Lab notebook
What was actually done, dated, including what failed. A research programme that only publishes conclusions hides the part that would be useful to someone repeating it.
4 November 2026 — later
Built. The hypothesis machine: 42 domains, 12 relations, 8 deducers, lockable slots, custom words, JSON export, mobile-first. Live at /hypotheses/.
Failed. The lock feature broke the app and I published it broken. Deleting a render() function left three calls to it; the script threw at startup and the page rendered an empty card with four blanks. node --check passed — the syntax was valid, the failure was a runtime reference. Caught only by screenshotting the page and looking.
Lesson. A syntax check is not a test. The check that works is rendering the artefact and looking at it, which is now the fourth time in this project that a change looked applied while being functionally dead.
4 November 2026
Published. Three books (60, 29 and 152 pages), eleven site sections, a 58-function toolkit and a browser audit tool.
Found. A floating-point comparison at a completeness threshold dropped 2,060 records and moved a headline number by 28%. Caught because two independently written scripts disagreed.
Failed. The first workbook upload silently re-sent the previous file under a new name, and the page claimed 26 pages while linking 21. Found only by checking the live artefact.
Failed. A chapter insertion did nothing because its anchor text had changed in a previous pass; the edit reported success and the page count barely moved.
Corrected. The closure-dating penalty turned out to be model-dependent: a third similar marker improved conditioning 26 → 9 under a two-parameter model. The claim was narrowed rather than withdrawn.
Next. A twenty-claim literature screen with a sampling rule, to replace the purposive sample of eight.
3 November 2026
Ran. 78,493 USGS earthquake records through a per-scale analysis. The pooled exponent is stable to ±2% across three years while the five magnitude scales span 0.70–1.74.
Found. Last-digit χ² of 3,150 against uniformity — a rounding convention that a mean would inherit.
Published. Sixteen corrections, including four that damaged my own strongest claims.
How to read this
Entries are added when something is finished or fails, not daily. The failures are the useful part: a defect found by an artefact check is a procedure other people can copy.