crate README a projection rather than a document — reported with three stated confounds and one design failure large enough that the headline result should not be acted on.
2026-08-16
Study: tier-true rename campaign report — the complete record and roadmap
Renaming a 395-crate workspace by planning every edit as database rows first: the machinery, the agents, the telemetry, and the roadmap for doing it again.
2026-08-16
E2 — name-only comprehension: current vs tier-true crate names — 2026-08-16
Show a model nothing but a crate name and ask where it belongs. The best model got 59% right, the stronger one 39% — and both read the brand prefix as an architectural claim it does not make.
2026-08-15
The Day We Measured Our Words
A plain-language account of twenty-four hours in which a team treated vocabulary as an engineering material, built a machine to package work for AI, and watched the AI's honesty find bugs in the machine.
2026-08-14
Vocabulary is measurable — a consolidation of the 2026-08-14 research programme
The consolidated research programme: whether models actually agree on what a word means, measured rather than assumed, with every number cited to a committed artifact.
2026-08-14
S1 results — vocabulary agreement (H8), run s1-2026-08-14
Put a candidate vocabulary to a panel of models and count where they disagree. The results, the rerun the operator ordered, and the candidate that was ratified.
2026-08-12
Semantic coherence and the harness plane
A session that began as a defect report on a documentation directory and ended as a methodology finding, with every claim marked by how it was established.
2026-08-04
Research — Observation, not presumption: an evidentiary standard for host instrumentation
Removing one word from a crate description forced the question of what the capability actually is. The answer reframed the program: instrumentation, not security tooling.
2026-08-18
Language as Architecture — thesis v2
The thesis, revised and narrowed: names are model inputs treated as facts, so a misleading name is not missing information but misinformation — and the vocabulary is therefore a system component with its own tests and ratchets.
2026-07-28
Language as Architecture
The long research agenda: whether language, ontology and constraint are load-bearing parts of software architecture, posed as questions to be tested rather than claims to be believed.
2026-07-24
Research — adversarial verification of the repair-shop framework-gap claims
Ten claims about what a framework must supply, each put to a verifier and two adversarial skeptics, then a completeness critic and a correction pass. Forty-five agents, and what survived.
2026-07-23
Research — The department model: a working repair-shop custody system
A working custody system from a computer-repair shop: departments as physical areas, the tub as the unit of custody, guards between stages, and why the intellectual property was the process rather than the code.
2026-07-23
Research note: the repair-shop operational model (field-proven blueprint)
The shop as it actually ran: intake photographed into the ticket, a consent link before any card is charged, the blind folder drag that keeps a technician from ever seeing a filename, and a locker combination released by payment.
2026-07-23
Research — Baseline attestation, informed consent, and custody records
A repair-shop intake agreement states the pattern more plainly than any specification draft: what is attested at the counter, what is consented to, and what custody record survives the job.
2026-07-20
Research note: blind data-handling (field-proven privacy pattern)
Move a customer's data off a failing drive without ever seeing a filename. A pattern the shop ran for years, protecting the customer's privacy and the technician's exposure at the same time.