01Models we use
SQEase routes different stages through several specialist providers, picked for accuracy on each task.
| Provider | Models | Used for |
|---|---|---|
| Gemini 3.1 Pro & Flash | Document analysis, question drafting, and verification. | |
| Anthropic | Claude Sonnet & Haiku | Question drafting, mathematical checks, and tutor explanations. |
| OpenAI | GPT-5.5 | Independent review of every drafted question. |
| Tavily | Web research | Extensive legal research grounded in real, current sources. |
02The 5-stage pipeline
- The Professor — reads source material and extracts subtopics.
- The Exam Writer — drafts the SBA question with numerical verification.
- The Critic — independent peer review.
- The Fact Finder — live web search verification.
- The Judge — final synthesis. Earlier drafts are discarded.
03Verification gate
A question failing any of these verification layers is rejected:
- Knowledge-base cross-check
- Document-grounded reasoning trace
- Mathematical verification via Claude Sonnet
- Live UK law search
- Targeted deep search on ambiguity
- Comprehensive escalation for high-stakes claims
04Known limits
- Hallucination risk — while the pipeline isn't proven to catch 100% of hallucinations, we put massive effort into ensuring questions are factually, legally, and mathematically correct, with multiple layers mitigating hallucination.
- Time lag — 24–72 hours to fully reflect legislative changes.
- Edge cases — we conduct extensive research with Tavily web research grounded in real resources. Every possible effort is taken to make questions as accurate as possible, and unverified questions are not published, though there is a slight possibility of minor inaccuracies.
- Handwriting OCR — accuracy varies with clarity.
- No legal advice — nothing on SQEase constitutes legal advice.
05No guarantee of results
06Your data, our knowledge base & AI training
What we don't do
- Your document content is never copied into our knowledge base or used as training data.
- Third-party providers operate under contractual undertakings that your input is not used to train foundation models.
What we do
We extract topic labels from your documents and use them as prompts for independent legal research. That research — our work, not your content — populates our knowledge base and fine-tuning data.
07Liability
- SQEase accepts no liability for errors in AI-generated content.
- Nothing excludes liability that cannot lawfully be excluded.
See our Terms & Conditions §10.
08Reporting an error
Use the in-app "Report this" button on every question. Reports are re-verified within 48 hours.