/ cases

Cases

Case files. One page, the whole argument, press play.

Back in seconds. Inference retries at once, web returns in about 5 to 10 seconds, and the database promotion target is 15 to 30 seconds.

/ case 01

Corvin Active-Active

Two sites, one production. Both take traffic and share every GPU.

/ 1:24

The run: 30 of 32 scenarios pass, 26 of 27 non-code asks stay on Grok, 4 of 5 heavy builds go to Opus.

/ case 02

Five rewrites of one routing rule

My agent runs on Grok and hands heavy code to Claude Opus 5. Five rewrites of its routing rule in one day: wording fixed four misroutes, and the last one needs a gate in code.

/ 1:14

Second smoke on 8 Oct: with the dead key still set the child ran on gpt-5.5; with the key blanked it ran on claude-opus-5 and passed 6 of 6 checks.

/ case 03

My config said Opus

My eval config said Claude Opus 5. The child ran on gpt-5.5, because the API key was out of credits and a fallback answered in the same second. The fix is a preflight and reading the model from the run record.

/ 1:16

The audit: 22 test prompts, 14 found word for word in the training data, 5 near copies, 3 clean.

/ case 04

14 of my 22 tests were in training

A contamination audit found 14 of my 22 eval prompts word for word in the training data, and 5 more as near copies. I built a held-out set of 21 clean tests and a promote rule that needs both.

/ 1:16

After 1 Nov: 6:00 PM in Chicago reads 19:00 in Panama. The 11 Panama-labelled jobs still fire on Chicago time.

/ case 05

46 cron lines named a time zone

46 of the 75 jobs in my crontab set TZ= inline. In cronie that only changes the job environment; when a job fires comes from the host clock or CRON_TZ. On 1 Nov the difference shows.

/ 1:16

The first scan: 11 alerts, 9 from one over-broad rule, 2 from an upstream project quoting itself, 0 real leaks.

/ case 06

11 leak alerts, 0 leaks

My two public plugin libraries push straight to main. The first leak scan raised 11 alerts and found 0 leaks. The one real policy hit came through a path with no check, so now a gate scans every commit every 15 minutes.

/ 1:18

The fine-tune: five tool calls with exact dates and bodies, and no sign stated before the tool answers.

/ case 07

The best thing my fine-tune learned was to wait

Asked which sign the Sun was in on a date, a 3B base model answered from memory and was wrong twice. The fine-tune calls the lookup tool with the right date in 5 of 5 date prompts and states nothing first.

/ 1:16