- Every station's auto labels reviewed by an analyst (/review-auto-labels), then a
cross-review per family across all stations (new skill /cross-review-labels),
a finishing pass, and descriptions rewritten in words (no numeric values).
- The three never-labeled ranges (WILDSTRUBEL 07/09-01/10, NORDEND 16/08-01/10)
labeled and reviewed: 991 auto labels on 11 stations, labeling/controle.py at 0 errors.
- Labeling knowledge gathered in labeling/: regles-d-etiquetage.md (new), familles.md
(settled boundaries), connaissances/ (notes moved from docs/, links updated),
the label map rebuilt from reviewed labels, events and cases; events tied to their
stations (PS_DES_COUDRES is JUNGFRAU). The labeler and reviewers read only labeling/.
- labels-map reads auto labels and events; decision D-130; living plan points 73-75.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Definitions of peak, flux, sharpness, fwhm, dominance and snr in docs/domaine;
agent-tools/qualite.py (ratio to a calm reference, per label, outside labels,
figures), inventaire/graphique/resume aware of the six columns, lire-les-donnees
and the labeler's instructions updated. LENZSPITZE 16.10-01.11 relabeled: 67
labels, 4 found by the metrics.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
One analyst checks neighbouring stations together (.claude/skills/review-auto-labels/coherence.md). The month's drift, shared by the three cameras (same step on 08/10, same recoil 13-16/10), is Displacement everywhere, cause not established (the building or its temperature); the 09/10 reference losses and the 15-16/10 gap are tied across stations, a common cause to look for in the building; the 26/10 cold morning added on BREITHORN and NEBRASKA; MICHIGAN 27/10 as Displacement. Proposed per-station files versioned.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Bounds checked target by target on both columns; a slowing drift stays one event; a phenomenon found once is searched over the whole batch; "the camera, not the targets" checked with ensemble.py --rigide; housing temperature, "d'après l'expert" in words. familles.md: a small error_value rise at the housing temperature trough, nothing on the images, is Thermal displacement, not Visibility. Map lessons left to the next /labels-map.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
All 26 BREITHORN October labels carry "revue": "2026-11-04"; the expert's thermal answer is said in the description. /review-auto-labels skips labels that already carry "revue" (count, analyst review, replacement); /labels-map reads all reviewed auto labels (D-128).
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
The analyst reviews labeling/auto/ labels on a range chosen by the user (bounds, one event, family, description, missed ones) and proposes corrected labels; the expert answers its questions, then the range is replaced. First run on BREITHORN October: 10 corrected, 25 and 26 merged, one missed night added; the expert confirms the camera's cold-night deviations are thermal (housing temperature, not air). Labels the expert settled carry "revue", read by /labels-map.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
30 labels written by an agent from E-001 to E-011, decoded with the anonymization registry and appended to labeling/<STATION>.jsonl, so the labeling tool shows them with the others; no source field. The tool accepts only data/ stations again.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
The labeling tool now accepts event folders (Add, Save), and a Save of a label's text drops its "revue" mark.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
docs/displacement/domaine/chaine-de-mesure.md: how a measurement is made and where each cause enters the chain (view, obstacle, target, lens, camera support, references, structure), the three thermal effects, the displacement reading rules; sources marked, Claude's hypotheses flagged. Draft, not yet read by the analyst. Plan: pipelines status at point 57, draft noted at point 67.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
- inventaire.py: what each source contains on the window (filled or empty columns, metric names, images), names and counts only
- analyser-lot: step 0 on the data, own scripts in <lot>/outils/, section 3 "Outils"
- tool-builder (the optimiser, rewritten): after the analyst, optimises or adds reading tools, never removes; no control yet
- /aiguiser skill, reader, corrector and voir.py removed; its sessions and spec stay as history
- docs: D-097, specs, CLAUDE.md, run.md, README, plan point 56; the shared files also carry another session's pending edits (D-096 entry, plan points 57 to 67)
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
- /boucle becomes /events-learner: skill and scripts in
.claude/skills/events-learner/, sessions in events-learner/, spec
vues-deplacements/contexte/events-learner.md, end page
out/events-learner_<session>.html, commits "Learn <STATION> ...";
history keeps /boucle, the review page reads both.
- No benchmark-fin.txt: the end benchmark is the last kept
modification's, on the same events as the start.
- The spec keeps only the why; the how lives in the skill and agents.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
- event-builder: called at every batch, reads the whole analysis and
decides on its own, against the event base, what is new; an event
keeps the bounds of the event, not of the batch (plus 6 h).
- the "Nouveautés" section leaves the analysis and the upgrader's
output; the summary lists contradictions with the base.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
- analyste: the most precise analysis of the batch, detectors allowed as
a tool; reading, labels, known events, novelties in analyse.md; skill
critiquer-lot renamed analyser-lot.
- detectors-critic: compares the detectors to the analysis event by
event, without looking at the data; verdict and at most two
suggestions in critique.md, read by the detectors-upgrader.
- /boucle step 2, builder, review page, /new-event and score.py import
follow.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
- detectors-upgrader: makes the detectors better on the batch from the
analyst's critique; adds, removes or modifies, removing first; targets
the significant events missed or badly bounded, not a perfect match.
- detectors-control: the guard against overfitting; never looks at the
batch, judges the diff and the benchmark on the event base, decides,
and sends its remarks back to the upgrader.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
The event base is the analyst's knowledge base, read before or during
its reading; a known event is a lead to check on the data, not a
conclusion. No measurement of the analyst's progress.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
- Catch-up: E-004, E-006, E-007, E-008, E-010, E-011 in the map; new
section 5 "Faux amis"; rejected E-009 no longer cited.
- /review-event and /new-event: what the event teaches goes into the
map at validation, shown to the expert with the fiche.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
- read-event: map and index first, then the card of close events only,
then their data with the reading tools; never all cards at once.
Diagnostic agent, critiquer-lot and builder point to it; the index
header too. Signature length: about 250 characters, a guideline.
- "Cas" becomes "event": events/E-xxx/event.md, skills new-event,
review-event, anonymize-event.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
The diagnostic agent is kept in case it is needed again, under a name that
says what it is; entry point still .claude/agents/diagnostic.md; paths updated
in skills, agents, spec, contexts, the living plan and CLAUDE.md; reports and
history keep old paths. Also: two stale tool paths in the living plan.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
The case base leaves agents/diagnostic/cas/ for a root folder of its own,
with the event map and its graph; paths updated in skills, agents, scripts,
case fiches, contexts, the living plan and CLAUDE.md; history keeps old paths.
Also: plan point 51, an analyst that improves from session to session.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
It zooms on the event, shows its details and its timelapse under the figure,
and keeps only its sensors on the curves (no second click). Bar dates are
given without the trailing Z, which Plotly does not parse.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
The GIF is the timelapse as shown (focused or global, sensors circled and
labelled), 1000 px wide at most, red border during the event, time at the
bottom. In file:// the browser cannot read the images' pixels, so the
detectors, end-of-session and case-review pages now open through a local
server (open_served, 127.0.0.1:8765, started in the background, serving
out/ and data/ only).
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>