Ask.app
Ask the case studies
Ask the 64 case studies a question. Every answer is a verbatim quote with its source, and it abstains when the answer is not there.
Answers you can check
Ask.app searches 868 passages from the 64 case studies with BM25, in your browser. Every answer is a word-for-word quote with a link to where it came from. No model writes anything, so nothing can be made up. When the best passage doesn't cover the question, it says “Not in the case studies” instead of guessing.
Five real questions and what it answered
On the bundled benchmarks: 100% precision and recall on structured PII, and 83% accuracy telling a grounded answer apart from a hallucinated one.
From guardrail-gate, The Number
Six target-species examples aren't enough to learn methylation from scratch. Transferring from related species looked like chance-to-100%, but that was one lucky seed. Reweighted properly, it never does worse than training alone across 90 seeds.
From methyloadapt, Summary
Recent work demonstrates practical privacy leakage through shared LLM KV and prefix caches, and motivates tenant-aware isolation. Read the NDSS 2025 KV-cache leakage study .
From cacheisolate, Research Basis
Always-picking-the-top-tool fires a wrong tool on all 5 genuinely ambiguous queries in the 20-tool benchmark. Confidence-gated routing abstains on all 5, and it does that at zero cost: clear-query accuracy stays at 100% either way. Caution that costs nothing on the easy cases is the whole point.
From toolrouter, The Number
On the adversarial set, keyword retrieval sits at 62% citation precision whether or not you bolt dense fusion on top. Add reranking and it jumps to 88%.
From citebench, The Number