Whitepaper · v0.1 draft · May 2026

Auditable Clinical Translation Orchestration. A verifier-backed approach to real-time medical language bridging.

The third category between human Video Remote Interpretation and AI-only translation. Defined by simultaneous performance on translation quality, latency, cost, and regulatory-grade auditability.

Freddy Rojas · Orquor · Lima, Peru

CC BY 4.0 specification · MIT verifier code · submission to arXiv cs.CL planned post community review

01 — Abstract

What this paper claims, and what it does not.

Abstract

We introduce Auditable Clinical Translation Orchestration (ACTO) as a new technical category for real-time medical interpretation systems that combine automated speech recognition, machine translation, and human-in-the-loop interpreters into a single orchestrated pipeline whose every output is verified by deterministic outcome-based tests and recorded in a cryptographically timestamped audit log. We argue that the current generation of Video Remote Interpretation (VRI) services and AI-only translation tools occupies opposite ends of a quality-cost trade-off without addressing the legal, regulatory, and forensic requirements emerging in healthcare systems across Latin America and the United States. We propose ACTO as a third category that resolves the trade-off through a cooperative inference stack, a verifier-first architecture, and a regulatory-grade audit primitive.

02 — Sections

Eight sections, two appendices.

The full document is the PDF. The web summary below mirrors the paper structure for navigation.

03 — Methodology

How a verifier is designed.

A verifier is a deterministic test function that evaluates one outcome property of system output. Atomic, weighted, versioned. Adapted from outcome-based rubric work in OpenClaw Atlas and HiL-Bench.

Principle 1. Verifiers evaluate outcomes, not traces. A verifier checks that the final translated output has a property, not that the system's intermediate steps followed a particular pattern. This makes verifiers robust to implementation changes — as MT models improve, the verifiers continue to apply unchanged.

Principle 2. Verifiers are atomic. Each verifier evaluates exactly one property. A verifier that combines multiple checks via conjunction obscures failure attribution.

Principle 3. Verifiers are weighted explicitly. A misidentified medication name is catastrophic (weight 5); a stylistic register deviation is mild (weight 2). The orchestrator's escalation logic is parameterized by the weighted sum.

Principle 4. Negative verifiers are required. A complete suite includes both positive checks (contains_dosage) and negative checks (does_not_contain_excluded_names). Without negative verifiers, the suite cannot detect leakage of disallowed content.

Principle 5. Verifiers are versioned. When a verifier changes, the version is recorded in the audit log. A translation verified under suite v1.2 remains auditable as such after the suite advances to v1.3.

The current Orquor Clinical production suite contains approximately 60 verifiers organized by clinical specialty (Emergency, OR, ICU, Pharmacy, Pediatrics, Psychiatry, Admissions, Legal, 911 Dispatch, General) and by register (Formal/Clinical, Simplified, Legal).

# A representative verifier from the production suite
# Appendix A · whitepaper v0.1

def verify_dosage_preserved(
    source: str,
    translated: str,
) -> VerifierResult:
    """Numeric dosage in source must appear
    identically in translated output.
    Weight: +5 (catastrophic if violated)."""

    source_dosages     = extract_numeric_dosages(source)
    translated_dosages = extract_numeric_dosages(translated)

    missing = source_dosages - translated_dosages
    if missing:
        return VerifierResult(
            passed=False,
            weight=5,
            reason=f"Dosage {missing} missing",
        )
    return VerifierResult(passed=True, weight=5)
04 — Preliminary results

Single-site, preliminary, not a competitive claim.

From a 500-utterance evaluation set spanning three specialties at a Lima medical-tourism clinic. Presented as evidence of architectural feasibility, not as production guarantees.

SpecialtyWERCOMET-22Verifier passLatency p50Human escalation
Emergency4.1%0.8794.2%0.9 s5.8%
Operating Room3.6%0.8895.6%0.8 s4.4%
Pharmacy3.2%0.8996.1%0.7 s3.9%

Baseline reference: Whisper-Large-v3 medical WER 11.8% on the same evaluation set without ACTO orchestration. Full methodology in PDF Section 5. Independent benchmark to be published 2026-Q4 with 1,000 utterances per specialty and IRB review.

05 — Author & citation

Cite this.

Freddy Rojas

Founder · Orquor · Lima, Peru

Founder of Orquor, the company building the orchestration substrate for verifier-backed clinical AI. Previously built ITERAMED. Author of the ACTO whitepaper and the open verifier-design specification. Active in the agent-evaluation research community.

research@orquor.com · x.com/Orquorai

BibTeX

@techreport{rojas2026acto,
  title  = {Auditable Clinical Translation Orchestration:
            A Verifier-Backed Approach to Real-Time
            Medical Language Bridging},
  author = {Rojas, Freddy},
  institution = {Orquor},
  year   = {2026},
  month  = {May},
  number = {v0.1 draft},
  url    = {https://orquor.com/whitepaper-acto-v0.1.pdf},
  note   = {Specification under CC BY 4.0;
            reference verifier code under MIT.}
}
06 — Engagement

Three ways to engage with this work.

If you are a hospital procurement or CIO office, include ACTO-aligned criteria in your next translation services RFP — require per-utterance audit logs with cryptographic timestamping, require verifier pass rates by specialty, require benchmark publication. Pilot inquiries: hello@orquor.com.

If you are a clinical informatics researcher, contribute to the public evaluation set we release in 2026. Independent benchmarks reduce vendor lock-in across the industry. Collaboration: research@orquor.com.

If you are another system builder, adopt the ACTO category. We are not pursuing this as a trademark. The category is open. The implementations will compete.

Open commitments

· Public evaluation set, 1,000 utterances/specialty, 2026-Q4
· Reference verifier suite open source, 2026-Q4
· arXiv cs.CL submission, post community review
· ACTO specification page at orquor.com/acto, 2026-Q3