Security Charter Effectiveness in Large Language Model Code Generation: A Multi-Phase Experimental Analysis Revealing Task-Dependent Responsiveness and Architectural Differences
Bibliographic record
Abstract
This study introduces regression discontinuity design to LLM security evaluation, analyzing charter effectiveness across 858 trials with Claude-3.5-Sonnet and GPT-4o. We discover that security charter responsiveness operates independently from baseline model performance: while GPT-4o's overall scores dropped 12.58 points between experimental phases, its sensitivity to security guidance increased dramatically through optimization ($12.8 \times$effect size improvement from$\text{d = 0. 1 9 1}$to$\text{d = 0. 3 4 6}$). Claude maintained stable performance with consistent charter responsiveness (+6.18 points,$\mathrm{p}=0.030$). Task-specific analysis reveals both models respond strongly to charters on 4-5 out of 8 vulnerability domains ($d \geq 0.8$), effects completely hidden in aggregate measures. Strategic placement comparison shows embedded and late charter positioning outperform early placement across models. Despite achieving perfect security compliance (no vulnerabilities across$\text{8 5 8}$trials), charter influence operates through security practice enhancement rather than vulnerability elimination. Our findings demonstrate that charter effectiveness depends critically on task characteristics and model architecture, with single outlier tasks capable of masking significant intervention potential. These results provide the first causal evidence that security guidance and model capability represent distinct architectural systems in LLMs.
Fetched live from OpenAlex and de-inverted. Abstracts are not stored in this database: the inverted indexes are 8.6 GB of the frame’s 9.3 GB of text, and the host has 13 GB free.
How this classification was reachedexpand
Full frame machine prediction
Teacher imitationNot calibrated prevalence, not ground truth. Human validation pending. The Gemma side is a direct model label for every work in the frame, read from the title-only record. The Codex side is a classifier learned from the 10,348 direct Codex labels and calibrated to design-weighted sample rates; fields without enough sample support carry no Codex call. Candidate is the union of the two sides; consensus is their intersection. These outputs are machine_predicted_unvalidated and are not human labels.
Distilled classifier scores by category (both heads)
| Category | Codex | Gemma |
|---|---|---|
| Metaresearch | 0.018 | 0.115 |
| Meta-epidemiology (narrow) | 0.001 | 0.001 |
| Meta-epidemiology (broad) | 0.001 | 0.001 |
| Bibliometrics | 0.001 | 0.000 |
| Science and technology studies | 0.000 | 0.001 |
| Scholarly communication | 0.002 | 0.003 |
| Open science | 0.001 | 0.002 |
| Research integrity | 0.001 | 0.002 |
| Insufficient payload (model declined to judge) | 0.002 | 0.001 |
Machine scores (provisional)
The two teacher heads of the student model, read on this work. A score orders the frame for review; it never asserts a category, and the validation status ships verbatim with every row.
Baseline scores from an immature model (maturity gate not passed, 7 training rounds). Scores rank; they never assert a category.
score_only:v0-immature-baseline · verbatim from the scoring run: score_only means the number may rank works, and no category label ships from itClassification
machine, unvalidatedMachine predicted; a candidate call from one source (direct Gemma or distilled Codex), not a consensus.
How this classification was reached, model by model and score by score, is at the end of the page under "How this classification was reached".