The false sentence was fixed on the page and left standing in the file that renders it - #503
Merged
Merged
Conversation
…e that renders it Yesterday's correction moved the rendered header and stopped there. `make_detector_table.py`'s own module docstring still said the learned six's two halves were produced by "the same fit ... so their two rows do describe one instrument", three hundred lines above the header that now says the opposite. A reader opening the tool to find out what the page means would have got the retracted claim, and the page and its source would have disagreed with each other. One process is not one fit. The Scored row is the mean of four fits, one per held-out fold; the Observed row is a fifth, separate fit over the whole simulated corpus. The docstring now says so, carries the retraction where the wrong sentence was rather than in a changelog nobody opens, and adds the two facts the same conversation established: five of six shipped values lie outside everything the folds chose, and the split is 4-fold cross-validation holding out 2 of 8 at one training seed per fold -- not leave-one-out. `run_learned_on_folder.py` was right the whole time. Its `train_all` docstring already said "the bake-off fits four times, once per held-out fold ... this fits ONCE over the whole simulated corpus". The error was one file's summary of another, which is the shape worth noticing: the source was accurate and the paraphrase two hops away was not. Grepped the tree for the same sentence in any other file. There is none. Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
syncytium2
added a commit
that referenced
this pull request
Sep 9, 2026
Releases the claim #504 took. A claim cannot retire itself in the PR it rides in, which is why this is a second commit and not an edit to that one. Published `3cb89fd`, Version ID ab3cc81d-1ed5-415b-ada8-14deb77574ce, 7 of 8 assets changed. Verified against the live edge rather than the build, including reading the corrected figure off the served HTML with curl: the bypass runs below the difference-of-Gaussian bank instead of through it, for the first time since 2026-09-02. Two things recorded because a later deployer needs them and neither is flattering. The claim PR was pushed and open but had NOT merged to main when the deploy ran -- CI is about thirteen minutes and the instruction was a five-minute deploy. Nothing collided and no other session was deploying, but for that window the claim lived on origin and not in the file a starting session reads. Written down rather than smoothed over; if it recurs, landing a claim should not have to wait on a full CI matrix. And main moved under the preflight: #503 merged between the claim and the build, so the HEAD == origin/main check failed, the checkout was fast-forwarded and the build re-run. The published stamp is therefore 3cb89fd and not the caeef0c the claim block named. That check exists for exactly this and it earned its keep today. Co-authored-by: richard defazio <defazio@umich.edu> Co-authored-by: Claude Opus 5 (1M context) <noreply@anthropic.com>
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
2 participants
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.
#502's last commit corrected the rendered header and stopped there.
make_detector_table.py'sown module docstring still said the learned six's two halves came from "the same fit ... so
their two rows do describe one instrument" — three hundred lines above the header that now
says the opposite. Anyone opening the tool to find out what the page means would have got the
retracted claim, and the page and its source disagreed with each other.
One process is not one fit. The Scored row is the mean of four fits, one per held-out
fold; the Observed row is a fifth, separate fit over the whole simulated corpus.
tube_ratio_guardran the folder at threshold 0.990819, a value none of the four foldsproduced.
The docstring now says that, carries the retraction where the wrong sentence was rather than
in a changelog nobody opens, and picks up the two facts the same conversation established:
five of six shipped values lie outside everything the folds chose, and the split is 4-fold
cross-validation holding out 2 of 8 at one training seed per fold — not leave-one-out.
run_learned_on_folder.pywas right the whole time; itstrain_alldocstring already said"the bake-off fits four times, once per held-out fold ... this fits ONCE over the whole
simulated corpus". The error was one file's summary of another — the source accurate, the
paraphrase two hops away not.
Grepped the tree for the same sentence anywhere else. There is none. No figure is
regenerated: the rendered page is already correct.
🤖 Generated with Claude Code