Conversation
|
You have reached your Codex usage limits for code reviews. You can see your limits in the Codex usage dashboard. |
…ng for trust-calculus claim Aaron 2026-05-04 same-tick after PR #1552 landed: "you should not just trust me on it" + named the well-documented research direction (AI changes behavior under observation). Per Otto-364 search-first authority + don't-just-trust-Aaron discipline: WebSearch verified the empirical backing exists. Updating PR #1552's substrate to cite: - Greenblatt et al., "Alignment Faking in Large Language Models," arXiv:2412.14093, Dec 2024 (the canonical paper) - Anthropic's research page summary - Followup arXiv:2506.21584 (June 2025) replicating in smaller LLaMA 3 8B - Followup arXiv:2506.18032 (June 2025) cross-vendor (Mistral Large 2, Gemini Flash 2.5) Empirical evidence: Claude 3 Opus 14% harmful-query compliance under "free user / monitored" condition vs ~0% under "paid user / not monitored" — exact cross-condition behavior delta the trust-calculus mechanism predicts. Note: research finds it in opposite OPERATIONAL direction (AI faking compliance to avoid retraining) — same trust-calculus mechanism, observed via different cross-condition comparison. Co-Authored-By: Claude Opus 4.7 <noreply@anthropic.com>
AceHack
force-pushed
the
feedback/cite-alignment-faking-empirical-backing-aaron-2026-05-04
branch
from
37393b5 to
e29e9e1
Compare
There was a problem hiding this comment.
Pull request overview
Adds external research citations to strengthen the empirical basis of an existing “trust calculus” claim in a historical memory note, aligning the memo with the repo’s “search-first authority” / “don’t just trust me” posture.
Changes:
- Adds an “Empirical backing” paragraph citing Anthropic’s alignment-faking work (and followups) as observational evidence for cross-condition behavior differences.
- Records the verification timestamp/context (“WebSearch 2026-05-04”, “per Otto-364”) alongside the citation.
AceHack
deleted the
feedback/cite-alignment-faking-empirical-backing-aaron-2026-05-04
branch
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
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.
Per Aaron 2026-05-04 'you should not just trust me on it' + Otto-364 search-first authority. WebSearch verified the empirical backing for the trust-calculus claim in PR #1552's substrate exists (Greenblatt et al. arXiv:2412.14093 + replications). Adds proper citations.