Agents

Reads

• study plan: phase, funding source, indication, target enrollment and window
• sites contracted to the study and their countries
• consented patients and months elapsed (active studies)
• latest protocol synopsis for restrictive eligibility criteria
• ClinicalTrials.gov termination comparator for the indication and phase
• existing feasibility assessments

Escalates

Low-feasibility assessments (predicted enrollment shortfall or termination-risk red flags), as a medium-priority review task

To: clinical operations lead

Literature & benchmarks

• Zhang & DuBois 2023 — odds ratios for early termination by funding source, phase and design, applied as the risk prior
• Dombernowsky et al. 2019 — site recruitment forecasts overshoot; projections are discounted accordingly
• ClinicalTrials.gov API v2 — comparator trials in the same indication and phase for enrollment and duration benchmarks

Data sources

• Study design fields (phase, design, blinding, funding source)
• ClinicalTrials.gov API v2 comparators
• CTX-NSCLC-301 enrollment (clinTrialData v0.1.2, CC BY 4.0 synthetic)

Reasoning written by

Reads

• the latest protocol synopsis on file (linted as written)
• study design, blinding, arms, target enrollment, primary endpoint, funding, registry id
• planned visit schedule, recorded dosing, enrolled demographics, contracted sites
• SPIRIT 2013/2025 checklist with Phillips 2019 harms items (priv/benchmarks/spirit_checklist.json)
• previous protocol_review documents and pending approval tasks

Escalates

Every auto-drafted protocol, listing the unresolved SPIRIT items and the predicted amendment risk, for medical review before use

To: medical reviewer

Literature & benchmarks

• SPIRIT 2013 (Chan et al., Ann Intern Med 2013;158:200–207) — 33-item protocol content checklist the draft is linted against
• SPIRIT 2025 (Chan et al., Nat Med 2025;31:1784–1792) — 34 items; gaps at v1.0 drive avoidable amendments, which sets the amendment-risk score
• Phillips et al. 2019 — harms-reporting items checked in the safety section

Data sources

• Study record and design
• Aldox Phase 1 synopsis (real, de-identified)
• CDISCPILOT01 protocol structure

Reasoning written by

Reads

• site contracts for the study: sent and executed dates, status
• site activation date, principal investigator and first-patient date
• consented patients per site
• regulatory submission and feasibility assessment on file for the study

Escalates

Sites whose contract or budget aging breaches the published Q3 benchmark, or that are missing essential documents, as medium-priority tasks

To: site start-up manager

Literature & benchmarks

• Martinez et al. 2016 (Trials 17:106) — 78 industry trials: contract negotiation median 33 d / Q3 77 d, budget 40 / 68 d, IRB approval 3–4 d; contract and budget, not IRB, are the critical path
• Kurbegov et al. 2021 (ASCO Research Statement) — a feasibility assessment on file is a prerequisite for activation

Data sources

• Site contracts and their aging
• Regulatory submissions
• Feasibility assessments on file

Reasoning written by

Reads

• patients: status funnel and consent dates per site
• enrollment plan: target, months and start date
• sites contracted to the study: status, activation date, accrual
• latest protocol synopsis (eligibility criteria) and the eligibility priors table
• pending candidate-import review tasks

Escalates

Protocol-matched candidates, batched for import; nobody is contacted or enrolled without a coordinator's approval

To: study coordinator

Literature & benchmarks

• Screen-failure funnel — referred → pre-screened → consented → eligible, with the eligibility pass fraction as the tracked indicator
• Eligibility pass fraction applied against the study's target enrollment and enrollment window

Data sources

• Eligibility criteria from the protocol
• Patient records and screening status
• CTX-NSCLC-301 DM / IE domains (CDISC SDTM)

Reasoning written by

Reads

• serious and grade ≥3 adverse events per patient at each site
• major/minor protocol deviations and open QA findings per site
• data queries open past the 14-day SLA
• subject visits against the protocol visit schedule (window compliance, missed visits)
• out-of-range lab flags, withdrawals, deaths and enrollment pace per site

Escalates

Sites whose key risk indicators breach threshold, with a proposed on-site visit a CRA must confirm (high priority)

To: clinical research associate / clinical lead

Literature & benchmarks

• ACRO RBQM landscape survey (Barnes et al. 2021) — KRI set: query rate, AE rate, protocol deviations, data-entry lag
• Centralised statistical monitoring (ICH E6(R2) §5.18.3) — sites compared against the study distribution rather than a fixed cut-off

Data sources

• Sites and monitoring visits
• Data queries and adverse events per site
• CTX-NSCLC-301 DV / SV domains

Reasoning written by

Reads

• adverse events against dosing, death and each other (onset, resolution, seriousness, grade, duplicates)
• demographics and disposition (consent, death, disposition dates)
• subject visits against the protocol schedule and windows
• exposure intervals and doses against the study median
• lab results against reference ranges, plausibility bounds and consent
• open data queries and their age against the 14-day JSCDM SLA

Escalates

Queries open longer than the 14-day SLA, and any query an automated edit check cannot resolve, as medium-priority tasks

To: data manager

Literature & benchmarks

• JSCDM (Amatya et al. 2021) — 14-day query-resolution SLA as the aging threshold
• Edit-check taxonomy — range, consistency and missing-value checks generated from the CRF

Data sources

• Data queries and their age
• CTX-NSCLC-301 LB / VS domains (CDISC SDTM)
• CDISCPILOT01 ADaM datasets

Reasoning written by

Reads

• adverse events: seriousness criteria, grade, investigator causality, onset/outcome
• patient demographics, dosing, arm (unblinded studies only) and deaths
• class reference label and FAERS counts from openFDA (committed snapshot)
• pending assess_serious_event review tasks

Escalates

Every serious or severe adverse event as a critical task; causality and expectedness are never decided by software

To: safety physician

Literature & benchmarks

• ICH E2A — 7-day (fatal or life-threatening, unexpected) and 15-day expedited reporting clocks
• openFDA drug labels — listed reactions as the expectedness reference
• FAERS disproportionality (PRR / ROR) as the aggregate-signal check

Data sources

• Adverse events (seriousness, severity, narrative)
• openFDA label and FAERS endpoints
• CTX-NSCLC-301 AE domain

Reasoning written by

Reads

• consented and enrolled patients (status, consent date, first dose, site)
• protocol visit schedule (TV) and attended visits (SV) per patient
• adverse events per patient: CTCAE grade and serious events in the last 60 days
• exposure intervals (EX) to detect stopped treatment
• site withdrawal rate within the study
• Luna engagement signal (check-in recency, mood, adherence) for linked participants
• disposition and withdrawal history for the cohort retrospective

Escalates

Participants above the dropout-risk threshold, with proposed outreach a coordinator approves before anyone is contacted

To: study coordinator

Literature & benchmarks

• Dropout-risk model — missed visits, overdue check-ins and AE burden weighted into a per-participant score
• Human-approved outreach only; the agent never contacts a participant

Data sources

• Patient visit schedule and Luna check-ins
• Adverse events per participant
• CTX-NSCLC-301 DS / SV domains

Reasoning written by

Reads

• enrollment plan (target, months, start) and every consent date
• funding source, phase, design and contracted-site countries
• patient dispositions, deaths, serious and grade ≥3 adverse events
• latest protocol for oversight, stopping rules and interim analysis
• ClinicalTrials.gov termination comparators for the indication and phase

Escalates

Studies carrying early-termination red flags, at the priority the flag count implies

To: sponsor study lead / steering committee

Literature & benchmarks

• Zhang & DuBois 2023 — red-flag catalogue: slow accrual, single site, non-industry funding, phase and design factors
• Lièvre et al. 2001 (BMJ) — governance safeguards against commercially motivated termination

Data sources

• Enrollment against plan
• Site count and activation status
• Funding source and design fields

Reasoning written by

Reads

• exposure records (treatment, dose, dates) per patient
• patient status, consent, first/last dose and disposition
• sites contracted to the study
• previous supply forecasts for the study
• planned visit schedule and protocol synopsis (startup fallback)

Escalates

Nothing — purely operational. The kit forecast is handed to the IRT/RTSM partner, where humans act on it

To: IRT/RTSM partner (integration boundary)

Literature & benchmarks

• Dosing-derived kit forecast — enrolled × retention × dosing schedule over the horizon, plus a 15% buffer
• IRT/RTSM partner boundary — GalenOps forecasts; the vendor dispenses and resupplies

Data sources

• Enrollment and withdrawal counts
• Dosing schedule from the protocol
• CTX-NSCLC-301 EX domain

Reasoning written by

Reads

• patient disposition, demographics and randomised arm
• study-drug exposure intervals
• adverse events by system organ class, preferred term, grade and seriousness
• time-to-event and best-overall-response efficacy records (unblinded studies only)
• TLF documents already awaiting validation and their batch hash

Escalates

Every generated TLF batch, for double-programming-equivalent validation before release

To: validating biostatistician

Literature & benchmarks

• ICH E3 §14 — the standard tables, listings and figures set
• Kaplan–Meier estimates for time-to-event endpoints
• Double-programming gate — nothing is released without independent validation

Data sources

• CDISCPILOT01 ADaM (ADSL, ADAE, ADTTE)
• CTX-NSCLC-301 ADaM

Reasoning written by

Reads

• study design, arms, endpoint, registry id and data cutoff
• patient disposition, demographics and deaths by arm
• protocol deviations (QA findings)
• adverse events: overview, frequent terms, serious events, fatal outcomes
• exposure summary and planned visit schedule
• efficacy records (closed-out studies only): time-to-event and response
• CSR drafts already on file and their review tasks

Escalates

Every CSR draft as a high-priority sign-off task; a draft never becomes a document of record without a signature

To: medical writer / medical monitor

Literature & benchmarks

• ICH E3 — CSR structure and content
• Narrator drafting — an LLM writes prose only from computed evidence, and the deterministic summary is kept alongside
• Mandatory human sign-off

Data sources

• Validated TLF outputs
• Protocol and study record
• Aldox Phase 1 synopsis for Phase 1 structure

Reasoning written by

Reads

• protocol versions, investigator's brochure, feasibility and regulatory submission on file
• per-site executed contracts, activation dates, monitoring visits and payments
• serious adverse event assessments, expedited-report clocks and aged data queries
• TLF and CSR documents, prior agent runs and open TMF findings

Escalates

Missing essential artefacts become open QA findings; the file is audited continuously rather than at inspection

To: quality assurance

Literature & benchmarks

• DIA TMF Reference Model v3.0 — expected essential-artefact set by zone
• ICH E6(R2) §8 — essential documents for the conduct of a clinical trial

Data sources

• Documents, contracts, submissions and feasibility assessments on file
• QA findings

Reasoning written by