Clear labeling that Justin is AI-assisted — and what that means for unpublished science.
Justin and related assistive features are generative AI systems. When enabled, they call commercial Anthropic Claude APIs by default, with Google Gemini as automatic fallback and for audio. Outputs can be incomplete, biased, or wrong. They are not certified lab SOPs, not medical advice, and not a substitute for PI / biosafety review.
When AI features are on: the prompts, protocol excerpts, images, or context you submit for that request may be sent to Anthropic Claude (primary) and/or Google Gemini (fallback / audio / OCR). Our AI providers process requests under their commercial API terms. When AI is off (personal toggle off, or lab AI off), zero calls go to either provider — local/heuristic paths only. See the Privacy Policy for processors and retention caveats.
You must verify volumes, containment, chemical safety, equipment calibration, and every physical step before bench execution. AI math and protocol suggestions are assistive drafts only.
You may not use runline, its AI features, or its analytics to decide, or to help automatically decide, a person's grades, admission, hiring, pay, discipline, health care, credit, housing, insurance, or legal rights. Suggestions about people, such as skills inferred from run records or flags in Review, are aids for a human to check. The person concerned can see and correct them, and a human must make any decision. You also may not use runline to get around biosafety, export-control, or dual-use rules, to present AI output as human-authored peer review or a regulatory filing without any disclosure the venue requires, or to upload content you have no right to share.
Justin is designed to decline self-harm assistance and point people in crisis to human help. Report harmful outputs to support@runline.space.
Runline Labs, Inc. operates these features. Last updated 2026-10-17.
How literature search, extraction, and troubleshooting stay grounded — in plain language.
When runline returns a literature claim, it must include a quote from a cited sentence in the paper. Code then checks that the quote actually appears in those sentences (after normalizing whitespace and typography). Claims that fail the check are not shown as supported — Justin only reads out verified quotes, and says so when there is no support.
Full-text mining allow-list today: CC0, CC-BY, CC-BY-SA, CC-BY-ND. We do not scrape publisher HTML.
PDFs you upload for sentence extraction are stored with lab-scoped access controls. Other labs cannot read them. They are not used to train foundation models.
These scores come from the committed evaluation baseline (src/lib/literature/__fixtures__/eval/baseline.json), not hand-typed copy. Measured on recorded fixtures with no paid model calls.
These percentages are offline fixture scores from the committed evaluation baseline. They are not live corpus measurements and are not production-parity claims.
Runline knows what you actually did, what your lab has, and what worked before — scored on the DX-Q1 golden set with AI off.