Repository navigation
fix(ci): two gates that could not fail, and the arithmetic that says so (#829, #559) - #882
Merged
Merged
Conversation
…so (#829, #559) Both report a verdict they did not measure. They are the same family as "un commentaire n'est pas un contrôle", one level up: a control that cannot fail is not a control, and a control that fails on a healthy subject is worse than none because it teaches the reflex of re-running it. ## The gofmt hook reported Passed when gofmt could not run `.pre-commit-config.yaml` read `test -z "$(gofmt -l .)" || …`. `gofmt -l` writes the offending files to stdout and its own errors to stderr, and the verdict was `test -z` over the captured stdout alone — the exit code was never read. Any failure of gofmt produced an empty substitution, `test -z ""` was true, and the hook exited 0. Not skipped, not red: **Passed, asserting it looked.** Its window is a fresh clone before `mise trust`, which is a new contributor's first commit, and it is the only hook in that file with the shape — `go vet` and `golangci-lint` propagate their own status, and both reddened at the same instant on the same cause, which is the only reason the gofmt green was ever noticed. The test runs the hook's own `entry:` line, read out of the YAML rather than copied into Go: a test asserting the entry *contains* `||` would pass on any rewrite that kept the characters and lost the property. Both directions are asserted in the same file, because a hook that always failed would satisfy the refusing test and reject every commit. ## A balancer verdict drawn from six coin flips The Outscale dataplane suite asked **six** connections whether a balancer reaches both of its machines. A balancer that works sends every connection to one side with probability 2 × 0.5⁶ = **3.1 %**, so the gate reddened about one run in thirty-two on a subject that was behaving — on the suite whose subject is this project's own differentiator. The same arithmetic runs the other way after an unlink: six probes that happen to avoid a machine the unlink never reached is 0.5⁶ = 1.6 %, a false **green** over a real defect. That site's comment already named it "#559's defect with a different number" and left the number. `DISTRIBUTION_PROBES=24` puts both tails at 1.2e-7 and 6.0e-8, for about ten seconds on a leg that measures in minutes. The liveness probe keeps six and says why: "did every connection get an answer" has no false negative that more probes would remove, so a bigger sample there buys wall-clock and nothing else. What the suite still does not prove is written beside the constant rather than left to be assumed: that the distribution is **fair**. Both machines being reached says the balancer has more than one backend, nothing about the ratio. The control reads the script, because the subject *is* the sample size — but it checks a property, not a value: every probe whose verdict names a machine draws from the declared constant, and that constant supports the claim. Changing the number stays free; going back to a literal does not. /falsify: 2 mutations on a-hook-that-cannot-measure-fails.json and 2 on a-verdict-drawn-from-six-coin-flips.json, all biting. Each spec mutates the over-correction as well as the defect, because a guard that refuses everything passes its own attack test. mise run prepush green. Assisted-by: Claude Code (claude-opus-5)
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
Add this suggestion to a batch that can be applied as a single commit.This suggestion is invalid because no changes were made to the code.Suggestions cannot be applied while the pull request is closed.Suggestions cannot be applied while viewing a subset of changes.Only one suggestion per line can be applied in a batch.Add this suggestion to a batch that can be applied as a single commit.Applying suggestions on deleted lines is not supported.You must change the existing code in this line in order to create a valid suggestion.Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be applied on multi-line comments.Suggestions cannot be applied while the pull request is queued to merge.Suggestion cannot be applied right now. Please check back later.
Both report a verdict they did not measure.
The gofmt hook reported
Passedwhen gofmt could not run. The entry wastest -z \"$(gofmt -l .)\", which discards the exit code:gofmt -lwrites its errors to stderr, so any failure produced an empty substitution and the hook exited 0. Not skipped, not red — green, asserting it looked. Its window is a fresh clone beforemise trust, which is a new contributor first commit, and it is the only hook in that file with the shape:go vetandgolangci-lintpropagate their own status and both reddened at the same instant on the same cause, which is the only reason the gofmt green was noticed.A balancer verdict drawn from six coin flips. The Outscale dataplane suite asked six connections whether a balancer reaches both machines; a balancer that works sends every connection to one side with probability 2 x 0.5^6 = 3.1 percent. One run in thirty-two, red on a subject that was behaving, on the suite whose subject is this project own differentiator. The same arithmetic runs the other way after an unlink: six probes that avoid a machine the unlink never reached is 1.6 percent, a false green over a real defect — and that site comment already named it. 24 probes puts both tails at 1.2e-7 and 6.0e-8, for about ten seconds on a leg that measures in minutes. The liveness probe keeps six, and says why.
Both tests run the thing rather than reading it: the hook test executes the
entry:line taken out of the YAML, with a PATH holding no gofmt. Both assert the ACCEPTING half too, because a guard that refuses everything passes its own attack test and teaches--no-verify./falsify: 2 mutations per spec, all biting, each spec mutating the over-correction as well as the defect.
mise run prepushgreen. The testplan prescribesfalsify:lintand the specs it names for this diff.Closes #829. Closes #559.
🤖 Generated with Claude Code