Complete pretrained checkpoint · Causal end-window adaptation

A small model’s causal preflight

The full time/frequency network was executed on authored histories. Explore what its masks, timestamps and reloads establish before a new well-data comparison.

1,084,330 parameters278 tensors · 4.37 MB original weights
Strict full load; no fitting or ignored keys

Complete incoming-signal software review is now available. Inspect all 33 recordings, rolling inputs and full packet-to-alarm timings. Features and eligible scores match bit for bit; both checkpoints still find no new known onset within 300 s. Original failures remain, all eight reserved groups stay sealed and industrial deployment is unqualified.

Conformance passed. Industrial qualification remains open.
This page displays saved results. It runs no model in your browser, accepts no uploads and receives no telemetry. There is no calibrated alarm or well-event performance result for this adapter.

Saved authored input

Past-window end (nominal seconds)

Local raw reconstruction error

Included CPU raw reconstruction error

Channel evidence

Same univariate network per channel; no trained multivariable mixer. These raw scores are not fault probabilities.
Channel	Local raw score	Included CPU raw score	Input support

Eight saved periodic decisions

512 past samples per decision; eight masked patches in the final 64. The aggregate is assigned only to the window end. Internal masked patch positions are not independently causal timestamps.

What was checked

18 local / 19 included CPU checks; every outcome retained.
Unchanged official stitching and separate single-channel calculations.
Future-append invariance, end assignment and fresh strict reload.
Hidden-value isolation and six refusals before any model call.
64 saved score/channel values match at original atol 1e-6 / rtol 1e-5.

Finite authored checks; author verification, without independent reviewer sign-off. Public checkpoint pretraining exposure to 3W remains unresolved.

A low score cannot qualify a sensor

Both constant and zero histories yield an error near 1.58 × 10⁻⁶. A steady process and a sensor stuck at one value can both be reconstructable. Independent measurements are needed to distinguish those physical explanations.

All eight reserved well groups remain closed to predictive evaluation. Earlier consumed confirmation and its failed alarm budgets remain separate and unchanged.

Original failure and separate recovery

Cloud version 1 remains ERROR: Python environment bootstrap failed before model inference. A separately frozen packaging-only recovery used the same weights, probes and 25 offline dependency wheels. Version 2 is COMPLETE, with no numerical or scientific changes.

Two private included CPU jobs in separate budgets; no GPU, TPU, internet, paid compute, field data or plant access.

All three warm-call timings retained. Process-lifetime RSS is descriptive; target-device deadlines and recovery remain unmeasured.
Host	First after load	Warm calls (seconds)	Lifetime RSS

Complete method and results Download weights + replay kit Original IBM checkpoint Primary TSPulse paper Original failure receipt

Kit SHA256: