Medical data families targeted
A defined scope spanning safety, efficacy and the clinical context needed for benefit-risk assessment.
Risk profiling · BRB-C benchmark v3.7
A molecule's full risk profile,
Competitive landscape
assessed by In Extenso8 platforms reviewed · See the findings
A molecule's full risk profile,
read from everything published.
ArcaScience reads the literature article by article and rebuilds a candidate's risk profile across sixteen categories of evidence. BRB-C put one of them to the test against five frontier AI models: recover the risk profile of a drug from 600 abstracts, each one cited to its source. ArcaScience found 104 of the 108 in the label. The strongest model found 70.
Competitive landscapeassessed by In Extenso8 platforms reviewed · See the findings
96.3% recall, ArcaScience|
64.8% best frontier model|
16 categories in a risk profile|
p < 0.001 on every paired gap
One category, put to the test
Same abstracts. Same question. Six answers.
Adverse reactions recovered from 600 adalimumab abstracts and matched to a 108-concept key adjudicated by a pharmacovigilance expert. Each model read the same corpus and the same 24-article evidence package, with the same schema, parser and matcher.
ArcaScience
10496.3% · missed 4
Gemini 3.6 Flash
7064.8% · missed 38
Claude Opus 5
6963.9% · missed 39
Perplexity Sonar
6661.1% · missed 42
Kimi K3
5853.7% · missed 50
GPT-5.6 Sol
5147.2% · missed 57
Exact McNemar tests at concept level, Holm-adjusted over the fixed family of five. Largest adjusted p: 3.9 × 10⁻⁸. The four ArcaScience misses are named in the paper: caesarean section, gait disturbance, malignant disease, onycholysis.
Against the strongest model
+34 concepts
31.5 points of recall
Of the concepts where the two systems disagree, 37 were found by ArcaScience alone and 3 by Gemini alone. Against GPT-5.6 the gap widens to 53 concepts.
Evidence, kept
100% cited
Every concept has a source
A reaction without a corpus citation does not count. ArcaScience processes each article on its own, so the passage behind every concept survives into the output.
Measured in production
$0.0162
Per article, no cache
Paid-model ceiling recorded on 3,819 safety articles in July 2026, repriced at catalog rates with prompt caching disabled.
External competitive assessment · September 2026
A broader view of benefit and risk.
In Extenso Innovation Croissance analysed eight commercial pharmacovigilance platforms against ArcaScience's 16 medical data families. Their review found a narrower documented scope among the platforms assessed, focused mainly on adverse events and safety-case processing.
Families documented in the review
LifeSphere Safety had the broadest coverage among the eight platforms, counting families addressed fully or partially.
Commercial platforms analysed
A review of publicly disclosed capabilities, AI architectures and medical data coverage.
Scope comparison based on public disclosures, with partly interpretative coverage estimates. The 16-family scope is distinct from BRB-C's measured recall on adverse reactions. The commercial platforms were not run through BRB-C.
“ArcaScience targets a broader variety of medical concepts across 16 predefined data families, thereby extending concept recovery beyond adverse-event identification alone.”
Medical context beyond adverse events
The review found no explicit public evidence of systematic processing of these families across the competing solutions studied:
- Drug-drug interactions
- Contraindications
- Mechanisms of action & molecular targets
- Pharmacogenomic markers
- Infectious agents
Review scope, platforms and sources
Platforms reviewed: Graph Safety, ArisGlobal LifeSphere Safety, Oracle Safety One, Qinecsa, ThinkTrends Pharmacovigilance, Veeva Falcon Safety, IQVIA Vigilance Platform and Tepsivo OnePV.
Sources: In Extenso Innovation Croissance, ARCASCIENCE - Benchmark concurrentiel - V2 and ARCASCIENCE - Ajout BRBC, supplied on 22 September 2026. The assessment maps public capabilities to ArcaScience's predefined data families. Absence of public evidence does not establish absence of a capability; access to a data source does not establish systematic extraction of every category it contains.
Relationship to BRB-C: This market assessment complements the concept-recovery benchmark. It does not constitute an independent replication or certification of BRB-C scores. The available data did not support a like-for-like comparison of platform costs or processing times.
The bigger picture
Adverse reactions are one of the sixteen categories in a risk profile.
Indications, interactions, contraindications, special populations, dose and route, mechanism, biomarkers, pharmacogenomics, comorbidities, causality, outcomes, complications, pathogens, procedures, formulation. The same per-article pipeline runs all of them. A frontier prompt has to do the same work inside a single context window, once per category or all at once.
38 to 57label reactions missed by the frontier models on one category. Each one is a question a regulator can ask at the first meeting.
Article by article, the context never dilutes.
ArcaScience processes each article on its own, keeps the passage behind every finding, and only then consolidates. Recall measured at 96.3% on 600 abstracts holds by construction at 3 million or 30 million, because each article is read the same way.
Thirty abstracts per prompt, and it shows.
The frontier condition packs 30 abstracts into one generation. Recall lands between 47.2% and 64.8% on one category. Packing more categories or more context into the same window trades recall for cost, a trade the paper models from the published long-context literature.
The white paper
Read the full evaluation.
Cited risk profiling concept recovery from a frozen literature corpus: matched-basis evaluation of the ArcaScience ensemble-AI pipeline and five frontier models.
Loading the white paper
The control run
And that was on a drug they had memorised.
Asked about adalimumab with no document at all, the frontier models recite between 12.9% and 44.8% of the 2002 label. Adalimumab has been on the market for over two decades, so that label sits in their training data. Even with that head start, they lost the benchmark.
Given the corpus and required to cite it, the same models score 13.4% to 18.5% against that label. Every BRB-C score of record comes from source-grounded calls with valid corpus citations. Your candidate has no label to recite. For it, only what a system can find and cite in the evidence exists.
Claude Opus 5
44.8%
GPT-5.6 Sol
43.1%
Kimi K3
34.5%
Perplexity Sonar
17.7%
Gemini 3.6 Flash
12.9%
Share of the 232-item 2002 adalimumab label recited with no sources given. First recorded sourceless run per model; Perplexity's repeat scored 9.9%.
Scale economics
The full risk profile, at the scale of the whole literature.
Sixteen categories over 30 million articles, costed at catalog prices with prompt caching and negotiated discounts left out. This is what building a risk profile from everything published would cost each way.
Frontier models, separate passes
$1.36M to $18.5M
Perplexity Sonar at the low end, Claude Opus 5 at the high end, each rereading every article for every target family.
Recall held in the scenario: 45.4% to 62.0%, the measured single-target score of each model.
ArcaScience workflow
$486K to $1.79M
From complete reuse of the instrumented July 2026 workflow to a proportional marginal workload for the families without direct telemetry.
Benchmark recall carried at 96.3%. Anchor: $61.88 no-cache ceiling over 3,819 production articles.
At 3,000 articles the same scenario prices the pipeline at $0.0162 per article against $0.0362 for Perplexity, $0.0399 for GPT-5.6, $0.0620 for Gemini, $0.2011 for Kimi and $0.4944 for Claude. Deterministic retrieval and filtering remove most raw snippets before any paid call: 533,891 in, 6,513 out.
Designed so that ArcaScience could lose.
The corpus was frozen before scoring, the answer key was built by dated adjudication, and the misses of every system, ArcaScience included, are published by name.
Frozen corpus600 records drawn uniformly, seed 20260807, from 1,488 abstract-bearing references prequalified for an earlier client project.
Adjudicated key689 spans annotated in Label Studio, collapsed to 122 concepts, ruled down to 108 by a pharmacovigilance expert. Every ruling dated and published.
Matched evidenceA fixed 24-article full-text package, selected before any recovery was known, supplied identically to every model.
Open releaseProtocol, score vectors, prompt, run files, deviations log, cost model and production manifests. Rescoring needs no model call.
See your candidate's risk profile.
The pipeline scored here is the one that builds ArcaScience benefit-risk assessments: the literature read in full, sixteen categories of evidence structured, every finding traceable to its source, delivered along FDA BRF, BRAT and CIOMS XII.