Files
aimsight-labeling/labeling-rules.md
T

22 KiB

The labeling rules

How to place a label: the only source of these rules for the aimsight-app analyst (before it, the Support-Agent etiqueteur) and for those who review their labels. The families, their boundaries and the catalogue of phenomena are in families.md; what the data mean, in reading-rules.md.

Sources in brackets: [instruction] the section "Les étiquettes" of .claude/agents/etiqueteur.md; [cross ] the cross review of 2026-11-04, auto/revue-croisee/2026-11-04/<family>/revue.md; [finishing ] its finishing pass, auto/revue-croisee/2026-11-04/finition/<folder>/revue.md; [] the station review, auto/revue/<STATION>_<start>/revue.md; [consistency] that of BREITHORN, MICHIGAN and NEBRASKA; [expert] a label or an answer of the expert.

1. What is labeled

  • Everything that stands out from the noise on corr_x or corr_y, over the whole batch: a movement (a spike, a jump, a drift, a wave, an agitation) is labeled when at least one target deviates from its level by more than 5 of its own noises, the threshold of the thermal cycle below (tools/scripts/ranges.py at its default --k 5, with --rolling under a drift or a thermal cycle); a missing point is labeled at any amplitude (the gaps and the rejections below). Under the threshold, it is the ordinary noise of the station: it goes in the analysis, with its deviation in noises, not in a label. An instant without a label means the curves show nothing beyond the noise. [instruction; WYOMING: an agitation under the threshold, labeled then removed according to the expert; the user, 2026-11-17]
  • A gap of a single measurement is labeled No data, at any rate: otherwise, the label of one and the same common gap would depend on the station's rate. [cross No data, CLARIDEN, FLORIDA, LENZSPITZE]
  • A target rejected for one instant is labeled: it is a missing point; the isolated rejections of a target right at the threshold each have their one-measurement label. [LENZSPITZE, SNCF01, cross Target issue]
  • Thermal movement is always labeled, even compensated: the camera pose that follows the temperature gets its label every time, even when the corrected curves are back, even without a daily cycle (tunnel, building: the pose against brief drops of the housing temperature), even on an overcast day (D-117). [NORDEND, WILDSTRUBEL, MICHIGAN, CLARIDEN, cross Thermal displacement]
  • No thermal label on a day when neither the pose nor the corr deviate by more than 5 noises from the night level: there is no cycle to label. [cross Thermal displacement, NIEDERHORN]
  • "No effect on the curves" is checked before labeling nothing: on the residuals of the corr at those instants, and on the number of targets jumping together against a calm moment of the same day; an agitation under the rejection threshold that makes targets flip is labeled. [JUNGFRAU, NORDEND]
  • A signature found once is searched over the whole batch: every cold night, every sunny afternoon, every step of the same shape, with the image of the same day; this is how the forgotten ones were found. [BREITHORN, FLORIDA, JUNGFRAU, NEBRASKA]
  • What falls under a long label is labeled apart: pose steps, isolated rejections, a passage within a day of turbulence have their own label. [SNCF01, LENZSPITZE, cross Obstruction]
  • Targets twinned with their references are not labeled for themselves: they are named with their reference in the labels that touch it. [cross Target issue]

2. The bounds

  • start is the measurement just before the first touched point, end the measurement just after the last, never start = end: a touched point moves or is missing; an anomaly of a single measurement goes from the point before to the point after (a passage at 11:00, 15 min rate: 10:45 → 11:15). [instruction, WILDSTRUBEL]
  • The bounds are measurement instants, at the minute of the measurement: at a 15 min rate, it is the quarter hour; at 10 min or 5 min, the minute of the measurement, like the expert's label. [cross Obstruction, CLARIDEN, expert]
  • The bounds are read on the curves, not on the rejections: a reference that drifts before being rejected, a target that comes back after the last rejection are touched; read the two or three measurements before the first rejection (the error_value of a reference and the pose often already move there), and, for a loss of references at night, the corr_x of the targets and of the neighboring references. [cross Visibility, FLORIDA, LENZSPITZE]
  • Each edge is checked sensor by sensor, in both columns, not on the median: all sensors, not only those named; a corr_x peak can touch corr_y longer. [instruction, NORDEND, NEBRASKA, MICHIGAN]
  • A measurement that has come back halfway is still touched; one more measurement, already back, is too many: read the value of the next step before closing, without rounding inward or outward. [instruction, BREITHORN, JUNGFRAU, CLARIDEN]
  • A range of rejections ends at the measurement following the last rejection of all touched targets: the first rejected measurement is not the start either. [instruction, BREITHORN, NIEDERHORN, SNCF01]
  • A reference rejected at a bound counts as a touched point, unless it is rejected there as usual: its usual rejections (by day, in the morning, a Target issue target) do not move the bound. [cross Obstruction, CLARIDEN]
  • At the edge of a gap, the bound is the first measurement after the gap, or the last before: an event that starts at the resumption starts at that measurement, which is also the end of the No data; a single measurement inside an outage is bounded one rate step on either side. [cross Obstruction, WILDSTRUBEL, JUNGFRAU]
  • Before writing, a script checks the bounds: start < end, they are measurement instants, and no sensor is touched there outside its usual rejections, references included. [cross Obstruction, WILDSTRUBEL]
  • The end of a nuisance specific to a target is read on its corr_x and corr_y, against its level before, over several nights: not on the return of its error_value; without a return, the label runs to the end of the batch; without a level before, it ends where the curves stop changing. [cross Target issue, NORDEND]
  • A return ends where the curve stops coming back: not at midnight; a plateau is not a return. [LENZSPITZE]
  • A label goes to the end of the batch only if points are still missing or moving there: otherwise it ends at the return. [cross Displacement]
  • A lasting jump is bounded on the jump: from the measurement before to the measurement after the one where the new level is reached; hidden in an outage, on the outage and the first measurement after; hidden under another label, at the first measurements where it is seen. [cross Displacement, expert]
  • The bounds of a thermal cycle are read day by day, on the turning points of the pose and of the corr: never at midnight or at a fixed hour, except at the start and the end of the batch; when the pose is noisy, on its hourly median. [cross Thermal displacement, FLORIDA, SNCF01]
  • The camera cycle starts at the morning turning point: the first instant when the tilt and the pan leave their end-of-night level without coming back, the later of the two; not the fast part of the movement, nor the first deviation of days.py --bounds; a deviation of a single rate step from the bound placed stays within the uncertainty. [cross Thermal displacement, NORDEND, finishing nordend]
  • It ends at the later of the two returns, the pose on its main axes and the curves at their night level: an axis returns to the level where it then stays, not under a threshold set on the night, which may differ; the slow relaxation of a secondary axis (often the scale) does not push the end back, as in the expert's labels; without rest, the cycle ends at the next start and the labels follow one another. [cross Thermal displacement, NORDEND, WILDSTRUBEL, expert, finishing niederhorn, finishing wildstrubel]
  • A nuisance at night stops the cycle only if it distorts the pose: a reference rejected, discarded, or that drifts against the targets (overview.py --rigid); a target bothered alone does not stop it. [finishing divers, finishing nordend]
  • A gap does not cut a cycle visible on both sides (so the gaps of one or two measurements); otherwise, a gap or a jump at night stops it: the cycle ends at the rest of the pose if it comes before, otherwise at the last measurement before the gap or at the measurement before the jump. [finishing divers, finishing nordend]
  • Before extending a cycle over a night that moves, read the housing temperature and the references: it must still drop there, or the pose follow its usual profile; a pose that moves at flat temperature, with a reference that drifts, is the nuisance of that reference, not the thermal return. [finishing divers]
  • The reference night of a cycle is chosen calm, not by civil day: when the pose comes back all night, a start is dated at the turning point, not at the first deviation from that night. [SNCF01, cross Thermal displacement]
  • The phases of the structure are cut at the turning points: when the trough is flat for several hours, the cut is placed in the flat and the two phases touch. [JUNGFRAU, cross Thermal displacement]
  • An event that touches the edge of the batch extends the label of the neighboring batch, if there is one: it is not redone; otherwise, it starts or ends at the batch edge. [JUNGFRAU, cross Target issue]

3. A single event

  • One label, a single event: two events, even close or mixed, make two labels; an event that repeats (a passage, the rain of each night, the thermal cycle of each day) makes one label each time. [instruction]
  • The outward and return of a drift are two labels: after each slow drift, look for its return, and do not join a descent and its rise because the description says so. [instruction, CLARIDEN, cross Displacement]
  • The thermal cycle of the camera, outward and return, is a single event per day; the structure makes two, the expansion and the contraction: this is what the expert does; the outward-and-return rule does not apply to the camera cycle. [cross Thermal displacement, expert]
  • An event in one piece over several days makes one label: a change of speed does not cut it, nor a tilt that continues in the same direction; only a stop followed by a return, or a change of direction, cuts it. [instruction, BREITHORN, cross Displacement]
  • A return that exceeds the starting level remains the same event: it does not start a new movement. [LENZSPITZE]
  • A lasting jump: the label bounds the jump, the return is another label: the step is stated in the description; a step that returns within the day remains one label (a jump then the return); two steps separated by days of plateau make two labels, even in the same direction. [cross Displacement]
  • A bump that is only the trace of a slow camera movement follows the cutting of that movement: it is not a separate event. [cross Displacement]
  • Two gaps separated by a single measurement that came back make a single No data label; separated by two measurements or more, two: a single measurement does not make a resumption. [cross No data, SNCF01]
  • A measurement where a target is rejected alone just after a passage is the same event; a clean measurement between the two makes two labels: the rejection that follows is the end of the passage. [cross Obstruction]
  • Two nuisance episodes separated by a few hours where targets stay shifted make only one: the shifted targets are still touched. [NORDEND]
  • An obstacle that stays makes one label per parking or per episode: an intermittent episode (a plant shaken by the wind) is cut at a pause of several measurements. [cross Obstruction]
  • A Target issue: one label per target and per batch, from the start to the end of its regime: the defect is the target's, the night is not an event. [cross Target issue, JUNGFRAU]
  • The isolated rejections of a target before its Target issue regime each have their label, in the family of their cause: the Target issue starts with the regime, not at the first rejection separated from the regime by clean days. [finishing divers]
  • A target whose error_value goes above its threshold at a jump and stays there has its Target issue, in addition to the Displacement of the jump: the two labels overlap. [finishing niederhorn]
  • A deviation specific to targets, mixed with a camera movement, is the same event if it follows it in both directions, on the same dates: it rises, stops and comes back with it; with its own dates (start, stop or return without the camera), it is a second event; it is read on the residual (overview.py --rigid), day by day, at equal tilt or temperature. [finishing divers]

4. The family

  • A single family per label, its key in families.md, following its boundaries: Visibility, Obstruction, Thermal displacement, Displacement, Target issue, No data; no other key, nor old names. [instruction, MICHIGAN, WILDSTRUBEL]
  • We decide by the signature, not by coincidence: a temperature drop at the same moment does not make a thermal, an error_value that moves a little does not make a Visibility. [cross Thermal displacement, cross Visibility]
  • A same target keeps the same family from one batch to the next, and a same phenomenon on neighboring stations the same family: otherwise the defect of a target or a site reads as two causes. [NORDEND, consistency]
  • The sun at sunrise in the field and the lit background behind a spot are the same nuisance: same family, same way of stating the cause. [cross Visibility]

5. The description

  • As the expert would write it, in English: what is seen, the cause and how it was known, the conclusion for the client: which targets or references with their exact names, which magnitudes, the return or not; then nothing to do, data not to be believed, or movement to monitor. [instruction]

  • Qualitative, without numeric values, even in words: "a tenth", "half a pixel", "a few hundredths", "up to almost 1" are numbers; say "a little", "clearly", "strongly". [instruction, FLORIDA, JUNGFRAU]

  • The targets named are those that move the most, read instant by instant, in the right column: not the usual list; a rejected target does not "jump"; the direction is stated as it is that day. [instruction, CLARIDEN, NEBRASKA, FLORIDA, NIEDERHORN]

  • A sentence taken from a model is reread on the range of each label: the hours, the weather, the targets, the image, the return, the overshoot; a precision true for a single label is not copied. [cross Visibility, cross Thermal displacement, SNCF01, LENZSPITZE, NORDEND, FLORIDA]

  • "Nothing on the images", "image without obstacle" are written only after looking at the image of the hour: a passage is often seen in transparency; the image cited is that of the day when it exists. [instruction, CLARIDEN, MICHIGAN, FLORIDA, WILDSTRUBEL, cross Visibility, JUNGFRAU]

  • "No image" is written only at an instant when the camera usually takes an image, and after searching for the file: a station that takes one only on the hour proves nothing at the half hour, and the hourly images often fall on the minute of the measurements. [cross No data, LENZSPITZE]

  • A fact given by the expert is written "according to the expert", in words, without his numbers: not "the map knows". [instruction, BREITHORN, MICHIGAN, NIEDERHORN]

  • The temperature is "the housing temperature" (of the camera): the column being either sensor_*_deg_c, or, failing that, temperature_deg_c; never of the air, the building, the plant, the station, the components, the camera alone, nor the name of the column; without any metric, "no temperature is available". [instruction, cross Thermal displacement, LENZSPITZE, NORDEND, JUNGFRAU, consistency]

  • The weather is not drawn from the housing temperature: "cold and overcast day" is seen on the images. [FLORIDA]

  • The moments of the day (morning, late afternoon, night) are judged on the images, never on the UTC hour: the station can be anywhere. Every hour written is in UTC, no numeric local hour. [cross Visibility, label map, expert's reading rule]

  • A level "back" or "a little below" is compared over several days: a single day after may be in the middle of a thermal cycle; a night level that changes from one night to another is compared with several nights; a target rejected at night, with the nearest neighboring night where it is measured, never across a step. [instruction, MICHIGAN, cross Visibility, WILDSTRUBEL, finishing niederhorn]

  • A thermal cycle label is described on its own day, never by copying the day before: the direction and magnitude, in words, of each axis of the pose and the hour of its return; the rise of the housing temperature; the targets that dip and those that rise that day, without those that a Target issue makes unreadable; the corr_y; the hour when the curves come back; "in the sun", "the references stay calm", "some days" only if that day shows it. [finishing niederhorn, finishing nordend, finishing wildstrubel, NORDEND]

  • "Mostly pan", an axis that moves "with" another, a pose that "comes back at once" are checked axis by axis: in degrees, at the hour of each extreme. [finishing niederhorn, finishing wildstrubel]

  • The agitation from one measurement to the next is written in the thermal label: when the pose step exceeds several times that of the night and the error_value does not rise; it does not make a Visibility. [finishing niederhorn]

  • A compensated thermal movement is stated as such: the description says the pose moves and the compensation absorbs it. [cross Thermal displacement]

  • A touched reference is stated, with what remains: the targets move together; two remaining references, "a very bad sign"; none, "the points make no sense". [cross Obstruction]

  • A Displacement says what moves in the image, checked target by target on the raw position, and concludes accordingly: the camera moves and the targets do not: "it is not a movement of the structure; the camera is to be monitored"; targets move alone: "a movement of these targets is not excluded: to check (what to look at), monitor"; nothing moves in the image, the base or the configuration changes: "it is not a movement of the structure; points to ignore if there is a return, to check with the operator otherwise"; the whole scene deforms while the camera moves: "movement of the structure not established: to monitor". [cross Displacement, NEBRASKA]

  • A drift of several days says whether it follows the housing temperature and whether it comes back with it, at equal temperature: otherwise, "the temperature does not explain it alone, link not established". [cross Displacement]

  • A lasting jump states its step: its duration, its return or not. [cross Displacement]

  • Each label of a sequence (outward, return, resumption) names the other: "return of the drift of…, labeled apart", "without return until the end of the batch". [cross Displacement]

  • A Target issue says when the rejections stop before the end of the label: the end is read on the curves, the rejections may stop before. [cross Target issue]

  • A No data states the metrics and the image of the instant: present or not, checked, for each gap. [NIEDERHORN, cross No data]

  • On a recent camera (image quality filled, tools/scripts/quality.py), an occultation is written only after reading the flux of the instant: a cut flux proves it, a full flux denies it; the hidden targets are named from their flux, each time, and a hidden target that stayed under its threshold is stated (its point stays, wrong). The metrics: tools/DATA.md. [LENZSPITZE quality review, D-129]

  • "Bigger spots" is written only after comparison with a calm day of the same image regime and reading of the fwhm: a spot that has lost its light can look swollen. [LENZSPITZE quality review, D-129]

  • A jump without rejection is in no list of rejections: on a recent camera, quality.py --labels --outside gives the degraded instants outside the labels; corr is checked at each. [LENZSPITZE quality review, D-129]

6. The stated cause

  • A cause is stated in three ways only: seen, deduced, not established: "seen on the image (or the thumbnail) of HH:MM", saying what is seen there; "deduced from…" (the signature, the other passages seen, the curves and the camera pose), saying why it is not seen; "not established", with its hypotheses; or "according to the expert". [instruction, cross Obstruction, cross Displacement]
  • A cause deduced from a known signature is stated "deduced", not "not established" followed by hypotheses: the signature is a proof. [cross Visibility]
  • "Image not looked at" is not an end when an image exists: it is looked at, or the cause is stated deduced; a signature that returns about twenty times is looked at on a thumbnail sheet, one measurement per episode. [cross Obstruction, NIEDERHORN]
  • Do not say more than what was seen: never "nothing to fear" nor "probably not a movement" when the cause is not established: they reassure without proof; with an established cause (seen, or given by the expert), "nothing to fear for the structure" is right. A nuance on a cause not established ("probably", "likely", "possible") remains allowed, next to "not established" or "deduced": it states the degree of the hypothesis, as with the expert. [instruction, cross Displacement, WILDSTRUBEL; the user, 2026-11-05]
  • A gap at the same minutes on several stations is stated the same on each: "a common cause outside the station (transmission or reception of the data), not a station failure; it is not established", with the other stations, then "nothing to do on the station"; with two stations only, the common cause is "possible"; the word "outage" is not written. [cross No data]
  • Each rejection of a measurement of a small group of targets, each loss of a reference, is looked at on the thumbnail, even in a day of turbulence: it is often a passage. [WILDSTRUBEL]
  • A gap specific to a station keeps its own cause: energy, mains or silent station, read on its metrics; without a signature, "cause not established". [cross No data]

The boundaries between families, and what is not labeled: families.md.