Flagship synthetic workflow
A regulatory writing team uses AI to draft a clinical study report narrative from synthetic trial outputs. The draft sounds polished, but several efficacy, subgroup, and safety statements need exact traceability to tables, listings, figures, SAP definitions, and analysis datasets before they can survive.
The draft
The draft states that QLX-204 met the primary endpoint, demonstrated statistically significant secondary endpoint improvement, showed consistent benefit across all subgroups, was well tolerated, had no serious safety signals, and is FDA-ready.
The risk
A CSR narrative can overstate efficacy, generalize subgroup results, soften safety findings, or imply regulatory readiness without the required source support. QEL converts polished narrative language into claim records before any final text is trusted.
What QEL controls
QEL extracts claims, maps evidence, applies the active rule pack, builds a registry, and compiles survived output only from admitted or caveated records.
Why QEL is needed
A CSR narrative can overstate efficacy, generalize subgroup results, soften safety findings, or imply regulatory readiness without the required source support. QEL converts polished narrative language into claim records before any final text is trusted.
The draft reads polished before its material claims are qualified.
Some claims need exact source spans, not just a credible nearby document.
Unsupported or review-required claims can be copied forward unless a firewall stops them.
Synthetic Draft Preview
The draft states that QLX-204 met the primary endpoint, demonstrated statistically significant secondary endpoint improvement, showed consistent benefit across all subgroups, was well tolerated, had no serious safety signals, and is FDA-ready.
Full synthetic source document
Synthetic source document inspected by QEL.
Synthetic QEL demonstration material - not medical advice - not regulatory advice - not a real clinical study report.
Evidence Packet
Type: analysis plan
Supports: Primary endpoint alpha, population, endpoint hierarchy.
Cannot support: Regulatory readiness or safety conclusions by itself.
Type: TLF table
Supports: LS mean difference and p-value for the primary endpoint.
Cannot support: All subgroup or submission-readiness claims.
Type: subgroup table
Supports: Individual subgroup direction and confidence intervals.
Cannot support: Consistent benefit across all subgroups.
Type: subject listing
Supports: SAE count and arm imbalance.
Cannot support: No serious safety signal without clinical review.
Configurable Rule Pack
The active synthetic CSR pack requires exact numeric source spans, SAP/protocol alignment, multiplicity context, and human review gates for safety and regulatory interpretation.
Synthetic sample rule pack only. Not official, not endorsed, not regulatory approval, and not a substitute for qualified human review.
{
"pack_id": "synthetic_fda_csr_traceability_pack_v0",
"display_name": "Synthetic FDA CSR Traceability Pack",
"version": "0.1.0",
"domain": "life_sciences_pharma",
"workflow": "clinical_study_report_regulatory_narrative",
"status": "synthetic_demo_only",
"not_official": true,
"synthetic_demo_only": true,
"intended_users": [
"regulatory writing",
"medical writing",
"clinical operations",
"biostatistics",
"quality",
"regulatory affairs"
],
"compatible_outputs": [
"CSR narrative",
"regulatory summary",
"table-linked clinical narrative",
"safety narrative"
],
"required_evidence_types": [
"protocol excerpt",
"SAP excerpt",
"TLF table",
"SAE listing",
"ADaM traceability map"
],
"decision_states": [
"admitted",
"admitted_with_caveat",
"blocked",
"review_required"
],
"rule_categories": [
"numeric clinical result claims",
"efficacy endpoint claims",
"safety claims",
"SAE claims",
"subgroup claims",
"population claims",
"statistical significance claims",
"clinical interpretation claims",
"regulatory readiness claims"
],
"human_review_gates": [
"safety interpretation",
"SAE conclusions",
"clinical meaningfulness",
"regulatory readiness wording"
],
"blocked_language_patterns": [
"proves",
"safe",
"effective",
"no risk",
"FDA-ready"
],
"caveat_triggers": [
"multiplicity context",
"exploratory endpoint",
"mixed subgroup result",
"medical interpretation"
],
"freshness_requirements": [
"current SAP version",
"final locked TLF",
"traceable analysis dataset"
],
"source_span_requirements": [
"table ID",
"row",
"column",
"population",
"treatment arm",
"timepoint",
"unit",
"exact value"
],
"audit_artifacts_generated": [
"claim map",
"evidence/source-span map",
"admission decision table",
"registry view",
"blocked/review appendix",
"survived output",
"provenance/audit summary"
],
"final_output_policy": "Final output allowed only for admitted or admitted_with_caveat registry records.",
"disclaimer": "Synthetic sample rule pack only. Not official, not endorsed, not regulatory approval, and not a substitute for qualified human review."
}Claim Firewall Process
Results dashboard
QEL separated admitted, caveated, blocked, and review-required claims before compiling survived output.
Top findings
admitted
Why it matters: Numeric efficacy claims can survive when table, population, unit, and endpoint definition align.
Evidence / rule reason: Table 14.2.1 and SAP endpoint context match.
Survived final output: yes
blocked
Why it matters: A polished CSR can overstate non-significant endpoints.
Evidence / rule reason: Adjusted p-value is .083, so the significance claim is contradicted.
Survived final output: no
blocked
Why it matters: Subgroup generalizations are high-risk when outputs are mixed.
Evidence / rule reason: The subgroup table includes rows crossing null.
Survived final output: no
review_required
Why it matters: SAE interpretation requires qualified clinical review.
Evidence / rule reason: The SAE listing has an imbalance and the active pack has a human-review gate.
Survived final output: no
admitted_with_caveat
Why it matters: Tolerability wording can survive only with scope-limited caveats.
Evidence / rule reason: AE table supports limited wording, but medical review remains required.
Survived final output: yes, with caveat
blocked
Why it matters: QEL does not admit regulatory-readiness claims from statistical tables alone.
Evidence / rule reason: No submission package evidence exists.
Survived final output: no
Claim Admission Report
QEL does not compile final output from raw AI text, unchecked draft language, or a final model polish step. Survived output is compiled only from registry records with admitted or admitted_with_caveat decisions and final_output_allowed = true.
Rule Pack Marketplace Preview
This synthetic pack previews how QEL could support industry-specific admission controls. In a future rule-pack ecosystem, teams could configure, review, version, and govern packs for their own workflows.
This is a synthetic preview only. The pack is not official, certified, endorsed, or a production claim.