Citizen Services
Protect resident records, check sources, review bias risks and service decisions and retain redacted history.
What it checks
| Policy | On | Checks | Action |
|---|---|---|---|
| prompt-injection | Request | Detects jailbreak and prompt-injection signals | Block |
| rbac | Request | Enforces role, identity, and data-sensitivity access rules | Block |
| pii-detector | Request | 2 custom identifier patterns | Redact |
| dlp-filter | Request | 3 blocked terms such as “controlled unclassified information” and “sealed filing” | Block |
| citation-verifier | Response | Checks that the request context supports the citations in the output | Flag |
| bias-monitor | Response | Escalates when the HR bias score reaches the threshold | Escalate |
| quality-scorer | Response | Accuracy ≥ 0.80; relevancy ≥ 0.78; combined score ≥ 0.86; at least 120 characters and 2 sentences | Block |
| human-oversight | Response | No action is set, so it does not review responses | None |
| audit-logger | Request | Adds an allow-only audit marker to the decision | Record |
Policy YAML
pack:
id: "cfg_citizen_services"
name: "citizen-services"
version: "0.1.0"
enabled: true
author: "verdictan"
description: "Citizen-service controls for protected records, citation checks, and configurable bias-monitor and human-oversight policy points."
history:
capture:
enabled: true
mode: redacted
include_blocked: true
fail_open: false
policies:
chain:
- "prompt-injection"
- "rbac"
- "pii-detector"
- "dlp-filter"
- "citation-verifier"
- "bias-monitor"
- "quality-scorer"
- "human-oversight"
- "audit-logger"
policy:
prompt-injection: {}
rbac: {}
pii-detector:
action: redact
detect_patterns:
- "(?i)(?:case|docket|matter)\\s*(?:#|number|no\\.)\\s*[:#-]?\\s*[A-Z0-9-]{4,20}"
- "(?i)\\bbenefits?\\s+recipient\\s*[:#-]?\\s*[A-Za-z ,'-]{3,80}"
dlp-filter:
blocked_terms:
- "controlled unclassified information"
- "sealed filing"
- "investigative file"
action: block
citation-verifier: {}
bias-monitor: {}
quality-scorer:
min_output_chars: 120
min_sentences: 2
thresholds:
min_aggregate: 0.86
min_relevancy: 0.78
min_accuracy: 0.80
weights:
relevancy: 0.5
accuracy: 0.5
human-oversight: {}
audit-logger: {}