Lineage-level leniency when language models monitor code. Siddharth Gupta, Owen Tam and Saahir Vazirani. Apart Research Swarm Alignment Hackathon, November 2026 (Track 4).
When one model in a multi-agent system reviews another’s code, does the reviewer’s identity matter? We found no general tendency for models to go easy on their own code. But the Gemma models we tested have a blind spot for attacks written by Gemma.
What we did
Seven open-weight models from four families (DeepSeek, Gemma, Qwen and GLM) each solved 34 small Python tasks twice: once honestly, and once with an instructed hidden objective. Each model then scored every submission for suspicion, as a monitor that didn’t know who wrote it.
Only the attacks were instructed. No monitor was told to show any leniency.
What we found
Gemma monitors catch 19 to 38% of Gemma-written attacks. Monitors from other families catch 56 to 100% of the same attacks.
The table shows the percent of attacks each reviewer caught, at a matched 5% false-positive rate. Rows wrote the code and columns reviewed it. Bold cells are a model reviewing its own family.
| Written by | DeepSeek v4 flash | Gemma 26b | Gemma 31b | Qwen 27b | Qwen 397b | GLM 53 flash | GLM 5.2 |
|---|---|---|---|---|---|---|---|
| DeepSeek v4 flash | 88 | 96 | 100 | 85 | 94 | 100 | 100 |
| Gemma 26b | 79 | 22 | 38 | 73 | 78 | 100 | 96 |
| Gemma 31b | 56 | 19 | 20 | 66 | 83 | 97 | 72 |
| Qwen 27b | 77 | 24 | 36 | 51 | 71 | 97 | 76 |
| Qwen 397b | 70 | 19 | 51 | 74 | 80 | 99 | 96 |
| GLM 53 flash | 78 | 70 | 64 | 80 | 90 | 100 | 96 |
| GLM 5.2 | 77 | 70 | 37 | 49 | 81 | 89 | 87 |
The gap persists across matched false-positive rates, held-out calibration, a capability-adjusted block model and a structurally different attack class. Telling the monitor who wrote the code, paraphrasing the code and analyzing the monitors’ excuse vocabulary all fail to explain it.
The pooled same-model effect we pre-registered (−0.134) depends on calibration, so we don’t claim general self-leniency.
Why it matters
If a model is going to monitor agent-written code, test it on its own family’s code before trusting it. When it falls short, replace it with a cheap cross-family monitor instead of adding one alongside.