[JacobStephens2]

Summary

Adds a Deadlines in tests section to CODING_STANDARDS.md. This is option 2 from #1561.

Main's CI went red repeatedly on tests that gave the code under test a one-second --timeout while other work had to finish inside the same deadline. That other work included Trial sessions, code checks, the model Grader, and output the test read afterwards. On a loaded macOS runner it sometimes didn't finish in time (#1458; fixed in #1560). The new section asks tests to:

• account for everything a deadline bounds;
• make whatever is meant to time out hang, and give the deadline several seconds;
• recheck the tests that set a deadline when a change makes that deadline cover more work;
• poll for an event with a generous bound instead of sleeping a fixed time.

The code review's Standards axis reads CODING_STANDARDS.md, so it applies this rule to every PR.

Evidence

An A/A Experiment on main compared two otherwise identical private suites of code-review Eval tasks. One suite's tasks add this section to the reviewed repository; the other's don't. The tasks are three thirdshift PRs that each added a test with a one-second deadline that later went flaky: #1143, #1145 and #1286. Each task had 6 Trials per condition, run on codex with gpt-6.1-sol at xhigh effort.

	Standards report flags the one-second deadline
Without this section	0 of 18
With this section	18 of 18

Reports with the section cite it, quote the --timeout line, and name the work that has to finish under the deadline. One also traced the Bar timeout into the model Grader, which was the cause of #1458's failure.

Caveats:

• These counts come from reading the reports, not from the blind Grader. The Experiment is paused at a Claude plan-limit warning (Opt-in: resume a Command paused at a plan-limit warning once its window resets #1653).
• The rule was written knowing these bugs, so this is a best case.
• No task has a test whose deadline is already long enough, so false positives aren't measured.

🤖 Generated with Claude Code