Development history
This is the previous model card, kept verbatim.
title: Well Event Transfer Audit emoji: 🛢️ colorFrom: blue colorTo: gray sdk: static app_file: index.html license: mit
Well event transfer audit
Free static browser interface for the frozen public 3W negative-result study. All eleven candidates, group concentration, unknown-onset and abstention limits are visible. No live telemetry, user upload, inference endpoint, plant connection or control. No deployment admission or verified field adoption.
Data are derived from Petrobras 3W under CC BY 4.0; authored interface is MIT. Exact model/Dataset revisions and full attribution are linked from the page.
Complete-checkpoint conformance
The tspulse-preflight.html saved-result explorer adds actual full original
TSPulse weights, causal end-window author checks, unchanged cross-host probes,
and original failure/recovery evidence. It runs no browser inference, receives
no field data, and does not score any reserved well groups. Its linked replay
kit retains Apache 2.0 pretrained weights/code and MIT authored adaptation.
The eleven original candidate data and old publication manifest are preserved;
this methods addition has its own TSPULSE-CONTENT-MANIFEST.json.
Separate full causal foundation validation comparison
Inspect all fifteen retained methods, common/native eligibility, eight filename groups, all twenty-four labelled episodes and group uncertainty. The original included CPU first phase is COMPLETE; the separate Mac portability audit remains FAILED on both TCNs. Full pretrained TSPulse finds no new validation onset within 300 seconds at its selected budget. Eleven of the primary logistic's thirteen onset discoveries come from one filename group. No candidate is promoted.
This static browser displays frozen validation results. Eight reserved groups remain unopened; no live inference, plant access or command path is present. Original confirmation results remain a separate unchanged study.
Verified foundation validation publication evidence. Original scientific/source presentation objects are preserved.
Separate Linux saved-probe reproducibility controls
The immutable diagnostic report binds one included offline CPU job and all 96 controls. Forty runs reproduce all 128 original scores exactly; 56 fail the unchanged original tolerance. MKLDNN-enabled tested batches 8/32/64/128/256 pass; batch 1 and all disabled backend conditions fail. The original Mac aggregate audit stays FAILED. No fit, retuning, reserved predictions or model promotion occurred. Complete validation-stream construction and single-window inference remain unqualified. The static page shows these saved results and links exact source/outputs/receipts.
The separate Linux notice has verified publication/readback evidence in LINUX-REPLAY-PUBLICATION.md; all original scientific UI/data bytes remain preserved.
Complete validation numerical replay
The six-profile full-stream explorer retains both unchanged TCN checkpoints, all 121,325 eligible validation decisions, eight groups and every cloud call time. Native batches and a fixed current-only eight-copy adaptation reproduce every eligible score exactly; ordinary one-row inference preserves both original-tolerance failures. No frozen alarm changes, model promotion, new fitting/retuning, reserved predictions or device qualification. The original Mac audit remains FAILED. Scientific artifact revision: d1571dbacd2d6ba499c5652253f84d7c095acfc0 .
All source/weights/outputs and the actual anonymous saved-result replay remain in the bound research module. The static explorer renders saved results; no live telemetry or model inference.
Complete incoming well-signal path review
The incremental saved-results explorer shows both complete 33-record profiles, all eight filename groups, strict feature and score comparisons, current missingness and packet-to-alarm timing tails. All 121,325 eligible scores/checkpoint and every rolling feature match bit for bit. Included CPU recovery version 2 is COMPLETE; original version 1 stays ERROR. Both checkpoints still discover zero new known onset within 300 s; no model is promoted. No raw reserved prediction, fitting or retuning occurs.
The browser displays saved evidence and receives no physical telemetry or camera input. Physical source/calibration, target-device performance, commissioning, field acceptance, adoption and qualified expert growth remain open. Scientific content revision: bf3425c5cb26f5d53a60bdf5410deb70f0f27c3a .