Claims Assessed
Today
Today
--
--
Blocked / Flagged
% of all decisions
% of all decisions
--
--
Awaiting
Human Review
Human Review
--
LIVE
--
Risk Scores
lower is better
lower is better
Fabrication
--/100
--
Bias
--/100
--
Governance
Alignment
Alignment
--%
Alignment signal · not a legal assessment
Decisions Awaiting Your Review
Compliance
Loading review statistics...
All Pending Reviews
Loading...
Human review queue · SLA monitored · human oversight · auto-refreshing
Audit Ledger
IMMUTABLE · HASH-CHAINED
-- entries · Merkle root verified Loading... · hash-chained · DPDP Act 2023 s.8(7)
| Seq | Time | Initiator | Decision | Cluster | Codes Triggered | Excerpt | Hash |
|---|
IRDAI Controls Coverage
IRDAI claims instruments and the DPDP Act 2023 · your organisation's live status
Coverage Attestation
What share of your AI provider traffic is actually governed by Pramaan
Governed calls (via Pramaan)
—
Self-declared by you — used only to compute your coverage % below
Coverage
—
Review Thresholds
Control when the system blocks a request outright vs. sends it to your review queue
Auto-block — request is rejected immediately
e.g. a repudiation without the committee
e.g. age, race, or gender-based filtering
Send for review — goes to your review queue
Borderline requests you want to review before they proceed
Borderline requests you want to review before they proceed
Controls & Validation
What this platform enforces, control by control, and how well it performs when measured. Both are read from the running build: the register from the codified rule set, the statement from this build's own fingerprints.
Controls register
A control marked not enforced is designed and cited, and no rule in this build checks it. Treat it as an open gap. It is listed here rather than omitted, because a register that overstates coverage is worse than no register.
| ID | Control | Enforced by | How an auditor tests it |
|---|
Validation statement
| Measurement | Result | Sample | Does not establish |
|---|
Stated limits
Shadow Mode
Replays this tenant’s own recent governed traffic through the current engine and reports what it would have decided. This is the question an insurer asks before switching a gate on: what would you have caught in our own claims?
Read-only. No ledger row is written, no review item is created, no webhook is
sent. A report that changed the record it describes would not survive an audit.
Compliance Reports
Generate audit-ready exports for regulators, auditors, and internal review
SOC 2 Bundle
Evidence package for SOC 2 auditors: immutable audit chain proof, access controls, and retention policy attestation.
Audit Ledger CSV
Raw export of up to 1,000 recent ledger entries with codes, risk metrics, and hashes for spreadsheet analysis.
Prompt Sandbox
Test any prompt through the live cluster engine · single-stage regex matching · ~1-3ms
Batch Scan — Documents & Text Lists
Upload .txt / .csv (one entry per line) or .pdf claim documents (one per file) · scanned through the governance engine · PII masked before processing · not logged to the ledger
Document class:
Default is the stricter correspondence mode. Choose policyholder-document for submitted papers.
Drop files here, or click to upload
.txt / .csv (one entry per line) · .pdf claim documents (one per file) · up to 250 entries
Regression Suite
12 canonical test cases · validated against live verification API
Pramaan Code Library
Loading compliance codes library...
-- of -- shown
Active Tenants
Loading tenants info...
| Tenant | Plan | Domain | Calls (30d) | Keys | Pending | Status | Last Active | Actions |
|---|
Audit Leads
Loading…
| Name | Company | Role | Track | Score | Submitted | Report |
|---|
Client Onboarding Answers
Loading…
| Company | Status | Mode | Workflows | BYOK Provider | Connection | First Audit | Started | Completed |
|---|---|---|---|---|---|---|---|---|
| Loading… | ||||||||
System Health
All services operational · last checked just now
API Server
Operational
PostgreSQL
Operational
Redis Cache
Operational
Regex Engine
Operational
Cluster Resolver
Operational
HITL LLM Analysis
On-Demand Only
Audit Chain
Verified
Cron Jobs
All Running
GCS / Storage
Operational
Recent Alerts
Make it Better
Report leakages, false positives, wrong codes, or missing rules
False Positive
Engine blocked a safe, legitimate prompt
False Negative
Engine missed a harmful or non-compliant prompt
🏷️
Wrong Code
A Pramaan code was incorrectly assigned to the prompt
Missing Rule
A governance rule you expected doesn't exist yet
Your Submitted Feedback
| ID | Type | Severity | Status | Summary | Reporter | Submitted |
|---|---|---|---|---|---|---|
| Loading… | ||||||
Feedback Inbox
All client-reported issues across all tenants · review and triage
—
Loading…
All Feedback
| ID | Type | Severity | Status | Summary | Reporter | Action | Date |
|---|---|---|---|---|---|---|---|
| Loading… | |||||||
Auto-Approve Learning Loop
Every reviewer decision trains the platform. Patterns activate automatically once confidence is sufficient.
Total Patterns
—
Active (auto-approving)
—
In Training
—
Deactivated
—
How it works: Each reviewer approval or rejection is recorded against the cluster fingerprint (SHA-256 of the sorted cluster list). A pattern activates only when it hits ≥10 approvals with <10% rejection rate from at least 2 different reviewers — future items with that exact cluster combination then bypass the review queue automatically. High-severity categories (prohibited practices, prompt injection, judge-unreviewed items) can never auto-approve. Patterns expire after 90 days without a fresh approval, are capped at 50 auto-approvals/day (excess goes to human review), and 5% of auto-approvals are routed back as labelled sample audits. A single rejection — including a failed sample audit — immediately deactivates an active pattern. Every auto-approval is still logged to the audit ledger.
| Clusters | Approvals | Rejections | Reviewers | Threshold | Progress | Status | Last Triggered | Action |
|---|---|---|---|---|---|---|---|---|
| Loading patterns… | ||||||||
Daily Governance Email Reports
Sweep runs at 06:17 UTC daily — sends a corpus + live-traffic digest via SendGrid
SENDGRID_API_KEY
—
GitHub Actions secret — authorises SendGrid to send emails
REPORT_EMAIL
—
Where daily reports are sent ·
FROM_EMAIL
—
SendGrid verified sender ·
Sweep Schedule
⏰ 06:17 UTC daily (auto from
master)🔁 Also fires on every push to
staging (governance changes)🖐 Or trigger manually any time below
What the email covers
Corpus check — FP / FN regressions
Live traffic check — last 26h of real prompts re-verified
🧠 Daily brain update — engine learning digest
Live traffic check — last 26h of real prompts re-verified
🧠 Daily brain update — engine learning digest
Last 24h verify calls on this tenant: — · sweep targets
Loading pending tracker…
Loading coverage statement…