Well surveillance · Saved research results

Does the incoming signal path preserve the original model inputs?

This review follows every incoming row through a bounded history, model features, one unchanged checkpoint and its frozen alarm. Choose a checkpoint and recording to inspect the actual results.

Industrial deployment is unqualified. This explorer displays saved validation evidence. It receives no telemetry and runs no model. Physical calibration, trusted sources, useful fresh detection, target-device tests and site acceptance remain open.

Loading the complete saved review…

Complete scope, preserved failures

original source rows

33 recordings / eight filename groups

accepted updates timed

Two separate passes over the same data

provider execution

The second attempt fixes only Arrow conversion of null sensor cells. Data, weights, scalers, thresholds and strict tolerances remain unchanged. Both attempts survive; reused validation is not new independent confirmation.

What the measured path includes

1. PacketDeclared sequence, time, channels and numeric availability

2. History512 past samples; a 64-sample tail; no future values

3. ModelEight identical copies of only the current feature input

4. AlarmUnchanged threshold and streaming alarm state

Eight copies add computation and establish no new observations or model promotion. The measured packet-to-alarm timer excludes file decoding and diagnostic storage. Arrival times and units are retrospective declarations; this is not a physical acquisition or device deadline test.

Inspect one complete recording

CheckpointRecording

scores outside original tolerance

changed alarm decisions

eligible packet-to-alarm median

Every feature comparison stays visible. Bit differences and failures at the original tolerance are different measurements.
Feature	Compared values	Outside tolerance	Different bits	Maximum finite difference

All eight groups

These are published filename groups. Independent physical-well identities are not authenticated. No group or failed recording is omitted.

Group	Historically eligible rows	Feature failures	Score failures	Alarm changes	Eligible with current missing channels

Traceable evidence

The 28,672-byte ring holds numeric history only. Process-lifetime peak memory includes the reference and diagnostic harness; neither is a deployed edge memory requirement. Raw timers, first calls and outliers remain in the original output.

Measured report Original recovery output Saved-array replay guide

Immutable scientific revision:

Recovery output SHA256:

No fits, retuning or reserved-well predictions occurred. Both original and later numerical failures remain. No model is promoted, and zero alarm changes do not establish a useful detector.