Well-event transfer audit
Inspect how compact models behave on new well groups. Alarm coverage, early coverage, false alarms and unusable inputs are separate measures.
Industrial admission failedAll eleven tried candidates are retained. This browser reads fixed research results; it receives no live telemetry and issues no commands.
Fifteen retained methods, full pretrained TSPulse, original compact weights and per-group validation outcomes. Both TCN portability failures remain visible; eight reserved groups stay unopened.
Inspect the validation comparison and original evidence
Validation and fresh confirmation
The frozen budgets are ≤1 false alarm per eligible normal day and ≤1% eligible normal time in alarm. Thresholds and the primary candidate were selected on validation, before confirmation. A disabled alarm is not evidence of a useful or safe detector.
Where the outcome is concentrated
| Group | Covered / episodes | Already active | Known onset ≤300s | False alarms | Normal time in alarm | Normal eligibility |
|---|
Coverage is not early warning
Unknown label prefixes censor event onset. A long alarm may overlap many events while also occupying normal time. Inspect the carried-in count and false-alarm exposure together.
What the study does not measure
These labels describe production events. They do not independently establish annulus barrier defects, pump remaining life, flowing-bottomhole-pressure accuracy or successful intervention. Field calibration, trusted clocks and target-device deadlines remain unverified.
The private included-CPU run fitted eight learned checkpoints and scored three references, saved selection and confirmation traces, then failed final JSON serialization. The provider remains ERROR. Report-only recovery used saved scores without retraining or retuning. Missing original resource fields remain unknown.
Complete source audit: 2,228 files / 76,587,318 source rows. Initial study: 641 nonoverlapping real recordings, 16 training / 8 validation / 8 confirmation / 8 reserved groups. Reserved groups have source-audit metadata, but no predictive evaluation in this study. None of these models receives operational or control permission.
All 18 local and 19 included CPU conformance checks pass on authored histories. Complete original weights, masks, end timestamps, saved probes and both cloud run records are available. Reserved well performance remains unmeasured.
Explore the causal adaptation and its sensor-health limit →