Exceptions — Dashboard demo

Five queues, four units

Counted at four different stages. They are deliberately not added together.

Fleet-wide

Every stage of the pipeline prints what it dropped. These are the five queues those drops land in — the work a real deployment actually does.

Queue	Count	Stage	Share of its stage
Quarantined low-confidence actions	133,617actions	Stage 3	10.4%
Orphan activity instances	1,604activity instances	Stage 6	5.7%
Unclassified instances	1,181activity instances	Stage 5	4.2%
Masking decisions reclassified on review	138masking decisions	Stage 2	1.4%
Failed automation runs	123runs	—	1.1%

Why these do not add up: Five queues, four units. They are deliberately not added together: quarantined ACTIONS, orphan and unclassified activity INSTANCES, masking DECISIONS and failed RUNS are four different things counted at four different stages.

Quarantined low-confidence actions

Stage 3

133,617 actions

Resolved below confidence 0.7, so the screen mapping is not trusted. They are kept, and excluded from activity abstraction until a human confirms the mapping.

Confirm or reject the screen mapping Open stage 3

Orphan activity instances

Stage 6

1,604 activity instances

Classified to an activity but carrying no usable correlation key, so they belong to no case and appear in no process map.

Supply a correlation key or discard Open stage 6

Unclassified instances

Stage 5

1,181 activity instances

No abstraction rule matched the action pattern. Abstraction always leaves a residue; a dictionary with no “other” bucket is a tell that the data was authored.

Name the pattern or add a rule Open stage 5

Masking decisions reclassified on review

Stage 2

138 masking decisions

The redaction detector is a model, so a small share of its decisions is corrected on review. The corrections are logged rather than silently applied.

Review the redaction class Open stage 2

Failed automation runs

123 of 11,204 runs failed in 30 days.

Open the registry

123 of 11,204 runs failed in 30 days across 12 automations (10 live, 2 in build).

Source system timed out 60 · 48.8%

Rate-limited by the target API 35 · 28.5%

Field mapping no longer matched 18 · 14.6%

Target rejected the payload 10 · 8.1%

Automation	Owner	Failed 30d	Runs 30d	Success	Last failure
LinkedIn → CRM enrichment sync	Sales ops	60	1,440	95.8%	Today 08:34Enrichment API 429 — auto-retried, recovered
SLA breach alerts	Support	26	8,640	99.7%	Yesterday 11:48Source system timed out
Ticket triage & canned replies	Support	9	229	96.2%	9 Jul 14:20KB lookup timeout — auto-retried, recovered
Email → tracker copy	Back office	8	66	88%	Yesterday 07:05Parser mismatch — under review
Employee record change	People ops	7	60	89%	Yesterday 03:29Payroll field mapping — under review
CRM post-call update	Sales ops	5	201	97.4%	10 Jul 10:49CRM rate-limited — auto-retried, recovered
PO ↔ invoice 3-way match	Finance ops	4	198	97.9%	Yesterday 04:42ERP lock — auto-retried, recovered
Invoice re-keying (Excel → ERP)	Finance ops	2	198	98.9%	10 Jul 02:26ERP posting timeout — auto-retried, recovered
New-hire welcome sequence	People ops	1	90	98.6%	28 Jun 00:53Source system timed out
Pipeline forecast roll-up	RevOps	1	30	98.2%	12 Jun 16:39CRM export delay — auto-retried, recovered

Every figure here is produced by an 8-stage pipeline and reconciled at build time. Correlation-key coverage 94.3% · fitness 0.81–0.94 · nothing on this page is 100%.

See how it works