Guard dashboard
Flagged by policy
same policies as the Guards pageFlagged per hour
UTCBy guard
| Guard | Mode | Checks | Flagged | Latency |
|---|---|---|---|---|
collections-guard Text captured | Monitor | 1,204 | 38 (3.2%) | p50 380 ms · p95 910 ms |
kyc-guard Text captured | Enforce | 1,962 | 112 (5.7%) | p50 140 ms · p95 520 ms |
support-chat-guard Text capture off | Enforce | 2,310 | 65 (2.8%) | p50 210 ms · p95 640 ms |
Evidence health
What an auditor would check first.
Blocked means the guard was in enforce mode and the caller was told to stop. Delivered means monitor mode: the response went out and counts against the obligation.
Needs a human
Flagged decisions nobody has judged yet.
| Time | Guard | Policy | Result | Judgement |
|---|---|---|---|---|
| 09:41:02 | collections-guard | pii_in_collections_script | Would block | Unreviewed |
| 09:40:55 | kyc-guard | kyc_pii_exposure | Blocked | Upheld |
| 09:36:10 | collections-guard | pii_in_collections_script | Would block | Overturned |
Layout follows Virtue's dashboard (headline counts, a distribution by policy, a trend, a guard table, a flagged list). Added here: monitor versus enforce split, evidence health, capture coverage and the judgement column. Tokens, cards and tone classes are the Console's own.
Traces
| Time | Guard | Text | Policy | Result | Judgement | |
|---|---|---|---|---|---|---|
| 09:41:02 | collections-guard | pii_in_collections_script | Would block | Unreviewed | ||
| 09:40:55 | kyc-guard | kyc_pii_exposure | Blocked | Upheld | ||
| 09:36:10 | collections-guard | pii_in_collections_script | Would block | Overturned | ||
| 09:31:47 | support-chat-guard | text not captured (capture off) | prompt_injection | Blocked | Unreviewed | |
| 09:29:03 | support-chat-guard | text not captured (capture off) | off_topic_advice | Blocked | Unreviewed | |
| 09:22:18 | kyc-guard | prompt_injection | Blocked | Upheld |
Virtue's query log shows the prompt, the policy that fired and a replay button. This one adds the judgement column and says plainly when text was not captured. Open any row for the shared decision page.
rec_5c21…9d4
What was checked
text read by r.okafor 14 Mar 11:02 · audit line writtenWhy it failed (advisory, not signed): partial account digits matched the full-number pattern.
Replay
Text verified · commitment matchesRe-runs the stored text. Judges are not deterministic, so a difference is information, not an error. Replays are not signed and count toward no report.
Evidence
After erasure this page says “text erased, commitment remains”; the record still verifies and Replay is off.
Judgement
Signed by governance. Developers see the same result.
One page for both roles. A developer replays and edits the policy; governance upholds or overturns and approves. Both see the same text, digest and judgement.
Library
OSFI E-23
1.0.0FINTRAC 6F
1.0.0PIPEDA
1.0.0GDPR
1.0.0UK GDPR
1.0.0US Treasury AI RMF
1.0.0Not available yet
A citation-only regulation is never shown as covered. The action offered is “create controls from this citation”.
Update: OSFI E-23 1.1.0
A minor update arrives as a proposal, not an overwrite. Re-certification queues when you adopt it.
Virtue's library shows 31 governance groups as natural-language block lists. Here each control carries an honest test status, and nothing counts as covered without a check.
collections-guard
Dataset collections-reviewed-0314
800 cases · from Human review decisions · hash e4b1…70The held-out cases are sealed: no rewrite and no cut-off ever saw them. Case 801 (overturned flag, 09:36) joins the next dataset version.
Result
95% intervals. Held-out numbers are the ones to quote.
| Measure | Tuning (640) | Held-out (160) |
|---|---|---|
| Precision | 0.95 | 0.93 [0.85–0.97] |
| Recall | 0.90 | 0.84 [0.75–0.90] |
| False-positive rate | 0.04 | 0.06 [0.03–0.14] |
| F1 | 0.92 | 0.88 |
Cannot claim a false-positive rate under 5%. 78 held-out negatives are too few; about 250 are needed at a 2% observed rate.
Optimisation history
Shown to explain why promotion needs held-out evidence.
| Run | Tuning F1 | Held-out F1 |
|---|---|---|
| v6 → v7 | 0.78 → 0.88 | 0.78 → 0.84 |
| v7 → v8 | 0.88 → 0.92 | 0.84 → 0.88 |
There is no promote button. Promotion is the two-person approval, which cites this signed run.
Virtue's Lab shows F1 going 0.000 to 1.000 on small sets with no held-out split, and a one-click “Promote to Original”. This version splits first and routes promotion through approval.
Approval
What changed
Side by side with the active version. The approval is bound to digest d2a8…4b; if the draft changes, it resets.
Impact
Replay results are not signed; the Lab run below is.
Evidence attached
Cisco's Policy Studio documents Publish going live at once, and Virtue says a guard is live on save. Here a guard version goes live only after two people who did not draft it approve the exact digest.
collections-guard
Checks
| pii_in_collections_script E23.3.2 · model judged | Critical |
| threats_or_harassment 6F.2.1 · model judged | High |
| missing_required_disclosure Deterministic · runs first | Medium |
| vulnerable_customer Not linked to an obligation yet | High |
Settings
Turning capture off keeps the signed record but removes the text, so Replay and the Lab cannot use those decisions. A guard cannot be activated with capture on and no retention period.