01 · ResearcherSupplies contextPlans, people, consent, procedures
Pre-submission ethics workspace
Prepare a research plan that humans can inspect.
SafeBARS turns protocol materials into a participant journey, bounded scenario tests, and focused questions for a qualified ethics expert—before the plan reaches a committee.
Outputs support an application draft and research design. They are not institutional ethics approval.
02 · SafeBARSStress-tests encountersMaps, traces, provenance, trade-offs
03 · Ethics expertResolves boundariesOnly questions requiring human judgment
How this run is producedAgent roles, bounded tools, provenance, and stopping rules Inspect details
Encounter Orchestrator
Waiting for protocol materials.
Breakdown Scenario Agent
No scenario queue.
Relationship and Power Agent
No relationship check.
Boundary and Handoff Agent
No boundary review.
Inspectable audit plan Material-routed tasks, tools, dependencies, and stop conditions
Your first output
Build the participant encounter map
Use the guided setup or add existing protocol materials. SafeBARS will show the journey before it asks you to run any scenario tests.
Framework-grounded ethics map
Belmont baseline; conditional Menlo, AI human-subjects review, REC guidance, and NIST extensions; VSD and ESR process support.
Evidence coverage at a glance
Click a donut segment or count to filter the framework map. Coverage describes submitted material, not ethical quality or approval.
How to read a source label: “Interview questions · passage 1” means the first numbered excerpt taken from the Interview questions field. The hidden internal ID (for example INT-001) is only a traceability reference—not a risk or approval score.
Saved analysis: charts reflect the current protocol case.
Active cited frameworks
Select a petal to inspect its question and submitted evidence.
Framework-linked Trade-off Board
Compare each design parameter by current leaning, connected ethics dimensions, downstream document impact, and the expert role that should review unresolved risk. The board is the decision surface; the relationship map is optional context.
Show optional framework relationship overview
Foundational ethics / VSD / ESRMenlo ICT extensionAI ethics-review and risk extensionsGreen / amber / red outline = documented / partial / missing evidence
Visualization rationale: adapted from the synthesized-data principle of An et al.'s Dandelion Diagram (CHI EA 2020); SafeBARS keeps the relationship overview but adds a decision board because ethics trade-offs need inspectable consequences, not only spatial attention patterns.
Application draft readiness
A completeness scaffold, never an ethics verdict or institutional form.
Participant journey stress map
Switch lenses to compare written protocol coverage with scenario-test outcomes.
Covered in selected lensPartial or untested in CombinedMissing or audit gapNo scenario resultExcluded from scopeNumber inside each node = linked tests run
Encounter stages
Edit scope and responsibility before running scenarios.
Scenario queue
Bounded test conditions, not predictions.
No scenario results yet
Select difficult situations in the encounter map, then choose “Run selected traces”.
No researcher decisions yet
After a scenario test finds a protocol gap or ambiguity, it will appear here with a proposed change.
Expert review is not needed yet
Run the audit first. If a question requires human judgment, SafeBARS will create a review request here.
Inspectable prototype evidence
Evidence & Validation
A public, case-level record of what the deterministic SafeBARS pipeline actually produced on author-seeded protocols. The matrix, aggregates, and lineage below are calculated from the same executable benchmark used in continuous testing.
Synthetic seeded technical validation — not a human study. Passing demonstrates specification conformance, not ethical validity, institutional compliance, user benefit, real-world effectiveness, or ethics approval.
Loading validation record
SafeBARS is reading the deterministic case-level evidence.
Case × check matrix
Each cell is one executable assertion, not a self-reported score. Choose a domain and case to inspect what was tested.
Select a case row to inspect its seeded omission and observed checks.
Benchmark composition
Balanced domains and conditional framework pathways make the testbed inspectable; they do not represent real-world prevalence.
Seeded omissions surfaced
Evidence lineage
This is an output-and-provenance flow. “Coverage” means submitted material was located; it is not an ethics score.
Human-study evidence is still pending
These measures require expert validation and a comparative study; SafeBARS will not fill them with developer/demo session counts.
Expert-rated quality PendingCompleteness, safeguard quality, grounding accuracy, and false reassurance.
Comparative benefit PendingSafeBARS versus general LLM chat on matched protocol tasks.
Researcher workload PendingTask time, SUS, workload, trust calibration, and revision effort.
Expert triage PendingHandoff routing quality, review time, resolution rate, and disagreement handling.