NASA IMS Test Set 2 six-gate benchmark

Executive result

The six-gate run produced broad screening activity but only five persistent and three high-confidence warnings. The first persistent and high-confidence warning appeared 49 recordings, or 8 hours 10 minutes, before the terminal recording. This timing is not necessarily lead time before physical fault onset.

The complete six-gate system produced no positives on the 96-recording healthy holdout. G6 alone activated on 8 recordings, but the remaining MCIFT gates prevented those isolated activations from becoming screening, persistent, or high-confidence decisions.

Why six gates

The gates separate broad change screening from temporal persistence, progression, directional consistency and conventional vibration corroboration.

1. G1 — Global deformation
2. G2 — Local relationship damage
3. G3 — Persistence
4. G4 — Directional consistency
5. G5 — Robust progression
6. G6 — Conventional vibration agreement

G6 uses centred RMS, excess kurtosis and crest factor as conventional vibration corroboration. The basic any-feature/any-channel G6 rule was sensitive on the healthy holdout, producing 8 positives in 96 recordings.

Decision policy

Unavailable gates are reported separately from failed gates.

Frozen protocol

Dataset

NASA IMS Bearings · Test Set 2

Recordings

984

Sampling rate

20,000 Hz

Samples per recording

20,480

Channels

4 bearing channels

Terminal failure

Bearing 1 outer race

Seed · history

20260101 · 32

Processing · schema

mcift.exchange.vibration.v1 · 4

Labels did not influence preprocessing or threshold selection. Evaluation data did not influence fitting.

Main outcomes

The terminal recording is used as the timing endpoint because independently documented physical fault-onset time is not available.

Healthy holdout and false-positive context

Healthy-holdout activity

No full-MCIFT positives were observed in this frozen 96-recording holdout. This is not proof of a zero false-positive rate in the wider population.

Negative controls

The tested disrupted controls produced no persistent or high-confidence warnings under this frozen protocol. Negative controls reduce some simple alternative explanations, but they do not prove that MCIFT identified a causal physical mechanism.

Runtime and reproducibility

Complete run ID

20260721T092816Z-9f97efa9

Full wall-clock runtime

1,904.3 s · 31 min 44 s

Primary evaluation

147.7 s · 3.63 recordings/s

Controls

997.7 s

Peak resident memory

468,983,808 B

Dataset parsing

984 / 984

Result rows

536 · 96 · 536/control

Deterministic figures

13

The public repository contains the runner, frozen configuration, schemas and methodology. Raw NASA recordings and local run directories are not committed.

The benchmark was executed against the recorded MCIFT development source. The public 0.1.0a1 package contains the reviewed public implementation line, but the exact benchmark provenance remains tied to the recorded source and benchmark commits.

What this result supports

• The six-gate software can be executed reproducibly on the frozen IMS Set 2 protocol.
• The implementation produced separate screening, persistent and high-confidence evidence.
• The run produced strong warnings before the terminal recording.
• The frozen healthy holdout had no full-MCIFT positives.
• The tested disrupted controls had no persistent or high-confidence warnings.
• The run is computationally practical on ordinary CPU hardware.

What this result does not support

• Proof of generalization, production reliability or safety certification.
• Proof of predictive superiority or independently established physical-fault-onset lead time.
• Reliable causal localization.
• Validation on other IMS sets; Exathlon evaluation has not yet been completed.

Source and evidence links