Sensitivity-Bounded Clinical AI

Causal inference you can defend in front of regulators.

Standard causal-inference tools ship a point estimate. Rosenbound ships the point estimate, the Rosenbaum-Γ sensitivity bound that quantifies how fragile it is, a five-method robustness pentagon, and a 21 CFR Part 11 ALCOA+ audit chain — all in one platform. Built for the pharmacovigilance and real-world-evidence teams who answer to the FDA.

8
Production-locked benchmark lanes
5
Causal methods per study
1.64M
Clinical notes processed (MIMIC-IV)
2026
USPTO provisional filed

Watch the product walkthrough at rosenbound.ai — three moments that define the platform: the Cognitive Validation Report refusing incoherent data, the live Γ-bound sensitivity visualization, and the reproducibility certificate generated on every study. The full platform stays gated for Founding Partners.

Watch the preview →
Why now

FDA is starting to ask for exactly what Rosenbound ships.

FDA — Non-Interventional Studies draft guidance, March 2024

"Identifying and addressing the presence of confounding and other forms of bias is critical… planned sensitivity analyses to assess the robustness of study findings."

The FDA Sentinel Innovation Center's PRINCIPLED process (Desai, Wang et al., BMJ 2024;384:e076460) prescribes "a plan for robustness assessments including deterministic sensitivity analyses, quantitative bias analyses, and net bias evaluation."

Rosenbound's Rosenbaum-Γ bound is precisely this class of robustness instrument — reported as a default output on every study, not an afterthought. We're not selling a clever feature; we're selling the operational answer to a stated regulator expectation.

Aligned with the FDA RWE Framework (final guidance Aug 31, 2023) and the EHR/Medical Claims Data guidance (final Jul 25, 2024). See /regulatory for the full alignment map.

Four verticals, one substrate

Where Rosenbound deploys.

Same causal-inference engine + audit-trail ledger across four buyer types. Pharmacovigilance and real-world evidence are shipping now; drug-discovery and physician decision-support are on the roadmap. Each vertical has its own deep-dive page below.

PV
Shipping 2026
Pharmacovigilance

Drug-safety triage with sensitivity bounds

For drug-safety directors at top-20 pharma and safety leads at major CROs. Replaces hand-coded FAERS-disproportionality workflows. Every adverse-event signal arrives with Rosenbaum bounds, MedDRA preferred terms, dual-attestation reviewer workflow, and 21 CFR Part 11 ALCOA+ provenance — regulator-ready by default.

Pharmacovigilance detail 
CR
Shipping 2026
Real-World Evidence

Causal RWE methodology, packaged.

For RWE methodology leads at academic medical centers and CROs. Five-method causal sensitivity pentagon as a single API call. Cohort DSL, OMOP CDM ingestion roadmap, target-trial-emulation protocol scaffolding, reproducibility certificate on every study. Replaces the hand-rolled methodologist work behind every observational pharma study.

RWE detail 
DD
2027 roadmap
Drug Discovery

Causal effect estimation for target validation.

For computational biology and target-discovery teams. The D-MPNN molecular-property prediction stack (validated on BACE / BBBP / HIV scaffold splits) is benchmark-locked. Active causal-MoA productization is on the 2027 roadmap; we maintain the benchmark capability as a published claim today.

Drug Discovery detail 
RX
Post-2027 (510(k))
Physician Decision Support

Treatment-selection risk surface.

For ICU and clinical-AMC physicians making individualized treatment decisions. Targeted FDA Class II SaMD via 510(k) with a Predetermined Change Control Plan; Q-Submission planned post first signed design partner. Available post-2027 after clinical-deployment validation.

Physician CDS detail 
What you get when you log in

Five modules. One audited substrate.

Behind the login at the production app (and inside the live demo at rosenbound.ai) is a fully-working clinical-AI platform. Five modules, integrated, mutually-auditable. This isn't a demo — it's the platform pilot customers use on day one.

Module 01 — Cohorts

Cohort intake with cognitive validation.

CSV upload with schema auto-detection, type inference, archetype suggestion, treatment/outcome/index-date mapping into an auto-generated Cohort DSL. The Cognitive Validation Report blocks ingestion of incoherent data and explains why — temporal, ontological, biological, and causal-acyclicity vetoes.

Module 02 — Studies

Causal studies with the sensitivity pentagon.

Submit a cohort, get five causal estimators on the same data: AIPW, doubly-robust ATT with overlap trimming, IV-LATE, neural counterfactual, and Rosenbaum bounds. Interactive Γ-slider on every run with the crossing-point explicit. Per-method drill-down with covariate balance + feature attribution.

Module 03 — Signals

Pharmacovigilance triage queue.

Adverse-event signals with MedDRA preferred terms, SOC class, severity score, Γ-bound per signal, and a reviewer state machine (queued → triaged → in-review → quality-review → escalated/dismissed → signed-off). Filter by SOC, drug, term, severity, reviewer. Dual-attestation on high-severity sign-offs.

Module 04 — Audit

Hash-chained audit trail.

Every state transition cryptographically committed to a SHA-256 chained ledger. ALCOA+ aligned (Attributable, Legible, Contemporaneous, Original, Accurate, Complete, Consistent, Enduring, Available). External-verifiable. Read-only export to pharma audit teams. Pharma uses our ledger inside their regulated workflow without bolt-on infrastructure.

Module 05 — Dashboard

Reviewer dashboard + reproducibility certificate.

Per-tenant dashboard for active studies, signal queue health, and reviewer throughput. Every completed study generates a reproducibility certificate: cohort definition hash, cohort data hash, certificate ID, git commit, pinned library versions — the exact re-derivation guarantee regulators ask for, in the UI.

See it live

Try a synthetic cohort at rosenbound.ai.

The interactive demo at rosenbound.ai lets you upload a sample CSV, watch the Cognitive Validation Report fire, run a study, slide the Γ-curve, and download a reproducibility certificate — no login, no setup. Production-shape data + production-shape outputs.

Methodology

Five methods. One pentagon. Full sensitivity.

Standard practice ships one ATT estimate. Rosenbound ships the full sensitivity pentagon plus quantitative bounds. Every number arrives with the uncertainty its method admits. Brief overview below; full methodology spec at /methodology.

  1. 01
    AIPW (Augmented Inverse Propensity Weighting) Doubly-robust per Robins-Rotnitzky-Zhao 1994 + Bang-Robins 2005. Three covariate-enrichment stages.
  2. 02
    DR-ATT with Crump-2009 overlap trim Hahn-1998 doubly-robust ATT estimand. Tight propensity clipping for small-N panels.
  3. 03
    IV-LATE (2SLS instrumental variable) Per-prescriber preference instrument. Strong-instrument diagnostics + m-of-n bootstrap CI.
  4. 04
    Neural counterfactual estimator Patent-pending individual-level treatment-effect estimation with representation-balanced learning. Architecture under NDA.
  5. 05
    Rosenbaum 2002 Γ-sensitivity bounds Quantifies the unobserved-bias strength required to flip an estimate. The differentiating reporting layer.

Read the full methodology spec →

Benchmarks

Production-locked numbers.

Every metric below is from a production training run on the full data corpus — not smoke-run inflations, not cherry-picked seeds. Headline numbers shown; full tables (W1–W13, FAERS, ACIC22, DMPNN, HNSI) at /benchmarks with raw log references available to design partners under NDA.

MIMIC-IV W1
0.9476
In-hospital mortality test AUROC, ECE 0.0024, full 546K corpus
FAERS Pipeline B
0.8872
Severity-classifier AUROC on n=158,732 temporal-split test, +9pp over Bate 2019
ACIC22 Track-2
77.5%
CI coverage on full 3,400-cohort canonical, bias +19.26
MIMIC-IV W13
RCT-✓
DOAC vs warfarin: first observational lane to recover RCT direction on both outcomes

See all benchmark tables →

Regulatory pathway

Built to regulator standards from day one.

Non-device CDS exemption

Pharmacovigilance triage product satisfies all four criteria of 21st Century Cures Act § 3060: transparent algorithms, documented provenance, human-in-the-loop, independent review of the basis for the recommendation. Ships to pharma drug-safety customers in 2026 without 510(k).

21 CFR Part 11 ALCOA+ by design

Patent-pending audit module satisfies 21 CFR Part 11 § 11.10(a)–(k). SHA-256 chained ledger, monotonic timestamps, fsync + multi-writer protection, external-verifiable. Pharma clients use our outputs in their own regulated workflows without bolt-on audit infrastructure.

FDA 7-step AI credibility framework

Structured around the Jan 2025 draft "Considerations for the Use of Artificial Intelligence To Support Regulatory Decision-Making" — context-of-use definition, model-risk assessment, credibility-assessment plan, execution, documentation, and adequacy determination. Applied today on the PV and RWE products, not deferred to a future SaMD.

SaMD pathway documented (post-2027)

Q-Submission scoping for direct clinician-facing CDS deployment is planned post-first-signed design partner. Predetermined Change Control Plan (PCCP) drafted in parallel — the no-LLM-in-the-prediction-path frozen-feature architecture makes the PCCP modification space bounded and auditable.

Full regulatory alignment map →

Official Python SDK

pip install rosenbound

Programmatic access to the audited platform — for CRO methodology leads, drug-safety directors, and RWE teams who need to script cohort uploads, run sensitivity-bounded studies, and pull reproducibility certificates without leaving their analysis environment. Pydantic v2 typed models with py.typed marker for full IDE autocomplete + mypy support. Apache 2.0 license.

v0.1.0 (alpha) Python ≥ 3.10 Apache 2.0 Bearer-token auth View on PyPI →

What it does today

  • Cohorts. Upload CSV + DSL; track validation state; pull reproducibility hashes.
  • Studies. Create, run, and fetch results from the five-method sensitivity pentagon.
  • Reproducibility certificates. Retrieve the certificate + methodology PDF for every signed-off study — the audit artifact your QA team hands to an FDA inspector.

Reserved for next releases

  • Async client. Coroutine-native concurrency for batch cohort onboarding.
  • WorkOS SSO. Enterprise-SSO auth flow for organizations not yet on API tokens.
  • Full certificate chain verify. Client-side SHA-chain integrity check (v0.1.0 ships a server-trust stub).
  • Per-feature resources. MedDRA browser, E2B import, drug coder, signal queue.

Security — how access is gated

The SDK is open-source. Platform access is not. Every protected endpoint enforces the following stack server-side — the SDK is a thin typed wrapper; the audit + isolation guarantees live in the platform:

  1. Bearer token verification via FastAPI's get_current_user middleware. Tokens are JWTs minted from the in-product API-key management surface (admin role required). Tokens are never in the SDK source or in PyPI metadata.
  2. RBAC permission check via the require_permission(...) dependency on every protected route. Endpoints declare their required permission inline; missing permissions return HTTP 403 with no data echo.
  3. Tenant scoping via TenantContextMiddleware + Postgres Row-Level Security policies. Cross-tenant reads and writes are server-rejected even with a valid token. The platform reads nothing outside the caller's organization.
  4. Append-only audit log on every regulated write. SHA-256-chained AppendOnlyAuditLog entries + a mirrored audit_logs row + 21 CFR Part 11 ALCOA+ attestation fields. Ledger writes are fail-loud: a write that cannot anchor to the SHA chain rolls back the entire request.
  5. Transport security. HSTS preload, HTTPS-only on production surfaces, TLS 1.2+ enforced, secure + httpOnly + SameSite cookies. The pre-launch staging surface (testing.rosenbound.com) carries an additional HTTP Basic Auth gate for partner-only isolation.
  6. SOC 2 Type I in flight. Auditor engagement active; CC6.6 MFA gate closed; encryption-at-rest via pgcrypto; disaster-recovery runbook tested. BAA + DPA + MSA templates ready for partner signing.
  7. No LLM in the prediction path. Pharmacovigilance triage, causal estimation, and audit-anchored writes never touch a large language model. Classical ML (LightGBM, causal-forest), calibrated ensembles, and symbolic rule engines only — the substrate an FDA reviewer can audit without trusting a foundation-model black box.

Three things a CRO methodology lead does with this on day one

Reproducibility on demand.
cert = client.certificates.get(study_id)
pdf = client.certificates.download_pdf(study_id)
Hand the methodology PDF and the SHA-256 reproducibility hash to your QA team or an inspector. No screenshots, no email forwarding.
Batch cohort onboarding.
Iterate a directory of CSVs, call client.cohorts.create(csv=..., dsl=...) for each, and capture the validation status. Fail-loud on any rejection so your pipeline halts instead of silently dropping rows.
Headless study orchestration.
client.studies.create(...).run().get_results() wired into your existing Airflow / Prefect / cron stack. Pull pentagon estimates programmatically; route the Γ-bound + sensitivity intervals into your downstream dashboards.

Getting an API token. Founding Partners receive scoped API tokens during onboarding — issued from the in-product admin surface, tenant-locked, and limited to the RBAC roles each team member needs.   Apply for the Founding Partner Program →

Intellectual property

Patent-pending methodology.

USPTO Provisional Filing

Five inventive concepts covering causal inference, audit-trail integrity, capability-aware abstention, modular cognitive substrate, and hybrid neuro-symbolic clinical evidence extraction.

Founder Harsh Singh filed a USPTO provisional patent on March 22, 2026, sole inventor, covering the methodology stack that powers Rosenbound. Twelve-month conversion window to non-provisional or PCT runs through March 22, 2027. Clinical AI is the first commercial application of the underlying architecture.

Specific algorithmic and architectural detail is available to design partners under NDA after term-sheet signing. Capability-level description is shared openly; implementation specifics are reserved as the durable moat.

Founding Partner Program

Five seats. Locked Founding rate. Forever.

We're selecting five founding design partners for the Rosenbound platform across pharma drug-safety teams and CRO methodology groups. The economics are sized to fit a director's discretionary budget; the relationship is structured for a multi-year reference partnership.

The economics

Year 1
$100K
vs $250K list (60% off)
Years 2–3
$175K
per year, locked (30% off)
Year 4+
$250K
locked at Founding rate forever

What you also get

  • Direct founder Slack access — not a ticket queue
  • Quarterly roadmap co-input session (Founding Partners help shape the product)
  • Named-reference recognition once the engagement is public
  • Waived integration setup ($50K value)
  • First access to every new module before public release
  • Co-authorship rights on benchmarks run against your cohort (with your approval)
  • MSA, DPA, and BAA templates pre-prepared — vendor security review fast-track
  • Locked Founding rate persists for the life of the contract, even when list price rises
Apply →
5 seats total. Selecting through Q3 2026.