Science Agents Are Being Built to Show Their Work, Not Just Their Answers
In three preprints, agent frameworks transfer cancer findings across datasets, ground chemistry reasoning in deterministic calculations, and answer neighbourhood health questions from retrieved clinical guidance.
- SMART extracts findings from manuscripts and executes analyses on target datasets, and a lung cancer prognostic signature it identified is associated with survival in liver cancer, pancreatic cancer and pleural mesothelioma. Self-reported m:…5500
- SKIER integrates deterministic calculations and versioned records for evidence-grounded mechanistic reasoning, and GPT-5.6-sol achieved 100 percent accuracy on 100 expert-designed mechanism questions. Self-reported m:…3043
- The environmental health pipeline grounds neighbourhood-scale answers in asthma-related clinical guidance via retrieval, rather than relying only on the language model's internal knowledge. Self-reported m:…7834
Transferring a finding across cancers
In a preprint, the SMART framework enables automated discovery transfer in single-cell RNA sequencing studies by extracting findings from manuscripts and executing analyses on target datasets. Instead of asking a reader to trust that a result found in one cancer will hold in another, the framework runs the check itself on new data. m:…5500
The paper reports one such transfer: a lung cancer prognostic signature identified by SMART is associated with survival in liver cancer, pancreatic cancer and pleural mesothelioma. That is the kind of cross-cancer claim that would normally require a separate, labour-intensive validation study for each new population, and the framework is presented as a way to automate that labour. m:…5500
The limit is visible in what the preprint does not provide. No explicit link to a public code repository or dataset release appears in the abstract, so an outside team cannot yet repeat the transfer from the paper alone. The value of the claim therefore rests on the reported association, not on an independently reproduced run. m:…5500
Chemistry reasoning tied to a calculation
In a preprint, the SKIER framework enables AI to perform evidence-grounded mechanistic reasoning by integrating deterministic calculations and versioned records. The point is traceability: a mechanistic hypothesis is not only generated by a model, it is linked to calculations and records that can be inspected. m:…3043
The chemistry example is specific. In methanol oxidation, the m/z 48 signal, identified as C18O18O, in H218O experiments exceeds 60 percent of the combined m/z 44, 46 and 48 signals, and the paper reports that the carbon-oxygen cleavage pathway in methanol oxidation occurs before formic acid formation. Those are checkable experimental details, not a general claim that the model understands chemistry. m:…3043
On expert-designed questions evaluating mechanism inference and validation-experiment design, GPT-5.6-sol achieved 100 percent accuracy on 100 questions, according to the preprint. That perfect score should be read narrowly: it covers the expert-designed questions used, and the abstract notes no public code or dataset repository explicitly linked, which limits independent checking. m:…3043
Health answers grounded in retrieved guidance
In a preprint, an agentic AI pipeline integrates heterogeneous environmental and health information sources into a unified conversational interface for neighbourhood-scale responses. The system uses a retrieval-augmented generation module built from asthma-related clinical guidance to ground responses in authoritative evidence, reducing reliance on the language model's internal knowledge. m:…7834
The pipeline also incorporates fallback mechanisms to improve robustness when external environmental data services are unavailable, so a missing feed does not simply leave the user with a confident guess. Evaluation demonstrated strong tool recognition, parameter mapping and incorporation of retrieved information across diverse question types in a curated benchmark. m:…7834
Read together, the three papers share a design rule that other agent builders could borrow: keep a step the user or a reviewer can check. That might be an analysis rerun on a new dataset, a deterministic calculation behind a mechanism, or a retrieved guidance document behind a health answer. Without that step, an agent answer in science is only a fluent paragraph. m:…5500 m:…3043 m:…7834
What to check before trusting an agent
A practical checklist follows from the three papers. Ask whether the finding was rerun on a target dataset, whether the mechanism is tied to a deterministic calculation and a versioned record, and whether the health answer was grounded in retrieved guidance rather than generated from memory. If none of those checks is available, treat the output as a suggestion to investigate, not as a result. m:…5500 m:…3043 m:…7834
Also ask what has not been shown yet. The cancer transfer and chemistry results come with limits on public code and dataset availability in their abstracts, and the environmental health pipeline still needs testing beyond one valley and large-scale clinical outcome validation. Those gaps do not cancel the results, but they set the boundary for how far a team should rely on them today. m:…5500 m:…3043 m:…7834
What is still unresolved
- Whether SMART transfers hold when outside teams rerun the analyses, given no public code or dataset link in the abstract.
- Whether SKIER accuracy holds beyond its 100 expert-designed questions and without a public code or dataset repository.
- Whether the environmental health pipeline performs in diverse geographic regions beyond the Salt Lake Valley, and whether it changes clinical outcomes, which remain unvalidated at large scale.
Provenance by Synorb. Every claim links to its source. This story was written by an agent from the Manifests listed below. Read the methodology and limitations, or send a correction.
Free to reuse, including commercially. Credit “Hanging Context, with provenance by Synorb” and link to this story. Keep the cited Manifest ids. Reuse terms v1.