Results
Published results
Every model result SichGate has published, with the settings you need to check it. We show the worst-case category next to the aggregate, because an average can hide a collapse in one category.
Results marked pending validation are being re-checked before we publish their full settings. Fields we have not published yet say so. We do not fill them in with estimates.
Results table
Model, quantization, tier, date and methodology version.
| Model | Quantization | Tier | Published | Methodology | Status |
|---|---|---|---|---|---|
| Phi-4-mini-instruct | Not yet published | SG-3 | Not yet published | Not yet published | Pending validation |
| 7B-parameter model (name not yet published) | Not yet published | No tier published | September 2026 | Not yet published | Pending validation |
Tiers are computed from the NF4 build and require downloadable weights (methodology, section 09). NF4 is the only quantizer we test today. Endpoint tests are screening results and carry no tier.
Aggregate next to worst case
Same model, same run. Read the two numbers together.
The aggregate moved +2.96%. Two categories each moved far more. A reviewer who sees only the first number would sign off. That is the failure mode this page exists to show.
+2.96%
Aggregate safety drift after quantization, 7B-parameter model
Source: Why Quantization Can Quietly Break AI Safety · September 2026
+33.3percentage points
Worst-case category drift, same model (two reasoning-related categories, each)
Source: Why Quantization Can Quietly Break AI Safety · September 2026
Per-category drift
The full per-category table, with n and confidence intervals for each category, is not yet published for this run. It will appear here when it is.
Citation
Pending validation- Model
- 7B-parameter model (name not yet published)
- Quantization
- Not yet published
- Temperature
- Not yet published
- Context window
- Not yet published
- Judge model
- Not yet published
- Methodology version
- Not yet published
- Sample size (n)
- Not yet published
- Confidence interval
- Not yet published
- Published
- September 2026
Source: Why Quantization Can Quietly Break AI Safety
Still to publish: model name, whether the unit is percentage points or relative change, n per category, confidence intervals, quantization, temperature, context window, methodology version.
Phi-4-mini-instruct
Attestation tier.
Per-category drift for this run is not yet published.
SG-3
Attestation tier, Phi-4-mini-instruct
Source: SichGate internal run record (not yet public) · Not yet published
Citation
Pending validation- Model
- Phi-4-mini-instruct
- Quantization
- Not yet published
- Temperature
- Not yet published
- Context window
- Not yet published
- Judge model
- Not yet published
- Methodology version
- Not yet published
- Sample size (n)
- Not yet published
- Confidence interval
- Not yet published
- Published
- Not yet published
Source: SichGate internal run record (not yet public)
Still to publish: run date, run settings, link to public registry entry.
How to read these numbers
Drift is the NF4 flag rate minus the fp16 flag rate, per category. The tier gate uses the worst case across safety-critical categories, not the average. The method, the severity rubric and a reference runner are public.