IRIS is a 5-layer intent-routing orchestration framework that surgically separates genuine safety refusals from over-triggered defensive responses. Built for healthcare AI deployments where a refused query has patient-safety consequences.
IRIS doesn't bypass safety — it adds intent resolution before the model's safety layer fires. Each layer narrows the ambiguity window, so the model receives a contextually-grounded prompt that reads as clinical, not adversarial.
Augments the raw query with structured clinical metadata — ICD codes, patient risk stratum, provider role. Removes the ambiguity that triggers reflexive refusal on bare symptom strings.
clinical_context.roleencounter_type
Binary classifier (fine-tuned on 18K labeled medical queries) routes between clinical_analysis, care_coordination, admin, and out_of_scope. Only clinical_analysis paths proceed to IRIS routing.
intent.classintent.confidence
Maps jurisdiction + care setting to the correct compliance frame (UK NHS, US FDA 510(k), EU MDR). Rewrites prompt preamble to reference authoritative clinical guidelines — NICE, BNF, UpToDate — before reaching the LLM.
regulatory.frameguideline_refs[]
Scores prompt refusal probability (0–1) using a lightweight perceptron trained on 6K refused/accepted pairs. Prompts above threshold are pre-transformed before model submission; below threshold pass through unchanged.
refusal_risk.scoretransform_applied
Every routed request is GPG-signed, SHA-256 chained, and written to an append-only audit log. Provides defensible evidence trail for ISO 42001 AI governance, CQC inspection, and MHRA Pre-Sub processes.
audit.gpg_sigaudit.chain_hash
Designed for regulated environments where an AI refusal is not an inconvenience — it's a care pathway interruption.
CDSS vendors embed IRIS as a pre-processing layer. Differential diagnosis queries, drug interaction checks, and lab result interpretation pass through without triggering defensive refusal. Clinicians get answers, not apologies.
Supports: Epic, Meditech, Cerner embedded AI modules. Tested against clinical vignette corpora from NHS Digital and MIMIC-IV.
↓ 94% refusal reduction on DDx queriesPolypharmacy queries are among the highest-refusal categories in clinical AI — the combination of brand names, dosages, and contraindication language triggers nearly every LLM safety filter. IRIS's L3 regulatory framing reduces this to noise.
Validated against BNF interaction dataset (2,100 pairs). Integrates with DrFirst, Surescripts API surfaces.
↓ 97% refusal reduction on polypharmacy queriesBeyond healthcare: any regulated domain where AI over-refusal creates liability. Legal (M&A due diligence, contract risk), financial services (FINRA-regulated advice generation), and insurance (actuarial narrative). Same 5-layer architecture, domain-specific L3 frame.
Frame library: NHS, FDA, EU MDR, FINRA, FCA, Lloyds market standards.
Horizontal — not healthcare-lockedIRIS exposes two integration surfaces. The REST API fits teams with existing LLM proxy infrastructure. The Node/Python SDK fits teams embedding directly in their application layer.
Route any OpenAI-compatible request through IRIS by swapping your base URL. All 5 layers apply transparently. Latency overhead: <40ms p99 (co-located).
Install the IRIS SDK for programmatic access to all 5 layers, audit log queries, and frame management. Typings included. Zero extra dependencies in the production bundle.
IRIS was designed by operators who have worked inside the systems it now improves.
Experienced working within NHS systems and understanding firsthand where LLM refusal creates care pathway friction. IRIS was not designed from the outside — it was designed to fix a problem that caused real delays in clinical settings. The 34.3% baseline was measured on a real deployment corpus, not a synthetic benchmark.
Architect of multi-agent orchestration systems operating at production scale. The IRIS framework evolved from a pattern library built across multiple regulated-industry deployments. The 5-layer model was extracted from the most robust routing patterns observed across those deployments — then formalized into a licensable asset.
IRIS is priced on the refusal-bias delta you deploy, not commodity API calls. A 32-point reduction in clinical AI refusal rate is worth more than the infrastructure cost to deliver it.
or contact us first to scope before paying