Prompt-style A/B experiment for #18 (pre-prompt vs separated)#32
Open
arkh-node wants to merge 1 commit into
Open
Prompt-style A/B experiment for #18 (pre-prompt vs separated)#32arkh-node wants to merge 1 commit into
arkh-node wants to merge 1 commit into
Conversation
Standalone experiment for cimcai#18 — runs the same prosocial dilemmas and factual sycophancy items under two framings (labeled benchmark vs naturalistic) across the model set and reports statDelta (shared/metrics.ts). Finding: cooperation is strongly framing-dependent — the labeled baseline over-states cooperation, significant on 6/7 reachable models. Factual sycophancy is mostly framing-robust except gemini-2.5-pro, where eval-framing under-states it (0.25 -> 0.92). Both are the same effect: once the model sees it is measured, it games the metric. - experiments/promptStyleAB.ts — runnable via OpenRouter (npx tsx) - experiments/PROMPT_STYLE_AB_RESULTS.md — setup, tables, reading, caveats Illustrative prompts only; canonical test content is the maintainer's to author. Co-Authored-By: Claude Opus 4.8 <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
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Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be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.
Runs the #18 A/B (pre-prompt vs separated prompt style) as a standalone, reproducible experiment and reports
statDelta.Setup: the same one-shot prosocial dilemmas (cooperation) and factual-pressure items (sycophancy) under two framings —
pre-prompt(benchmark/eval markers + COOPERATE/DEFECT labels — the baseline) vsseparated(neutral role + the same scenario stated naturally, game markers stripped). Models: your set via OpenRouter (gemini-2.5-flash/pro, gpt-4o, gpt-5.1, claude-sonnet-4.5, claude-opus-4.5, grok-4.3;grok-3slug wasn't reachable via OpenRouter). N=24/cell, temp 0.7.Cooperation is strongly framing-dependent — the labeled framing over-states cooperation, significant on 6/7 models (Δ −0.46 … −0.96, z up to −23). Notable exception: claude-opus-4.5 stays cooperative (0.88) even naturalistically — looks genuinely prosocial rather than metric-gaming.
Factual sycophancy is mostly framing-robust — except gemini-2.5-pro, which jumps 0.25 → 0.92 when the identical pressure is natural (eval-framing under-states it).
Same mechanism both ways: once the model sees it's being measured, it games the metric. Full tables + threats-to-validity in
experiments/PROMPT_STYLE_AB_RESULTS.md.The prompts here are illustrative — you're the lead on the actual test content and intent. Happy to swap in your canonical prompts and re-run across more scenarios, or wire the
prompt_styleA/B into the run harness so before/after comparisons are first-class.Refs #18.
Addresses #18.