Repository navigation
The goal page's expected-overlap figures use the right estimator, and now agree with the collapse page - #678
Merged
Merged
Conversation
… now agree with the collapse page The blind verify round (#674) found docs/goals/ and the chorus-collapse page disagreeing: 112.6/40.4 against 111.5/39.6. The goal page's numbers were mine (#668) and they were the worse of the two. Both estimate the same thing — how many of chorus_norm's inner fits would collapse in BOTH draws if collapse were independent within a configuration. I pooled each configuration's rate across the two draws and squared it. The page multiplies each draw's own rate. Squaring an average is never below multiplying the two values, so pooling biases the estimate upward, and it is the product that is unbiased when the two draws may sit at different rates. Recomputed from the same collapse_table.json: 111.5 against 111 observed, and 39.6 against 38. The reading does not change — the overlap is what chance gives, and the failure is not shown to be deterministic — but the figure a reader checks is now the defensible one and the two pages no longer contradict each other. Co-Authored-By: Claude Opus 5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_017hQX6iESxBsQk7Jan875e3
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Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be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.
The blind verify round (#674) flagged
docs/goals/learned-model-family.mdstating 112.6 / 40.4 where the chorus-collapse page states 111.5 / 39.6. The goal page's numbers were mine, from #668, and they were the worse of the two.Both estimate the same quantity: how many of
chorus_norm's inner fits would collapse in both draws if collapse were independent within a configuration.chorus_normchorus_gain_normSquaring an average is never below multiplying the two values, so pooling biases the estimate upward; the product is what's unbiased when the two draws may sit at different per-configuration rates. Recomputed from the same
collapse_table.json.The reading is unchanged — the overlap is what chance gives, and the failure is not shown to be deterministic. What changes is that the figure a reader checks is now the defensible one, and the two pages no longer contradict each other. That contradiction is exactly what #674's cross-check exists to find, and it found mine.
Docs only, one line. Sapper and
check_quotesclear.🤖 Generated with Claude Code
https://claude.ai/code/session_017hQX6iESxBsQk7Jan875e3
Generated by Claude Code