Roles

Roles: reasoning view of MAFAULDA-rotor — annot is the source label (imbalance / misalignment), kept machine-parseable as the gold for verification and reward parsing; the model reads query + image, where the image is a vibration spectrum with the shaft-order positions marked. The reasoning column is filled on all 706 records and is the SFT imitation target for this repo; the query enumerates the closed set of labels the answer must come from, and annot remains the string that is scored. metadata is provenance (split keys, source identities) and must never be fed to a model.

MAFAULDA-rotor-annotated

Shaft-order amplitude spectra from the UFRJ MaFaulDa rotor rig, as a binary classification task: imbalance vs misalignment. 706 records. reasoning is filled on every record; the twin repo MAFAULDA-rotor is identical except that field is empty.

The reading

Add the heights of 2x, 3x, 4x and 5x — call that the harmonic ladder — then ask how many times taller 1x is than that sum.

R = 1x / (2x+3x+4x+5x)	verdict
≥ 2.663	imbalance — 1x owns the energy, the run behind it has collapsed
≤ 1.064	misalignment — the run has risen above 1x
in between	not decidable; the record is not in this corpus

Both quantities are on the picture: every height is a multiple of the noise floor, the grass between the lines. Physically, an eccentric mass throws once per revolution and feeds only 1x; two shafts out of line are bent twice per revolution by the coupling, which lifts 2x and with it 3x and 4x.

On the shipped set the two sides do not overlap — imbalance all ≥ 2.663, misalignment all ≤ 1.061, a factor of 2.5 clear.

Retention — read this before using the accuracy

1,258 confirmed spectrum records minus 80 normal leaves 1,178 candidates. The rule speaks on 811 of them (69%) and is right on 0.871 of those; on an assembly group it has never seen, 0.844. The 706 shipped are the ones it spoke on and got right — the rest were dropped by looking at the gold.

This is the corpus's own precedent (OTTAWA-VARSPEED keeps records whose blind verdict is consistent with the gold; the five bearing sets keep evidence_tier == confirmed), but it means the rule is not an off-the-shelf detector: quoting 100% on this set would be quoting the selection.

What was dropped, and why

normal is not separable on this figure. Every rotor carries residual imbalance, so 1x is the tallest line on 80% of healthy records too, and the healthy 1x range (p10–p90 of 8.1–30.8 × noise floor) sits almost entirely inside both misalignment classes.

horizontal vs vertical misalignment is a static preload direction. A piezoelectric accelerometer is AC-coupled (−3 dB at 0.27–0.5 Hz) and blind to a d.c. quantity; the one vibration plot that shows preload direction — the shaft centreline plot — needs the d.c. gap voltage of a proximity probe, which this rig does not have. The published diagnostic vocabulary has no such pair of categories: ISO 13373-1:2002 Table C.1, the Technical Associates diagnostic chart and the SKF Vibration Diagnostic Guide all split misalignment into angular and parallel only, and both MaFaulDa classes are parallel. Measured here: four blind judges reading these figures scored 41.7% on the four-way task (chance 25%), swapping 12 of 30 misalignment images between the two subtypes; and a random forest over six accelerometer channels collapses to 0.000 on one held-out severity group while a random split reaches 0.938.

The upstream four-class label is kept in metadata.upstream_label_4class as provenance. It is not recoverable from the image.

Figures — one of 11 drawing styles per record

Every reasoning image in this corpus used to be drawn one way, which makes "dark blue line on white with amber dashes and a centred title" a constant of the training set. Each record here draws one of 11 styles from a hash of its own image path. What varies: colours, canvas size, title placement, tick density and direction, grid, spines, fill, marker form (line or band), label layout, font family, scientific notation. What never varies: the curve, the plotted band, which marks are drawn, and the y-limits (autoscaled once per record and reused by every style, so a peak is the same height whichever style drew it). No style draws a threshold line or a noise-floor line, and no line is styled differently because it happens to be the answer.

Ten figure-surface checks ran on the rendered pixels of all 1,258 images — name collisions, title over name, curve through glyphs, name off the axes or off the canvas, legend over a mark, exponent over title, degenerate y-ticks, name over tick labels, apex occlusion — all zero.

How reasoning was produced

This revision (2026-09-25) replaces the reasoning column; earlier versions remain in the repository history.

A deterministic renderer, not a language model. Each text is assembled by fixed code from the record's metadata and its gold label. No API, no sampling, no model output: rendering the same record again gives the same bytes.

Four paragraphs, separated by blank lines, with no headings inside the text:

1. Figure reading — what the figure is (the amplitude spectrum of the rotor vibration to eight times the shaft rate, with marks at 0.5x, 1x, 1.5x, 2x, 3x, 4x and 5x) and the height at every mark as a multiple of the noise floor, which mark is tallest, which of 2x–5x is tallest, and what 2x–5x add up to. Description only: no explanation, no verdict.
2. Physics — both candidate classes explained in one continuous derivation: imbalance as a centre of mass off the axis, a rotating centrifugal force seen as a sine at the shaft rate, so its energy sits on 1x; misalignment as a parallel or angular offset carried by a coupling whose stiffness repeats every half turn, so the force rises and falls twice per revolution and, being uneven, spreads from 2x into 3x, 4x and higher orders. It adds that residual imbalance keeps a 1x line in every rotor and that the half orders belong to looseness and rub, not to either class. Every term is defined where it is first used. The content of this paragraph does not depend on the record's label.
3. Linking — 1x on its own settles nothing, so the four harmonics 2x–5x are weighed together against 1x; the paragraph states how their sum compares with 1x for this record and why that share of energy fits one mechanism and not the other.
4. Conclusion — one sentence, then a blank line and FINAL ANSWER: <annot>, equal to the record's gold label byte for byte.

Variation comes from the image, never from the label. Paragraphs 1 and 2 swap order for roughly half the records (the exact share is given below); the linking paragraph and the conclusion always come last. Each paragraph has at least three phrasings; in this set the phrasing is chosen once per paragraph, and the order in which the marks are read in paragraph 1 varies with it. The swap and every choice of phrasing are taken from the record's image hash, image_sha256 in metadata, never from the label.

No criterion in the text. No decision threshold, class boundary or per-class range appears anywhere in the text, and no reading is compared against one. Heights are multiples of the noise floor rounded down to a whole number (in steps of five from 50 up), and the comparison of 1x with 2x–5x is given as a rounded-down ratio or in words; the same printed comparison never occurs under both labels.

Numbers for this set. 706 records; median length 542 words (whole text, FINAL ANSWER line included); the physics paragraph comes first in 360 of 706 (51.0%). All 706 texts are distinct, and every text ends on its gold label (checked on all rows).

Example. The linking paragraph (paragraph 3) of test record 59 (0-based row of the test split, gold imbalance), verbatim:

Set against the physics, the height of 1x on its own settles nothing, because every rotor keeps some 1x. The difference lies in how much energy the harmonic orders carry relative to 1x, and since a coupling spreads its energy over several orders, 2x to 5x are weighed as a group. Here 1x stands about 7 times as high as the four harmonics together. The energy has stayed on the once-per-revolution line, as it does when the only new force is an eccentric mass turning with the shaft; a misaligned coupling would have pushed a far larger share into 2x and the orders after it.

Fields

field	meaning
query	the classification instruction (one of 30 deterministic paraphrases)
image	the rendered shaft-order spectrum (bytes embedded)
annot	imbalance or misalignment
reasoning	the chain of thought, closing on a line FINAL ANSWER: <class>
cate / task	C / T-C1
metadata	JSON: representation, channel, fs, fr_hz, window, harmonic_profile, noise floor, ladder_sum, rule + R + band, upstream_label_4class, severity, style_variant, y-limits, image_sha256, split

Split

Severity-wise, inherited from the producer: one severity level is one physical assembly, so a file-wise split would leak. train / test counts: test 139, train 567.

Rig

SpectraQuest Alignment-Balance-Vibration trainer (UFRJ MaFaulDa): 1/4 CV DC motor, 16 mm × 520 mm shaft, two bearing housings, 8 channels at 50 kHz for 5 s. The spectrum here is the overhang tangential accelerometer, 1 s window, raw amplitude — not an envelope: a rotor fault is an additive synchronous component and Hilbert demodulation would discard it. Drawn to 8× the shaft rate with the seven shaft orders 0.5x / 1x / 1.5x / 2x / 3x / 4x / 5x marked. Every file runs at a different speed (12.3–61.8 Hz), so every record's abscissa is a different length.

Licence

The upstream page states no licence (checked: no copyright notice, terms or citation requirement). This repo is therefore gated manual and carries this warning: clear the upstream licence before any onward redistribution. Source: http://www02.smt.ufrj.br/~offshore/mfs/page_01.html