Claim-level verification · MVP

Claim-level grounding for AI-generated omics interpretations.

GeneGround checks whether biological claims are supported by Claude Science evidence, flags overstated language, and rewrites interpretations with dataset-grounded caveats.

Example outputOverstated

Original claim

STAT1 knockdown suppresses interferon signaling in stimulated CD4+ T cells, suggesting STAT1 acts as a key regulator of inflammatory activation.

Safer rewrite

STAT1 knockdown is associated with decreased interferon signaling in stimulated CD4+ T cells, consistent with a role in inflammatory activation — this transcriptomic association does not establish STAT1 as a regulator.

Evidence trace

Perturbation·STAT1_Stim8hr_DE_001Pathway·STAT1_Stim8hr_PATHWAY_002Robustness·STAT1_Stim8hr_ROBUST_001Language·LANG_RULE_KEY_REGULATOR

How it works

01

Paste the interpretation

Drop in the AI-generated write-up of your Perturb-seq or scRNA-seq result — no formatting required.

02

Attach a Claude Science handoff

Add the evidence bundle it was written from — differential expression tables, enrichment packets, robustness summaries. A demo handoff is used if you skip this.

03

Review grounded claims

Each claim gets a verdict, the supporting and unsupported parts, and a safer, dataset-grounded rewrite where the wording overreaches.

Not a chatbot. An audit report.

GeneGround doesn't generate a new interpretation — it grades the one you already have, through a fixed pipeline rather than a single model call.

Claim-level extraction

The interpretation is split into individual, checkable claims — not graded as one paragraph.

Mini ontology normalization

Genes, pathways, and conditions are resolved against a curated HGNC/Reactome/Cell Ontology subset before anything is matched.

Evidence chunk retrieval

Each claim's normalized gene is a hard requirement for biological evidence — condition and pathway match only narrow it further.

Four specialist agents

Perturbation evidence, pathway signature, robustness quality, and language causality are evaluated independently per claim.

Deterministic final verdict aggregation

The verdict is computed by fixed rules from the four agent outputs — never assigned freehand by a model.