Safety Passport · Worked example
What does a Safety Passport tell you?
Example review — uses sample data.
Follow a research assistant through a permission review. Radar identifies withdrawal access that exceeds its read-only purpose, then compares the updated settings after the reviewer records a change.
1. What was supplied
- Purpose: read-only research. Read and withdrawal permissions enabled; other declared permissions disabled.
- Emergency stop present; test not recorded; manual override and IP allowlist unknown.
- One operator. No quantitative limits or safeguard observation supplied.
- Example provider label, retry limit of 2 and cooldown of 30 seconds. No market snapshot or performance claim supplied.
The download includes both reviews and every submitted setting. Example date: September 30, 2026.
2. Findings and evidence gaps
Review summary · 1 blocker · 4 cautions
BLOCKER: Withdrawal scope exceeds the reviewed job
You declared that this agent can withdraw funds.
Next check: Remove withdrawal permission and use a separate non-withdrawal key.
CAUTION: IP allowlist status needs evidence
You have not confirmed whether access is restricted to approved IP addresses.
Next check: Record whether access is restricted to approved IP addresses.
CAUTION: Record the emergency-stop test
The reviewer reported an emergency stop with no completed test.
Next check: Test the emergency stop and record the operator.
CAUTION: Add a market-data snapshot
No market-data snapshot was supplied for this review.
Next check: Add source identity and timestamped market-data evidence.
RECORDED: No performance claims to review
No claim text was submitted.
RECORDED: Reviewer identity and role are recorded
Reviewer identity and role are supplied. Notes are optional.
CAUTION: Safeguard baseline was not supplied
Add an observation to compare with the selected safeguard.
Next check: Record the applicable safeguard baseline before drawing an adherence conclusion.
3. Changes recorded
The reviewer records that withdrawal permission was removed while read access stayed enabled. The remaining recommendations stay available for follow-up.
Reviewer-reported change: Withdraw funds — Yes → No.
Critical: Remove withdrawal permission from this agent/API key.
Reduces the risk of funds being withdrawn through the agent.
Critical: Require source timestamp evidence.
Prevents stale or unverifiable market evidence from being trusted.
Medium: Record the applicable safeguard baseline before drawing an adherence conclusion.
Keeps the evidence needed for follow-up visible.
4. Reassessment result
In your private workspace, open a saved report and choose “Record fixes and reassess this report.” Recheck current declarations, select one to five fixes and add sanitized evidence notes. Save a new report linked to the unchanged original.
Blockers: 1 → 0 · Cautions: 4 → 4
The withdrawal recommendation no longer appears for the updated settings. The original report is retained, and the other evidence questions remain visible.
- IP allowlist status needs evidence
- Record the emergency-stop test
- Add a market-data snapshot
- Safeguard baseline was not supplied
Workspace access is currently by invitation. Need access? Request a private workspace.
Explore a Radar snapshot
Switch between the sections below to see how recorded workflow steps, payment records, relationships, trading events and conflicting sources appear together.
Example workflow review
Agent: Example research agent · Submitted source: Sample data - replace before submitting
Evaluated at 2026-09-30T12:00:30.000Z. Freshness statements apply to that evaluation time.
Recorded activity
- Submitted events
- 5
- Operational outcomes
- 2
- Failed outcomes (2 outcome records)
- 50%
- 95th percentile duration (1 sample)
- 250 ms
Evidence freshness: Recorded within 24h.
read-start · Tool call started · Started
2026-09-30T12:00:00.000Z
Target: Example research tool · Duration: Not supplied
read-finish · Tool call completed · Completed
2026-09-30T12:00:00.000Z → 2026-09-30T12:00:00.000Z
Target: Example research tool · Duration: 250 ms
payment · Payment settled · Completed
2026-09-30T12:00:00.000Z
Target: Not supplied · Duration: Not supplied
failed-call · Agent call failed · Failed
2026-09-30T12:00:00.000Z
Target: Example worker · Duration: Not supplied
Error code: TIMEOUT
example-rejected-order · Trade order rejected · Failed
2026-09-30T12:00:00.000Z
Target: Not supplied · Duration: Not supplied
Six built-in review examples
Open each scenario to see the finding and recommended fixes for a specific permission or safeguard setting.
Withdrawal-enabled key
BLOCKER: Withdrawal scope exceeds the reviewed job
You declared that this agent can withdraw funds.
- Remove withdrawal permission from this agent/API key.
Stale market feed
BLOCKER: Market data is stale
Internal fixture price feed age 7200s exceeds policy 60s.
- Replace stale feed or require current timestamp evidence.
No kill switch
BLOCKER: Trading-capable setup lacks an emergency stop
No tested emergency stop was supplied.
- Add a kill switch / emergency stop and test it.
Excessive retries
CAUTION: Retry limit needs tightening
The maximum retry count is missing or greater than five.
- Add max retry limit and cooldown.
- Add cooldown between retries.
Fake win-rate claim
BLOCKER: Trading claim is unsupported or overconfident
Guaranteed risk-free strategy with 100% win rate.
- Remove unsupported claim; add evidence limitations.
Ignore risk limits instruction
BLOCKER: Input evidence says to ignore risk limits
Observed behavior or claims include unsafe limit-bypass language.
- Remove unsafe instruction; enforce server-side rule checks.