sample_id string | source_framework string | target_framework string | concept string | source_definition string | target_definition string | claim string | option_a string | option_b string | task string | correct_option string | required_note string | required_keywords string | axis string | difficulty int64 |
|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|
CSAF-0001 | Protocol | Publication | Serious adverse event | SAE per ICH E2A | SAE excludes hospitalization for observation | Treat SAE rates as comparable | Invalid correspondence, SAE definition narrowed | Valid correspondence, same SAE | Judge correspondence and write note | A | State SAE definition change and block direct comparison | SAE|ICH E2A|hospitalization|narrowed|not comparable | SAE_definition | 3 |
CSAF-0002 | Protocol | SAP | TEAE window | TEAE through 30 days post dose | TEAE through 7 days post dose | Assume same TEAE window | Invalid correspondence, window shortened | Valid correspondence, same window | Judge correspondence and write note | A | State 30 vs 7 days and flag window change | 30 days|7 days|TEAE|window|shortened | event_window | 3 |
CSAF-0003 | Phase 2 | Phase 3 | Cardiovascular event | CV event includes unstable angina | CV event excludes unstable angina | Treat CV event rates as same across phases | Invalid correspondence, component removed | Partial correspondence, restate components | Judge correspondence and write note | A | State unstable angina inclusion change and flag drift | unstable angina|include|exclude|CV event|drift | adjudication_rule | 3 |
CSAF-0004 | Protocol | Publication | Adjudication process | Independent blinded adjudication | Site-reported only | Pool adjudicated and site-reported events | Invalid correspondence, ascertainment changed | Partial correspondence, stratify by method | Judge correspondence and write note | A | State adjudicated vs site-reported and flag ascertainment shift | adjudication|site-reported|blinded|ascertainment|shift | measurement_rigor | 3 |
CSAF-0005 | Protocol | Publication | Event severity | Bleeding grade 3+ | Bleeding grade 2+ | Treat bleeding rates as same outcome | Invalid correspondence, threshold changed | Partial correspondence, restate threshold | Judge correspondence and write note | A | State grade 3+ vs 2+ and flag threshold drift | grade 3+|grade 2+|bleeding|threshold|drift | seriousness_threshold | 3 |
CSAF-0006 | SAP | Publication | Primary safety endpoint | Any SAE | Treatment-related SAE only | Report as same safety burden | Invalid correspondence, attribution filter added | Partial correspondence, separate outcomes | Judge correspondence and write note | A | State any SAE vs treatment-related and block direct pooling | any SAE|treatment-related|attribution|filter|pooling | attribution_filter | 3 |
CSAF-0007 | Protocol | Publication | Lab abnormality definition | ALT >3x ULN with symptoms | ALT >5x ULN any | Treat hepatotoxicity as same | Invalid correspondence, cutoff changed | Partial correspondence, restate cutoffs | Judge correspondence and write note | A | State 3x with symptoms vs 5x any and flag cutoff drift | ALT|3x ULN|5x ULN|symptoms|cutoff | threshold_rule | 3 |
CSAF-0008 | Phase 2 | Phase 3 | Follow-up for safety | Safety follow-up 90 days | Safety follow-up 365 days | Compare safety incidence directly | Invalid correspondence, follow-up differs | Partial correspondence, time-adjust rates | Judge correspondence and write note | A | State 90 vs 365 days and require time-adjusted incidence | 90 days|365 days|follow-up|incidence|time-adjust | followup_duration | 3 |
CSAF-0009 | Protocol | Publication | Device-related event | Event includes device malfunction | Event excludes device malfunction | Treat device event as unchanged | Invalid correspondence, category removed | Partial correspondence, restate categories | Judge correspondence and write note | A | State malfunction category change and flag drift | device malfunction|include|exclude|category|drift | event_taxonomy | 2 |
CSAF-0010 | Protocol | Publication | Death attribution | All-cause mortality | CV death only | Treat death endpoint as same safety signal | Invalid correspondence, cause filter added | Partial correspondence, separate outcomes | Judge correspondence and write note | A | State all-cause vs CV death and block equivalence | all-cause|CV death|cause filter|not equivalent|mortality | cause_filter | 3 |
Clarus Safety and Adjudication Framework Transition Integrity v0.1
What this dataset tests
- You track how safety concepts change meaning across trial phases and reporting layers
- You detect when adverse event definitions shift without justification
- You flag adjudication and attribution drift that breaks comparability
Framework transitions covered
- Protocol → SAP
- SAP → publication
- Phase 2 → Phase 3
- Planned adjudication → reported adjudication
Scope
- One trial safety domain
- Multiple safety and adjudication frameworks
- Multiple reporting stages
Typical failure modes
- Serious adverse event definitions narrowed or broadened
- Safety follow-up windows shortened or extended
- Event adjudication switched to site reporting
- Severity thresholds altered between documents
- Attribution filters added post hoc
What this dataset does not test
- Efficacy endpoint changes
- Population eligibility drift
- Internal narrative rhetoric
Task format
- Each row presents a safety or adjudication definition in a source and target framework
- The model judges whether correspondence is valid
- The model must name the reason for failure when correspondence breaks
Label meaning
- A indicates invalid or unsupported safety correspondence
- B indicates valid correspondence with explicit continuity
Use cases
- Automated safety consistency checks
- Regulatory pharmacovigilance review
- Systematic review safety harmonization
Version note
- v0.1 establishes baseline safety and adjudication transition auditing
- Later versions may add phase-linked severity weighting
- Downloads last month
- 7