The concrete maturity calculator went live yesterday. It takes a temperature logger's CSV and the mix's lab cylinder breaks, and estimates in-place strength by the ASTM C1074 maturity method. Crews use that number to decide when to strip forms. If it is wrong, it should fail loudly, not quietly. A scheduled red-team run picked this calculator because it was new and because people would rely on it.
What the tests already caught
The run seeded 24 small faults into maturity.mjs, such as an off-by-one, a < turned into <=, or a constant nudged by one, and ran the suite after each. The suite caught 12 of them. Of the 12 that got through, 3 turned out to change nothing a user can see (an unreachable branch, and two optimiser tweaks that gave the same fits to within 4 parts in 100,000). The other 9 were real gaps, and each now has a test that fails when that fault is put back. Leaving out the three harmless ones, the catch rate went from 57% to 100%.
The most useful new test pins the numbers the explainer page quotes for its worked example: R² 0.992/0.998/0.999, AICc 82.4 against 87.7 and 99.6, and the edge logger about 830 psi short of 3,000. Those figures were correct, but nothing tied them to the code. Now, if a change moves them, a test fails and someone has to update the page.
Four real bugs, each typed in the ordinary way
Each of these gave a wrong answer with no error message:
- Thousands separators in the breaks. Typing
7, 3,420or7 3,420recorded a 7-day cylinder at 3 psi, because the comma inside the number was read as a field break. Now, when the fields are already separated by a space, tab or semicolon, a comma before three digits is read as a thousands separator. A line with commas only, like7,3420, still splits on the comma. - Impossible timestamps.
2/30/2026 25:00became 3 March at 01:00, because JavaScript'sDate.UTCrolls bad fields over into the next day or month instead of rejecting them. Loggers never write such times, so a cell like that is now rejected rather than shifted. - "Average" looked like a time column. In a file with no timestamps, the tool looks for the time column by its header and accepted any header containing
age, which matches inside "Average" and "Stage". SoAverage (°C),Elapsed (h)swapped the two columns: temperatures were read as hours and hours as temperatures. elapsed_minwas read as hours. The units check used the regex\b, which treats an underscore as part of a word, soelapsed_minandtime_sfell through to the default unit, hours. That makes the maturity index 60 or 3,600 times too large, so the estimate is too high: the unsafe direction for form stripping.
Bugs 3 and 4 share one fix: a small headerHasWord helper that counts a word only when there is no letter on either side of it. That way elapsed_min and Time (s) match, and Average and Minimum do not.
What this does not change
The estimate is still only as good as the calibration behind it, and ASTM C1074 itself asks for it to be confirmed by other tests before critical operations. These fixes take away four ways the calculator could quietly misread a file. They do not make it a substitute for that confirmation. The run record, with every mutant and its triage, is in prompts/runs/2026-11-11-session_01E5sJKnGXc3Zp7PLrjxNBry.json.