Skip to content

~95% on SimpleQA (e.g. Qwen3.6-27B on a 3090). Supports all local and cloud LLMs (llama.cpp, Ollama, Google, ...). 10+ search engines - arXiv, PubMed, your private documents. Everything Local & Encrypted.

1.6 KB
Raw Blame

Security contract registry — seam → test file → PR

Seam Contract test PR
GET /benchmark/api/results grader_response CWE-209 boundary (failure-prefixed raw exception detail genericized, including the historical 🔑-decorated prefix; normal/NULL pass through) tests/web/routers/test_benchmark_results_grader_response_boundary.py #7591
journal-quality dataset fetch record-count ceilings (hostile-upstream bound on accumulated entry count; HostileSnapshotError raised inside the accumulation loop, before storing, for OpenAlex sources, institutions (ids and combined id/ROR/name-index entries, incl. per alternative name), DOAJ, JabRef, predatory lists (lazy CSV rows); fleet-sized payloads pass. Not covered: field lengths, whole-body buffering of DOAJ/JabRef/predatory responses, which only the transport size cap bounds) tests/journal_quality/test_snapshot_record_ceilings.py #7973
Chat research-start boundary is_safe_custom_llm_endpoint pre-flight (send_message + retry_attempt 400 before any row/worker; router-wide structural fence flipped from recorded xfail) tests/chat/test_chat_custom_endpoint_gate.py + tests/llm/test_provider_adapter_contracts.py::TestCustomEndpointBoundaryGate #7992
POST /metrics/api/domain-classifications/classify domain input bound (length capped at 255 — the DomainClassification.domain column's own declared length — and control characters rejected at the route boundary, ahead of prompt build/persistence/error echo; pass-through guards pin the 255 edge, ordinary domains, and batch mode) tests/web/routers/test_metrics_domain_input_bounds.py #7641