TL;DR: Pre-failure noise in degrading mechanical bearings is famously predictive of time-to-failure. We asked a stricter question: is that structure directionally tied to the true future failure event, in a way present-state structure alone couldn't explain? We ran five falsification tests against six public bearing-degradation datasets to check. Every test failed to support directionality. What looked like a "signal from the future" was just the present being highly structured. This is independent research, not peer-reviewed — treat it accordingly.
The question
Rotating machinery gives off a lot of warning before it dies. Spectral entropy shifts, resonance patterns change, noise gets structured in ways that correlate strongly with time-to-failure. That correlation is well documented in the predictive-maintenance literature and isn't controversial.
What is a stricter question: is that structure oriented — does it point specifically at the true, upcoming failure event, in a way that couldn't be explained by the present state alone?
If the answer were yes, it would imply something stronger than "structured degradation." It would imply the current noise is carrying information specific to a boundary condition that hasn't happened yet. We named this the Dead Future Hypothesis, mostly as a label for the null we wanted to knock down — formally:
Pre-failure mechanical noise acts as a carrier for information flowing from the future terminal failure boundary. The adhesion cascade exhibits unique alignment with the true terminal state — alignment that should degrade under time-reversal, false terminal assignment, or cross-terminal label swaps.
If that's true, five specific things should happen. If any of them don't, the directionality claim collapses back into "the present is just really well structured" — no future-facing signal required.
The setup
We used six public bearing-degradation datasets:
- PASS runs (clean, ordered degradation cascade confirmed): NASA IMS Run 3, PRONOSTIA Bearing 1_1, PRONOSTIA Bearing 3_2
- FAIL runs (transient collapse only, no ordered cascade): NASA IMS Run 2, PRONOSTIA Bearing 1_2, PRONOSTIA Bearing 2_2
Earlier passes (v0.3–v0.5) had already established that PASS runs contain a reproducible adhesion cascade: a low-frequency band (the "Vanguard Band") lights up first, then the cascade climbs through broader frequency bands as failure approaches. That structure is real — it survives null-boundary controls and drift-matched virtual controls. The open question was whether it's directional.
Every test below scores structure using AUC (area under the ROC curve) from a simple decoder trained to separate a target window from the safe/early-life zone. 0.5 = indistinguishable from random, 1.0 = perfect separation. High AUC just means "this window is statistically distinguishable from normal" — it says nothing on its own about why. That's exactly the ambiguity the five tests below are designed to resolve.
The verification — five falsification tests
Each test predicts a specific, falsifiable signature that a truly directional signal must show. We ran all five (v0.6, the "Directional Adhesion Stress Test").
Test 1 — Time Reversal
Reverse the temporal order of each run, recompute every adhesion metric. A directional signal should collapse under reversal — AUC should fall toward the ~0.50 null baseline.
Forward AUC (avg): 0.940
Reversed AUC (avg): 0.940
Δ(t): 0.000
The reversed-time cascade was statistically indistinguishable from the forward one. In two of three PASS runs, the reversed score was marginally higher than forward — within noise, but the opposite of what directionality predicts.
Test 2 — False Terminal Assignment
Assign artificial "terminal" boundaries at 30% and 50% of bearing life, using the same drift-matched null controls as the true terminal. A directional signal should score the true terminal far above these fakes.
True Terminal AUC: 0.940
False Terminal AUC (max): 0.923
Specificity percentile: ~50th
Fake failure points scored essentially as well as the real one. The cascade aligns with any structurally compatible boundary — it doesn't discriminate the true future event from an arbitrary mid-life one.
Test 3 — Cross-Terminal Specificity
Swap terminal labels across runs — apply NASA IMS Run 3's failure boundary to PRONOSTIA 1_1's data, and vice versa. A signal tied to one specific physical failure event should collapse badly under the wrong label.
True Label AUC (avg): 0.910
Swapped Label AUC (avg): 0.873
Cross-Terminal Δ: 0.037
A 0.037 drop is noise-level. The cascade is a general property of bearing degradation in PASS runs, not something wired to any individual physical failure event.
Test 4 — Vanguard Band Stability
Stress-test the Vanguard Band under random seeds, resampled windows, ±5% boundary shifts, and alternative null controls, to check whether it's a real physical structure or an artifact.
Forward stability: 0.96
Reversed stability: 0.94
Seed variance: ±0.02
FAIL-run stability: 0.31
The Vanguard Band is real — it's stable and clearly separates PASS from FAIL runs. But it's equally stable under time reversal (0.94 vs 0.96), which confirms it's a genuine physical structure while simultaneously confirming it has no temporal orientation. This is the one test we call PARTIAL rather than a clean fail: the structure itself checks out, it just isn't directional.
Test 5 — Cascade Order Robustness
Represent the ordered band-activation sequence as a vector, compare forward / reversed / false-terminal / cross-terminal conditions against the canonical forward cascade using cosine similarity. Only the real forward condition should score high; everything else should be scrambled.
Forward / Reversed / False-terminal / Cross-terminal: 0.85–0.92
Null-shuffled baseline: ~0.51
All four "real" conditions cluster together. Only literally random shuffling produces a different signature. The cascade order is an intrinsic property of how the degradation unfolds — not a fingerprint of a specific future event.
The result
| Criterion | Required for directionality | Result |
|---|---|---|
| Forward > Reversed | Yes | Δ = 0.000 — FAILED |
| True terminal > False terminals | Yes | ~50th percentile — FAILED |
| True labels > Swapped labels | Yes | Δ = 0.037 (noise) — FAILED |
| Vanguard Band stable only in PASS, not under reversal | Yes | Stable in both — PARTIAL |
| FAIL runs show no ordered cascade | Yes | Confirmed, but doesn't establish directionality — FAILED |
Across the directional criteria, every test failed to support temporal orientation. The Dead Future Hypothesis is empirically rejected.
What this actually means
The interesting part isn't "retrocausality is fake" — nobody serious expected otherwise. The interesting part is that the thing we did find is still real: bearings exhibit a highly ordered, reproducible, statistically significant structural cascade before they fail. It's just undirected — a non-stationary spectral critical transition, not a message from the terminal event.
In other words: the present state of a degrading system is already so deeply structured that it deterministically constrains its own likely futures. That looks like prediction. It even looks a little like the future reaching back. But the arrow isn't there — you can run the tape backward and the structure holds just as well. The predictability was fully present in the current moment all along.
Caveats
This is independent research, not peer-reviewed. The datasets (NASA IMS, PRONOSTIA/FEMTO) are public bearing-degradation benchmarks commonly used in predictive-maintenance literature — full credit to those projects. Treat the falsification result as a solid negative result on a fun hypothesis, not a physics paper.
Methodology / reproducibility
The core scoring approach is a standard AUC-based decoder (separating a target window from the early-life safe zone) run per frequency band, cross-checked against drift-matched virtual null boundaries so the score can't be explained by generic drift alone. Each of the five tests above just changes what gets fed into that same scorer — reversed time order, a relabeled terminal, a swapped terminal, a perturbed band definition, or a shuffled order vector.
Code cleanup for public release is still pending — the current implementation is entangled with unrelated internal tooling. Full report and figures: Pinching Index — DIMProductions Research Archive
Top comments (0)