Repository navigation
R13: position-effect pairing, one character measure, un-softened zh README - #33
Merged
Merged
Conversation
…ten zh README Three items left open by the pre-merge review. P2 — the position effect dropped every judge after the first. forward/reversed were keyed by regime alone and kept the first entry, so with two judges only one contributed. Reproduced: codex scores both arms 7 in both passes while a second judge swings 8 points on the same civ, and meanAbsDelta came back 0. That number is the noise floor a paired score gap must clear in the E1/E2 analysis, so under-reporting it marks order-sensitive results as resolved — exactly backwards. Passes are now paired within a provider, perProvider breakdown is reported, and maxAbsDelta is added because the floor a gap must clear is the worst judge's movement, not the average across judges. Revert-verified. P3 — 72,828 and 72,830 were both in the docs for the same transcript. Both are correct: JavaScript String.length counts UTF-16 code units, Python len() counts code points, and the transcript contains two astral characters (🔄). The docs now use the figure the code reports throughout, and E1-results states which measure that is, because mixing the two in one document is the actual defect. P3 — the Chinese README dropped the §9.3 table entirely and kept only "found 5 real errors". Four of those five are still unfixed, and a reader of the Chinese page could not know which. The table is restored with the same four⚠️ pending rows the English page carries. The rule this violated is one I set: the Chinese page is a translation of the same claims, not a softened summary. CI: 519 backend + 25 frontend + 57 regimes + smoke.
LeoLin990405
pushed a commit
that referenced
this pull request
Aug 6, 2026
feat: install.sh 跨平台支持 (macOS/CentOS/RHEL/Alpine)
LeoLin990405
added a commit
that referenced
this pull request
Aug 6, 2026
…ten zh README (#33) Three items left open by the pre-merge review. P2 — the position effect dropped every judge after the first. forward/reversed were keyed by regime alone and kept the first entry, so with two judges only one contributed. Reproduced: codex scores both arms 7 in both passes while a second judge swings 8 points on the same civ, and meanAbsDelta came back 0. That number is the noise floor a paired score gap must clear in the E1/E2 analysis, so under-reporting it marks order-sensitive results as resolved — exactly backwards. Passes are now paired within a provider, perProvider breakdown is reported, and maxAbsDelta is added because the floor a gap must clear is the worst judge's movement, not the average across judges. Revert-verified. P3 — 72,828 and 72,830 were both in the docs for the same transcript. Both are correct: JavaScript String.length counts UTF-16 code units, Python len() counts code points, and the transcript contains two astral characters (🔄). The docs now use the figure the code reports throughout, and E1-results states which measure that is, because mixing the two in one document is the actual defect. P3 — the Chinese README dropped the §9.3 table entirely and kept only "found 5 real errors". Four of those five are still unfixed, and a reader of the Chinese page could not know which. The table is restored with the same four⚠️ pending rows the English page carries. The rule this violated is one I set: the Chinese page is a translation of the same claims, not a softened summary. CI: 519 backend + 25 frontend + 57 regimes + smoke.
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
1 participant
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.
The three non-blocking items from the PR #32 pre-merge review.
Position effect dropped every judge after the first (P2).
forward/reversedwere keyed by regime alone and kept the first entry, so with
--judges 2onlyone provider's swap movement survived. Reproduced: codex scores both arms 7 in
both passes while a second judge swings 8 points on the same civ, and
meanAbsDeltacame back 0.That figure is the noise floor a paired score gap has to clear in the E1/E2
analysis. Under-reporting it marks order-sensitive results as resolved, which
is the opposite of what the position check exists for. Passes are now paired
within a provider;
perProvideris reported;maxAbsDeltais added because thefloor a gap must clear is the worst judge's movement, not the average across
judges. Revert-verified — collapsing providers into one bucket reddens it.
72,828 vs 72,830 (P3). Both correct for the same transcript: JavaScript
String.lengthcounts UTF-16 code units, Pythonlen()counts code points, andthe transcript contains two astral characters (🔄). The docs now use the figure
the code reports, and
E1-results.mdstates which measure that is. Mixing thetwo in one document was the defect, not either number.
The Chinese README softened §9.3 (P3). It dropped the findings table and kept
only "found 5 real errors" — four of which are still unfixed, and a reader of the
Chinese page had no way to know which. The table is restored with the same four
pending rows the English page carries.
CI: 520 backend + 25 frontend + 57 regimes + end-to-end smoke.