task #550: third dial point at band [9,13] nat#430
Draft
superkaiba wants to merge 3242 commits into
Draft
Conversation
## Results — task #534 (4-fraction b [task.py]
[task.py]
…gent=upload-verif [task.py]
[living_docs.py]
… leak-predictor, leak-be [task.py]
[living_docs.py]
… leak-predictor, leak-be [task.py]
…entation v13 --> ## Implementation [task.py]
…98 phase=done after 122s. [task.py]
…arent #521 clean [task.py]
[task.py]
…sarial planning. [task.py]
[task.py]
[task.py]
[task.py]
## Upload Verification [task.py]
…entation v1 --> ## Implementation [task.py]
…ent=experimenter [task.py]
…de change 1 (disp [task.py]
…minated after upload-ve [task.py]
[task.py]
…ubagent=analyzer [task.py]
…lore-persona- [task.py]
[task.py]
…onomous plan-gate (tas [task.py]
Generated with [Claude Code](https://claude.ai/code) via [Happy](https://happy.engineering) Co-Authored-By: Claude <noreply@anthropic.com> Co-Authored-By: Happy <yesreply@happy.engineering>
… contract holds: the [task.py]
…n.md (gpu_hou [task.py]
[task.py]
…onomous plan-gate (tas [task.py]
[task.py]
…f-response mar [task.py]
## Code-Reviewer Verdict — PASS [task.py]
…reshold extended [task.py]
…indings are cleanly closed by [task.py]
[task.py]
[task.py]
|
Deployment failed with the following error: Learn More: https://vercel.com/thomasjiralerspong-5386s-projects?upgradeToPro=build-rate-limit |
[task.py]
…inal reviewer step; abs [task.py]
…issue follow-up into th [task.py]
[task.py]
…dl phase=done after 182s. [task.py]
…ubagent=experimen [task.py]
…de-review v5 PASS [task.py]
…d preflight behind-origin/main gate
|
The latest updates on your projects. Learn more about Vercel for GitHub.
|
Synced from pod-side eval_results/issue_550/ (analysis, eval emission+shift JSONs, sweep, anchor_smoke) after HF public-storage quota 403 blocked the HF-side uploads; makes the science durable in git regardless of quota. Co-Authored-By: Claude Opus 4.7 <noreply@anthropic.com>
Adds figures/issue_550/bystander_slope_by_persona.{png,pdf,meta.json}
plus scripts/issue550_bystander_slope_figure.py. Re-reads the shift
JSONs across the three dial points (#527 / #550 / #538), computes the
bystander-only per-persona slope (source-persona reads excluded), and
plots one dot per persona sorted descending by slope. Source-adjacent
personas (medical / security / investigation) form the high-slope
tail; semantically distant personas form the low-slope tail. The
median bystander slope sits at ~0.96 nat per nat of dial; the spread
is ~1.8x.
Anchors the analyzer's round-2 fix for clean-result-critic Lens 9
(quantitative finding needs one inline figure).
superkaiba
added a commit
that referenced
this pull request
Jun 10, 2026
Branch issue-550 is 2296 commits behind origin/main, unsafe to blind-rebase (carries stale copies of ~50 foreign tasks/ folders). Cherry-picked this task's own added/updated files only: patched analysis.json (band metadata 14/20 -> 9/13), the new bystander-slope figure (3 files), and its task-namespaced figure script. Shared src/ / scripts/run_issue538_train.py left for the deferred full rebase; PR #430 stays open. Generated with [Claude Code](https://claude.ai/code) via [Happy](https://happy.engineering) Co-Authored-By: Claude <noreply@anthropic.com> Co-Authored-By: Happy <yesreply@happy.engineering>
…nding 3 follow-up)
Free-analysis follow-up: tests whether per-persona bystander-only slope of
marker log-prob shift across the dial tracks each persona's base-model
representation distance to the four source personas.
Distance metric reuses issue_527/pair_selection.json::cos_centered_L20
(extraction_layer=20, global-mean centering — the body's existing pair-
selection metric). Slopes reuse build_bystander_data +
compute_persona_slopes from scripts/issue550_bystander_slope_figure.py
(bystander-only convention, two-point endpoint over landings
5.89 → 17.24 nat).
Spearman ρ (slope vs distance):
mean-over-sources, n=19: -0.391 (p=0.098)
min-over-sources, n=19: -0.181 (p=0.459)
mean, drop sources, n=15: -0.311 (p=0.260)
min, drop sources, n=15: -0.161 (p=0.567)
Negative correlation is in the predicted direction for source-persona
generalization (closer to sources => larger slope), but only the
mean-over-sources reading reaches a marginal n=19 trend; descriptive at
this n. min-of-sources is a weaker predictor (assistant reads very close
to one source but has the second-lowest slope), consistent with the
mechanical-spread reading being a co-driver.
Outputs:
- scripts/issue550_slope_distance_correlation.py (new)
- eval_results/issue_550/analysis/slope_distance_correlation.json
- figures/issue_550/slope_vs_distance_{mean,min}.{png,pdf,meta.json}
Co-Authored-By: Claude Opus 4.7 <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
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 task #550. Single-variable change vs #538: band-stop window [14,20] → [9,13] nat. Code review: ensemble PASS (reconciled). Base = issue-538 (parent pipeline branch).