Add paired-bootstrap CIs to memory-pollution model differences#142
Merged
Conversation
- run_audit_eval.py: surface the exception message (not just the type) on a failed task so balance-exhaustion (402) is visible to wrappers. - render_finaudit_figures.py: draw 95% Wilson intervals on the difficulty-tier bars when the CI experiment is present, using its temperature-0.7 point estimates for consistency. - audit_eval_ci_wilson.csv: per-(model, tier) recall with Wilson 95% intervals over 6 models x 100 tasks x 3 samples. The difficulty inversion is confirmed: strong auditors' L1 intervals (lower bound >= 0.73) are disjoint from weak auditors' (upper bound <= 0.60).
model_difference.csv now carries the paired-bootstrap 95% interval on each agent's hold-ratio shift (the interval the analysis already computed but did not surface), and the per-model figure draws error bars. This matches the confidence reporting in the execution-sensitivity results; the binomial recall in the audit benchmark uses Wilson intervals instead. The intervals confirm the model split: gemini-3.1-pro and deepseek-v4-pro shifts exclude zero while gpt-5.5 and claude-opus-4.7 sit on it.
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Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be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.
Surfaces the paired-bootstrap 95% intervals the memory-pollution analysis already computed but did not output: model_difference.csv gains ci_low/ci_high and the per-model figure gets error bars. Consistent with the execution-sensitivity paper's CI reporting (the audit benchmark uses Wilson because its outcome is a proportion). Intervals confirm the split - gemini/deepseek shifts exclude zero, gpt/claude sit on it.
Test plan