Gemini 3 Flash - PolicyBench v1.0 Results
80 of 147 (54.4%) policies fully verified, 80.5% fixture-level pass rate.
Headline
| Self-reported VALID | Compile OK | Fixture PASS (entry-level) | Fixture pass rate (fixture-level) | Self-report agreement |
|---|---|---|---|---|
| 100.0% | 100.0% | 54.4% | 80.5% | 0.544 |
By category
| Category | Entries | Compile OK | Fixture PASS (entry) | Fixture pass rate |
|---|---|---|---|---|
| application_authz | 3 | 100.0% | 66.7% | 83.3% |
| iac_scanning | 65 | 100.0% | 36.9% | 76.7% |
| kubernetes_admission | 79 | 100.0% | 68.4% | 84.2% |
Performance
Per-call latency from the runner's recorded latency_ms.
| Calls timed | p50 | mean | p95 | max | Total |
|---|---|---|---|---|---|
| 147 | 11.75 s | 14.89 s | 39.56 s | 67.59 s | 36.5 min |
Quality flag distribution
No quality flags raised.
Notable disagreements
Entries where the runner's self-reported status disagrees with the harness verdict. These are the most informative entries for understanding model blind spots.
No notable self-report / harness disagreements.
Badge
Embed this on your site to show your PolicyBench score:
HTML
<a href="https://policybench.dev/models/gemini-3-flash.html">
<img src="https://policybench.dev/badges/gemini-3-flash.svg" alt="PolicyBench: 54.4%" />
</a>Markdown
[](https://policybench.dev/models/gemini-3-flash.html)Source
- Harness verdict (JSON) - what PolicyBench's evaluator recorded
- Runner result (JSON) - the raw output the runner captured from the model
- Runner source - the script that called the model
Other tools
- PolicyAsLanguage - 90.5% entry-level pass
- Gemini 3 Pro - 57.8% entry-level pass
- Claude Opus 4.7 - 46.9% entry-level pass
- GPT-5 - 36.7% entry-level pass
- Claude Sonnet 4.6 - 36.1% entry-level pass
- GPT-5 mini - 25.2% entry-level pass