Reproducibility · Public well-event research

Replaying every validation decision.

33 recordings, eight filename groups and 121,325 eligible decisions. Compare original offline scores with two ways to process one available input.

One included cloud CPU job4 exact profiles · 2 failed profilesIndustrial use unqualified

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.

The current-only eight-copy method matches every examined score.
Ordinary single-reading inference fails the unchanged numerical tolerance. Eight copies of the same current input reproduce the original scores, with extra computation. This finite validation result promotes no model; all eight reserved groups remain unopened.

Loading the frozen replay results…

Saved checkpoint

Processing method

Scores outside tolerance

Alarm changes

Median model call

Maximum retained call

All six frozen profiles

Every profile is retained. Passing profiles match all 121,325 eligible scores exactly; failing profiles keep the original atol 1e-10 / rtol 1e-9.
Checkpoint	Processing method	Failed scores	Alarm changes	Median call	P95 call	Maximum call

Check each filename group


Public filename group	Eligible decisions	Scores outside tolerance	Maximum score difference	Alarm changes

What these matches establish

Original offline inference processes all 140,708 grid rows in batches of 256, then masks unusable inputs. The final batch has 164 eligible rows. Both current-only methods use an explicit contiguous boundary; the eight-copy method consumes only the same one available feature, with no future or neighboring row, waiting or cached substitute.

Model, scaler and threshold bytes stay unchanged. Full grid identities and first-128 stored feature bytes match. This is a numerical adaptation on observed validation data, with added compute. A complete original feature tensor was not stored; the newly recorded full feature hash is not an original full-tensor comparison.

The original Mac aggregate replay stays FAILED. Internal kernel causes remain unisolated. Neither useful fresh detection nor physical calibration is established; no inference profile is automatically admitted.

Compute and reproducibility boundary

Every raw call time and first call is retained. Profiles run once in fixed order; timings exclude acquisition and rolling raw-feature work. Lifetime memory includes full-stream matrices and runtime, and is not incremental edge-model memory or a target-device deadline.

Separate saved-array review verifies 727,950 eligible scores, 844,248 full-grid alarm values, 486,400 raw call times, 48 group rows and twelve common/native metric reports. It runs no learned inference or raw measurement decoding and has no independent human sign-off.

Inspect and reproduce

Full methods and failures Original output archive Saved-result replay guide All saved profile results

Original output SHA256:
Scientific artifact revision: