//OPERATIONS
Inspectable operations, at portfolio scale.
Every number below is read live from the same PostgreSQL tables the pipeline writes to while processing requests. Small numbers are expected: this is a portfolio sandbox, not a production tenant with thousands of leads per day. The point is to make latency, spend, retries, and stuck work inspectable.
The snapshot includes the seeded 60-case evaluation run and guided demo replays; it is not customer traffic.
//OPERATING SIGNALS
No stuck workflows
All leads resolved within the 2-minute expected-duration window.
No duplicate writes prevented recently
No replay/duplicate-key collisions in the last 24 hours.
Unsupported-fact rate 0.0% — within target
0/40 drafts had an unsupported claim (target: below 2%).
//BREAKDOWN
Leads by status
- completed 64 (100%)
Leads by qualification band
- Sales-ready 17 (27%)
- Needs review 15 (23%)
- Insufficient evidence 13 (20%)
- Disqualified 11 (17%)
- Nurture 8 (12%)
Identity resolution outcomes
- New (no match) 57 (89%)
- Confident match 6 (9%)
- Possible match 1 (2%)
Outreach draft decisions
- pending_review 39 (98%)
- approved 1 (2%)
//LATENCY PER STAGE (real, from audit_events)
| Stage | Runs | Avg duration |
|---|---|---|
| draft_outreach | 40 | 2.48s |
| extract_facts | 44 | 2.07s |
| verify_facts | 44 | 1.49s |
| classify_message | 64 | 1.48s |
| check_draft_claims | 40 | 1.05s |
| identity_resolution | 64 | 1ms |
| find_research_source | 56 | 1ms |
| qualify | 64 | 1ms |
| propose_crm_change_set | 64 | 1ms |
Scope note
This is the v1 scope: real audit-event data plus 3 alerts (stuck workflows, duplicate-write prevention, unsupported-fact rate). CRM API error tracking, qualification drift over time, and operational controls (pause outbound, replay dead-lettered workflows) are documented as design, not built — an honest cut line, not a silent shortfall.