Frozen causal well comparison: verified research release
This release adds the complete fifteen-method training/validation first phase to the existing well research model and static result explorer. The original eleven-method consumed-confirmation study stays separate. No reserved values or predictions were opened, no candidate was promoted, and industrial deployment remains unqualified.
Immutable scientific content
| Artifact | Content revision |
|---|---|
| Model, checkpoints, source and complete result audit | 8f5707571840145eec1b53da22dd53adfd8814d7 |
| Static source and saved-result data | 56f1ec79461672d52b5294d0da587b226ab1c4d9 |
Open the live comparison. Current live hosting is a mutable presentation; original scientific content and replay inputs are pinned above. Subsequent receipt additions preserve all these payloads except explicitly appended repository-card prose. Final receipt commit identities are recorded in the local publication manifest; each public receipt document is identified by its enclosing Hub revision.
Original output ZIP SHA256:
12a07749800af0926b7055b9d4d6c2a7228c0c4e1b4b08cc18367663ce02d459.
Original selection SHA256:
e3be4520f81ba67c4d4c100f9d895d928ce0ae86b1a20a2d6d519207f9923954.
All seventy original members, byte-identical input kits, eleven fitted compact
numeric checkpoints, full pretrained TSPulse and source/license notices remain
available. Replay without fitting.
What the study found
The validation-selected signal logistic covers 17/24 episodes and discovers 13/24 known onsets within 300 seconds. Eleven discoveries are in one filename group; the other seven groups contribute two discoveries across eleven episodes. At common support it has two false normal alarm episodes over 12.573 eligible normal days and 0.622% eligible normal time in alarm. Group resampling gives a wide onset-discovery interval, approximately 5.0%–85.0%; it does not authenticate physical well independence.
The complete pretrained causal TSPulse adaptation covers 1/24 episodes and discovers zero new known onsets within 300 seconds at its selected budget. Disabled arms, unsupported episodes, warmup, unknown labels, common/native eligibility, all searches and every group's outcome remain visible. This is an original causal adaptation and selected validation analysis, not a paper replication, fresh confirmation, pre-onset forecasting or field diagnosis.
Verification and failures
- All 85 model payloads and five Space payloads match anonymous fixed-revision downloads. Prior scientific/source objects are unchanged.
- Actual execution from those public downloads rebinds all seventy archive members, verifies 437 threshold searches, all fifteen 140,708-decision grids, both eligibility cohorts and their 1,000-replicate group intervals.
- The aggregate Mac model replay remains FAILED. Thirteen arms pass; both
TCNs fail original
atol 1e-10 / rtol 1e-9. Full pretrained TSPulse passes its separately frozen tolerance. The original Linux provider is COMPLETE; these are distinct outcomes. - Forty-eight frozen saved-checkpoint batch/determinism controls reproduce exactly on the observed Mac host: all 148 saved arrays / 133,504 values match. Every tested configuration still fails the original Linux saved-score tolerance. Determinism changes no scores; batch size does. No configuration is adopted and exact kernel/architecture contributions remain unisolated.
- Actual local and live browsers each pass 38 result cases, plus two navigation cases. All fifteen methods, both cohorts, eight group filters, episode rows and uncertainty denominators match the frozen data. Desktop 1280-pixel and phone 390-pixel layouts pass. Live observed warnings/errors: zero; document/data return HTTP 200, the revalidated script HTTP 304.
- The static runtime is RUNNING with null current/requested hardware. Its only source-byte addition is the declared platform creator-ID script. No CPU or paid Space is created. Original eleven-candidate confirmation behavior and its displayed 4.60 false alarms/day remain intact.
- The first anonymous verifier remains FAILED after catching Hub's
generated storage rule for
foundation-validation/source-corpus-audit.json. A separate reviewed verifier checks the exact added rule and every payload; it does not relabel that failure. Both local browser presentation failures and the original requests text-decoding observation are retained separately.
Remaining gates
The eight reserved groups remain unopened until the inference/reload boundary is qualified and a separate no-fitting confirmation runner is reviewed/frozen. Public-pretraining contamination and physical group identity remain unresolved. No target-device deadline/recovery, physical calibration, trusted site mapping, plant/control acceptance, field adoption or qualified-expert growth is established.
Authored source is MIT; original upstream TSPulse Apache-2.0 notices are preserved. Original 3W source attribution/rights remain in the existing model card. No client architecture, private plant tags or identity documents are published.