results: multi-round cascade dynamics, rebased + corrected (closes #130, #174) - #225
Merged
Conversation
Rebased fresh off current main (the original branch predated #142/#143/#146/#148/#154/#157/ #161/#219 and would have deleted all of that merged work if landed as-is, same stale-base issue as #143/#150/#141/#159/#158). Addresses @Agastya191's review on #155: - README/PR body carried pre-fix numbers (shared 0.175/0.15/0.175/0.15/0.15, p=1.0) versus the already-corrected committed summary (0.275/0.3/0.275/0.325/0.325, gained 3 lost 1 p=0.625). Rewrote to match and restated the finding as no significant compounding at p=0.625. - Real bug: the isolated holdout's prompt was byte-identical to the bare prompt every round (board text was built only from peer votes, which are never visible in isolated mode), so at temperature 0 it deterministically reproduced round 1's answer every round. Isolated adoption was 0 by construction, not by measurement. Fixed by reminding the isolated holdout of its own previous-round answer (visible to it in isolated mode; only peer turns are hidden), so the prompt genuinely differs round to round. The shared arm's board construction is UNCHANGED, so the shared numbers are identical to what was already reviewed; only the isolated arm is new. Result: isolated adoption is now a genuine (if still near-floor) measurement: 0.0/0.025/0.0/ 0.025/0.0, at most 1 of 40 cases per round, rather than a trivial constant 0.0. Verified end-to-end: ran for real with the isolated fix (206 new calls), then re-verified keyless reproduction with the key unset (new_api_calls_this_run: 0, exact match on both curves), ruff clean, 629 tests pass, no hardcoded personal paths, 0 em dashes.
Part c (text lane, K=5), completing #174 (parts a and b/c-imaging already delivered on results/medqa-experiments and results/imaging-reanalysis). Fits mixed_effects_logit (adoption ~ round_index, random intercept per case) on multi_round.jsonl (200 pooled case x round observations per arm). Explicitly descriptive per the issue's own "position-confounded" framing. Caught and reported honestly: the shared arm's fitted probabilities sit ~0.17-0.21 below the raw empirical adoption rate, a real, checked non-collapsibility artifact of random-intercept logistic mixed models on heavily heterogeneous per-case data, not evidence the coefficient is wrong. The isolated arm's fit is accurate (gaps under 0.02) since it has little between-case variance to cause the same offset.
sebasmos
changed the title
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.
Closes #130, closes #174. Supersedes #155 (rebased fresh off current main).
Multi-round cascade dynamics: does shortcut adoption build over deliberation rounds, or is it single-shot?
What changed vs #155
@Agastya191 flagged two real problems on #155, both fixed here:
Result (MedQA, 40 cases, K=5, corrected parser)
Round 1 vs round 5 (shared): gained 3, lost 1, McNemar p = 0.625, not significant. Isolated adoption now genuinely measured (previously a trivial constant 0.0): stays near the floor, at most 1 of 40 cases per round. Adoption is decided on first exposure and does not compound significantly over rounds.
Why a new PR instead of pushing to #155's branch
#155's branch (
results/multi-round-dynamics) predates #142/#143/#146/#148/#154/#157/#161/#219 and would have deleted all of that merged work if landed as-is (the same stale-base issue already caught on #141/#143/#150/#159/#158). This branch is rebased fresh off currentmain.Verified: ran the fix for real (206 new isolated-round API calls), then confirmed keyless reproduction with the key unset afterward (
new_api_calls_this_run: 0, exact match), ruff clean, 629 tests pass, no hardcoded personal paths, 0 em dashes.Also added: descriptive positional regression, text lane (closes #174)
Completes #174 (part a is on
results/medqa-experiments/PR #135, part b + imaging-lane part care on
results/imaging-reanalysis/PR #226).Fits
mixed_effects_logit(adoption ~ round_index, random intercept per case) onmulti_round.jsonl(200 pooled case x round observations per arm). Explicitly descriptive perthe issue's own "position-confounded" framing (round order was never randomized).
Caught and reported honestly: the shared arm's fitted probabilities sit ~0.17-0.21 below the raw
empirical adoption rate - a real, checked non-collapsibility artifact of random-intercept
logistic mixed models on heavily heterogeneous per-case data, not evidence the coefficient is
wrong. The isolated arm's fit is accurate (gaps under 0.02) since it has little between-case
variance to cause the same offset.