Manual baseline (A) vs Automated (B)
Measured off the same audit stream. First-pass yield up and rework down while the exception rate (the guardrail) stays low = quality maintained, not traded for speed. The manual baseline uses representative hand-key timings (labelled, not canned). A and B are aggregate labelled cohorts, not guaranteed one-to-one matched cases; the live run (C) below is the real measured number off the case you drive.
Provenance: Run A is the labelled representative manual baseline; Run B is the labelled automated comparison. Both are computed live from LedgerSink rather than values stored in this page. Run C, when present, contains the case actually driven in the demonstration.
Inspect A · 0 requests / 0 events ↗ Inspect B · 0 requests / 0 events ↗
| Metric · expand for definition and evidence | A (manual) | B (auto) | Δ | ||
|---|---|---|---|---|---|
Touch-time / requestDefinition: Human touch duration per request. Formula: sum(duration_ms where actor = human) ÷ distinct request count ÷ 1,000. Evidence basis: Human-authored LedgerSink events and their recorded duration_ms. Aggregate labelled cohorts—not a matched-pair experiment. A includes 0 request(s) / 0 event(s); B includes 0 request(s) / 0 event(s). Open A evidence ↗ · Open B evidence ↗ | 0 | 0 | — | ||
Transcription steps / requestDefinition: Manual transcription actions per request. Formula: count(action = keyed_field or copied_value) ÷ distinct request count. Evidence basis: LedgerSink action labels; automated pulls are not counted as transcription. Aggregate labelled cohorts—not a matched-pair experiment. A includes 0 request(s) / 0 event(s); B includes 0 request(s) / 0 event(s). Open A evidence ↗ · Open B evidence ↗ | 0 | 0 | — | ||
First-pass yieldDefinition: Share of submitted requests whose audited request history contains no warning or rework. Formula: clean submitted requests ÷ all submitted requests. Evidence basis: All retained LedgerSink events for each request; requires an action = submitted. Aggregate labelled cohorts—not a matched-pair experiment. A includes 0 request(s) / 0 event(s); B includes 0 request(s) / 0 event(s). Open A evidence ↗ · Open B evidence ↗ | 0% | 0% | +0 pp | ||
Rework rateDefinition: Recorded rework events per request. Formula: count(outcome = rework) ÷ distinct request count. Evidence basis: LedgerSink outcomes. Multiple rework events on one request remain multiple events. Aggregate labelled cohorts—not a matched-pair experiment. A includes 0 request(s) / 0 event(s); B includes 0 request(s) / 0 event(s). Open A evidence ↗ · Open B evidence ↗ | 0% | 0% | +0 pp | ||
Cycle time / requestDefinition: Average measured active processing time per request. Formula: average(sum(duration_ms per request)); if no durations exist, first-to-last event elapsed time. Evidence basis: LedgerSink timestamps and durations; requests without enough timing evidence contribute no cycle value. Aggregate labelled cohorts—not a matched-pair experiment. A includes 0 request(s) / 0 event(s); B includes 0 request(s) / 0 event(s). Open A evidence ↗ · Open B evidence ↗ | 0 | 0 | — | ||
Exception rate (needs human) guardrailDefinition: Share of requests that required rejection or return-to-sender handling. Formula: distinct requests with returned_to_sender/rejected action or reject outcome ÷ all requests. Evidence basis: LedgerSink actions/outcomes. The retained human-review control is not counted as an exception. Aggregate labelled cohorts—not a matched-pair experiment. A includes 0 request(s) / 0 event(s); B includes 0 request(s) / 0 event(s). Open A evidence ↗ · Open B evidence ↗ | 0% | 0% | +0 pp |
Live run (C) — measured off the case you just drove
Not a canned demo number: these are computed from the same
run="C" audit events the chat-driven (and page) walkthrough just emitted to
LedgerSink, the moment you ran it. 11 case(s), 87 events.
Open the 87 contributing events in LedgerSink ↗
The one-liner
Operator touch-time is cost of revenue (delivery is bundled into the licence). Cutting it isn't just efficiency — it's a gross-margin lever, and because the same validation now runs at the front, control goes up at the same time.