Verified full-weight causal preflight release
Verified 4 October UTC / 5 October 2026 Dubai time. Research conformance on authored histories; no measured well-event performance or industrial admission.
- Saved-result explorer
- Immutable complete method and results
- Complete pretrained weights and replay kit
- Exact saved Space source
All 278 original tensors / 1,084,330 parameters actually load strictly.
The checkpoint is 4,374,640 bytes, with source SHA256
57fa03b67d1473a7253ac37801b24c391d1fa931395db986743e2f59e556b9ac.
The declared causal end-window adaptation executes 18 local and 19 included
offline CPU checks. All 64 saved score/channel values across eight probes match
at unchanged atol 1e-6 / rtol 1e-5; largest absolute cross-host difference is
9.085998231572745e-8. Both original source and adaptation scopes are named
precisely in the report.
Original cloud version 1 remains ERROR, before inference at Python environment bootstrap. A separately frozen packaging-only version 2 is COMPLETE, with the same three input hashes, full original weights, implementation, probes, score rule and tolerance. Two private included CPU jobs used separate budgets, with no GPU, TPU, internet or paid resources. No fitting, alarm calibration or real source-feature access occurred.
Actual public readback and execution
All 42 model content files and four Space content files match anonymous fixed-revision downloads. The 69 earlier model payloads outside five reviewed documentation changes and the three earlier Space payloads outside the updated index/card are unchanged. Initial fitted weights, scalers, data, confirmation traces, reports and manifests remain preserved at their original revisions.
The actual anonymously downloaded kit, including complete weights and
source, verifies all 44 members / 43 payload hashes, then executes all 19
preflight/reload checks. It performs no refitting, threshold search or source
reacquisition. Kit SHA256:
06096abdaf30f922ce25219c743597c5d897a4f7aa3af15bb71242cfec8f5e2f.
Local and live browser executions each pass 12 saved-input cases, every one of the 11 earlier candidate views, two actual navigation checks and three desktop/phone layout checks. Seven channel rows and all three original warm timings are retained. No warning/error was captured. Both live HTML responses match reviewed source after only the declared platform creator-ID script. The existing Space is static and RUNNING, with null requested/current hardware; no new CPU or paid Space was created. The browser displays saved outputs and performs no model inference or physical sensing.
The first local browser check preserved a failed link-navigation assertion. The link was given a contiguous padded hit area; the harness now explicitly scrolls it into view and waits for the destination. Exact browser-internal causation was not isolated. The original failure, correction and successful checks remain separate; no scientific artifact was changed by that UI work.
Receipt-only follow-up publishes the linked verification records while keeping
the scientific/content revisions above immutable. Its result is recorded in
RECEIPT-PUBLICATION.json; publication receipts describe their original
revisions, rather than pretending a later receipt was present at first release.
Qualification boundary
Zero and constant authored histories yield approximately 1.58e-6 raw error. This cannot distinguish legitimate steady operation from a stuck sensor without independent evidence. No model or sensor is promoted. All eight reserved 3W groups remain closed to predictive evaluation; earlier initial confirmation is consumed and its failed alarm budgets remain unchanged. Public pretraining exposure to 3W, target-device deadlines/recovery, calibrated source mapping, field acceptance, actual adoption and qualified expert growth remain unverified.
Review and freeze the real-cohort comparison separately, including training-only scaling, strong conventional references, bounded candidates/search/compute, validation-only alarm calibration, gap/warmup support and common/full eligibility denominators before opening reserved predictions. Author verification is not independent reviewer sign-off. Original IBM weights/code are Apache 2.0; the authored adaptation and checks are MIT. Private client architecture is absent.