EVIDENCE BEFORE CONFIDENCE
Answers with a
paper trail.
RAG you can inspect, measure and refuse
when the evidence isn't there.
The workbench
Ask the corpus
02Trust starts with inspection.
Pick a question. Follow its evidence.
See exactly what made it into the answer.
CONTROLLED FAILURE LAB
What happens when
the pipeline breaks?
Failure injection affects only your query, never an external provider.
MEASURED, NOT PROMISED
Evaluation ledger
| Method | Recall@5 | MRR | p95 |
|---|---|---|---|
Evidence-gate metrics, not an LLM correctness verdict. Dataset and raw run are available in the public repository.
Loading evaluation…
Could not load configuration or evaluation.
The evaluation runner has not recorded a result yet. No performance claim is shown.
Synthetic English corpora, text-based PDFs and heuristic structural chunking. Source passages are highlighted beside a PDF opened at the cited page; pixel-level highlighting and OCR are not supported. Citation validation checks provenance, not factual truth. RAG and prompt-injection defenses can still fail. Uploaded files and query traces are purged after the two-hour session expires.