English: README, CLAUDE.md, archive note; sessions/ (ex seances); the session of the 2026-11-16 trial (NORDEND, S1 kept)

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
This commit is contained in:
maximeandClaude Opus 5.5 committed 2026-11-16 12:40:03 +01:00
1 parent 2cb0c88a2b
commit 63bebbc415
19 files changed
+689 -39

No files matched your search

+2 -2
View File
@@ -15,11 +15,11 @@ Given by the caller, in the request; any other context or rule it gives (files t
- `DETECTORS`: the path of the detectors repository (default: `detectors/`, a link in Optimizer). - `DETECTORS`: the path of the detectors repository (default: `detectors/`, a link in Optimizer).
- `STATION`: the station's data folder (read by `DETECTORS/scripts/data.py`); or, several stations on the same range: `DATA`, the folder of the station folders, and `STATIONS`, `A,B`, the stations to improve. - `STATION`: the station's data folder (read by `DETECTORS/scripts/data.py`); or, several stations on the same range: `DATA`, the folder of the station folders, and `STATIONS`, `A,B`, the stations to improve.
- `FROM`, `TO`: the range, UTC days (YYYY-MM-DD, `TO` excluded). - `FROM`, `TO`: the range, UTC days (YYYY-MM-DD, `TO` excluded).
- `LABELS`: the station's labels file (JSONL: `id`, `debut`, `fin`, `categorie` = the family, `description`), the truth on the range; with `STATIONS`, a folder of `<STATION>.jsonl`, one per station. - `LABELS`: the station's labels file (JSONL: `id`, `start`, `end`, `family`, `description`), the truth on the range; with `STATIONS`, a folder of `<STATION>.jsonl`, one per station.
- `REFERENCE`: the reference truth for the non-regression, `--labels <folder of <STATION>.jsonl> --data <folder of <STATION>/>`. - `REFERENCE`: the reference truth for the non-regression, `--labels <folder of <STATION>.jsonl> --data <folder of <STATION>/>`.
- `FAMILIES`: the document defining the families and their boundaries (one detector per family). - `FAMILIES`: the document defining the families and their boundaries (one detector per family).
- `CONTEXT` (optional): documents explaining the range's labels (an analyst's report, for example). - `CONTEXT` (optional): documents explaining the range's labels (an analyst's report, for example).
- `OUT`: the output folder of this run (default: `seances/<analysis>_<YYYYMMDDTHHMMZ>/` in Optimizer). - `OUT`: the output folder of this run (default: `sessions/<analysis>_<YYYYMMDDTHHMMZ>/` in Optimizer).
A required input missing, or `DETECTORS` not a git repository: say so and stop, changing nothing. A required input missing, or `DETECTORS` not a git repository: say so and stop, changing nothing.
+14 -13
View File
@@ -1,17 +1,18 @@
# CLAUDE.md # CLAUDE.md
Optimizer : l'optimisation des détecteurs AIMsight sur les étiquettes de `labeling/auto`, et rien d'autre (`README.md`). Le workflow : `.claude/skills/improve/SKILL.md` (`/improve`), lancé depuis la racine d'Optimizer avec le plugin des agents (`--plugin-dir agents`) ; ses agents `agents:upgrader` et `agents:validator` ; le benchmark `detectors/bench/`. Optimizer: the optimisation of the AIMsight detectors on the labels of `labeling/auto`, and nothing else (`README.md`). The workflow: `.claude/skills/improve/SKILL.md` (`/improve`), launched from the root of Optimizer with the agents plugin (`--plugin-dir agents`); its agents `agents:upgrader` and `agents:validator`; the benchmark `detectors/bench/`.
- **Les liens**, hors de git : `detectors/` → `~/detectors` (la séance y commite) ; `data/` → `~/aimsight-data/data` et `labeling/` → `~/labeling` (lecture seule) ; `agents/` → `~/agents` (le plugin, chargé en lecture). - **The links**, out of git: `detectors/` → `~/detectors` (the session commits there); `data/` → `~/aimsight-data/data` and `labeling/` → `~/labeling` (read-only); `agents/` → `~/agents` (the plugin, loaded read-only).
- **Les séances** : `seances/<analyse>_<AAAAMMJJTHHMMZ>/`, non commitées par la séance ; Diagnostic-App y lit le bilan (`summary.md`). - **The sessions**: `sessions/<analysis>_<YYYYMMDDTHHMMZ>/`; Diagnostic-App reads their summary (`summary.md`).
- **Le workflow reste générique** : le skill ne décrit que le workflow (plage + étiquettes → upgrader → validator → bilan) ; ce qui est propre à un appelant (ses chemins, ses règles git, `Caller's context`) arrive dans la requête. - **The workflow stays generic**: the skill describes only the workflow (range + labels → upgrader → validator → summary); what is specific to a caller (its paths, its git rules, `Caller's context`) arrives in the request.
- **Vérifier un changement du skill** : `env -u ANTHROPIC_API_KEY claude -p "/improve" --plugin-dir agents --model haiku` doit dire qu'il manque les entrées et s'arrêter sans toucher à `detectors/` ; puis une séance réelle par le bouton de l'app. - **Check a change of the skill**: `env -u ANTHROPIC_API_KEY claude -p "/improve" --plugin-dir agents --model haiku` must say the inputs are missing and stop without touching `detectors/`; then a real session through the app's button.
- `archive/` : l'ancien batch et son scanner, arrêtés ; ne pas les relancer. - `archive/`: the former batch and its scanner, stopped; do not run them again.
Règles : Rules:
- **Les agents tournent sur l'abonnement Claude, jamais sur une clé d'API** (`ANTHROPIC_API_KEY` retiré du lancement). - **The agents run on the Claude subscription, never on an API key** (`ANTHROPIC_API_KEY` removed from the launch).
- Toute heure est en UTC. - Every time is UTC.
- Lecture seule sur `data/` et `labeling/` ; ne jamais recopier les identifiants des métriques (`SIM_CCID`, `ip*`) ; ne jamais lire `.env`. - Read-only on `data/` and `labeling/`; never copy the metric identifiers (`SIM_CCID`, `ip*`); never read `.env`.
- Pas de règle propre à une station, une cible ou une date. - The project uses the labels of `labeling/auto` only; `labeling/manual` is the user's alone.
- Commits par chemin, jamais `git add -A`. - No rule specific to a station, a target or a date.
- Langue : français ; le skill `improve`, comme le dépôt detectors et ses agents, reste en anglais. - Commit by path, never `git add -A`.
- Language: English.
+21 -20
View File
@@ -1,41 +1,42 @@
# Optimizer # Optimizer
L'optimisation des détecteurs AIMsight, sur les étiquettes de `labeling/auto`, et rien d'autre. Sur la plage d'une analyse approfondie (faite dans l'app Diagnostic-App), ses étiquettes pour vérité, le skill `improve` : The optimisation of the AIMsight detectors, on the labels of `labeling/auto`, and nothing else. On the range of an in-depth analysis (made in the Diagnostic-App app), its labels as the truth, the `improve` skill:
1. mesure le score des détecteurs sur la plage, et le benchmark sur toutes les étiquettes de `labeling/auto` ; 1. measures the detectors' score on the range, and the benchmark on every label of `labeling/auto`;
2. fait corriger les détecteurs par l'`upgrader`, une correction par commit dans `~/detectors` ; 2. has the detectors fixed by the `upgrader`, one fix per commit in `~/detectors`;
3. les fait juger par le `validator`, qui garde, fait retirer ou annule : aucune station ne recule, pas de sur-apprentissage ; 3. has them judged by the `validator`, which keeps, has some removed, or cancels: no station goes back, no overfitting;
4. écrit le bilan de la séance. 4. writes the summary of the session.
L'`upgrader` et le `validator` viennent du dépôt `agents` ; le benchmark (`bench/`), du dépôt `detectors`. Rien n'est poussé. The `upgrader` and the `validator` come from the `agents` repository; the benchmark (`bench/`) from the `detectors` repository. Nothing is pushed.
## Installer ## Install
```bash ```bash
ln -s ~/detectors detectors # les détecteurs, partagés avec Diagnostic-App ln -s ~/detectors detectors # the detectors, shared with Diagnostic-App
ln -s ~/aimsight-data/data data # les données pompées, hors de git ln -s ~/aimsight-data/data data # the pumped data, out of git
ln -s ~/labeling labeling # les étiquettes : la vérité et les familles ln -s ~/labeling labeling # the labels: the truth and the families
ln -s ~/agents agents # les agents génériques (plugin Claude Code) ln -s ~/agents agents # the generic agents (Claude Code plugin)
``` ```
`claude` (Claude Code) doit être sur le `PATH`, connecté à l'abonnement Claude (`/login`), jamais à une clé d'API. `claude` (Claude Code) must be on the `PATH`, logged in to the Claude subscription (`/login`), never to an API key.
## Lancer ## Run
Par le bouton « Optimiser les détecteurs » de Diagnostic-App, sur une analyse finie ; ou à la main, depuis la racine d'Optimizer : From the "Optimize the detectors" button of Diagnostic-App, on a finished analysis; or by hand, from the root of Optimizer:
```bash ```bash
env -u ANTHROPIC_API_KEY claude -p "/improve DETECTORS: detectors/. STATION: data/NORDEND. FROM: 2026-11-04. TO: 2026-11-08. \ env -u ANTHROPIC_API_KEY claude -p "/improve DETECTORS: detectors/. STATION: data/NORDEND. FROM: 2026-11-04. TO: 2026-11-08. \
LABELS: labeling/auto/NORDEND.jsonl. REFERENCE: --labels labeling/auto --data data. FAMILIES: labeling/familles.md. \ LABELS: labeling/auto/NORDEND.jsonl. REFERENCE: --labels labeling/auto --data data. FAMILIES: labeling/families.md. \
CONTEXT: labeling/auto/analyses/<analyse>/rapport.md. OUT: seances/<analyse>_<AAAAMMJJTHHMMZ>/." \ CONTEXT: labeling/auto/analyses/<analysis>/report.md. OUT: sessions/<analysis>_<YYYYMMDDTHHMMZ>/." \
--plugin-dir agents --add-dir ~/detectors --add-dir ~/labeling --add-dir ~/aimsight-data/data \ --plugin-dir agents --add-dir ~/detectors --add-dir ~/labeling --add-dir ~/aimsight-data/data \
--setting-sources project,local --strict-mcp-config \
--allowedTools Read,Write,Edit,Bash,Agent --permission-mode acceptEdits --allowedTools Read,Write,Edit,Bash,Agent --permission-mode acceptEdits
``` ```
`TO` est exclu. Plusieurs stations sur la même plage : `DATA: data. STATIONS: A,B. LABELS: <dossier de <STATION>.jsonl>` à la place de `STATION` et d'un fichier d'étiquettes. `TO` is excluded. Several stations on the same range: `DATA: data. STATIONS: A,B. LABELS: <folder of <STATION>.jsonl>` instead of `STATION` and a labels file.
## Ce qu'une séance laisse ## What a session leaves
`seances/<analyse>_<heure>/` (non commité par la séance) : les mesures avant et après (`range-*.txt`, `benchmark-*.txt`), `modification.md` (l'upgrader), `validation.md` (le validator), `summary.md` (le bilan, que l'app affiche sous le rapport de l'analyse), `tools/` (les scripts que les agents ont écrits). Les corrections : des commits `improve: S<n>` dans `~/detectors`, sur sa branche courante. `sessions/<analysis>_<time>/`: the measures before and after (`range-*.txt`, `benchmark-*.txt`), `modification.md` (the upgrader), `validation.md` (the validator), `summary.md` (the summary, shown by the app under the analysis report), `tools/` (the scripts the agents wrote). The fixes: `improve: S<n>` commits in `~/detectors`, on its current branch.
`archive/` : l'ancien batch quotidien et son scanner (l'auto-étiquetage par les détecteurs), arrêtés le 2026-11-16. L'architecture des dépôts : `~/Diagnostic-App/docs/architecture.drawio`. `archive/`: the former daily batch and its scanner (auto-labeling by the detectors), stopped on 2026-11-16. The architecture of the repositories: `~/Diagnostic-App/docs/architecture.drawio`.
+4 -4
View File
@@ -1,7 +1,7 @@
# Archive : l'auto-étiquetage par les détecteurs (arrêté) # Archive: auto-labeling by the detectors (stopped)
Ce qu'Optimizer faisait du 2026-11-13 au 2026-11-16 : un batch par jour UTC (`batch.py`), toutes les stations ; les détecteurs posaient leurs étiquettes, le `scanner` (`scanner.md`, Sonnet) les corrigeait de loin face aux courbes, et `wf-improve` apprenait les détecteurs sur ces corrections. Les batches, les outils de l'audit et les journaux d'essai sont ici tels qu'ils étaient (`batches/`, `outils/`, `docs/`). What Optimizer did from 2026-11-13 to 2026-11-16: one batch per UTC day (`batch.py`), every station; the detectors wrote their labels, the `scanner` (`scanner.md`, Sonnet) corrected them from afar against the curves, and `wf-improve` trained the detectors on these corrections. The batches, the audit tools and the trial logs are here as they were (`batches/`, `outils/`, `docs/`), in French.
Arrêté le 2026-11-16 par l'utilisateur : beaucoup d'étiquettes sans la qualité, et un effet circulaire, les détecteurs appris sur leurs propres étiquettes à peine corrigées. Les détecteurs s'améliorent désormais depuis les analyses approfondies de l'app (le skill `improve`, à la racine du dépôt). Stopped on 2026-11-16 by the user: many labels without the quality, and a circular effect, the detectors trained on their own labels barely corrected. The detectors are now improved from the in-depth analyses of the app (the `improve` skill, at the root of the repository).
Gardé pour trace, non maintenu : les chemins qu'il cite (le plugin `/detectors:wf-improve`, `skills/wf-improve/scripts/`) n'existent plus. Kept as a record, not maintained: the paths it quotes (the `/detectors:wf-improve` plugin, `skills/wf-improve/scripts/`) no longer exist.
@@ -0,0 +1,101 @@
BREITHORN: 2026-10-03T00:00Z → 2026-11-01T00:00Z, 27 labels
found correct
Displacement 168/168 (100 %) 168/171 (98 %)
No data 2/2 (100 %) 2/2 (100 %)
Obstruction 46/48 (96 %) 46/46 (100 %)
Thermal displacement 4/21 (19 %) 4/4 (100 %)
Visibility — 0/25 (0 %)
CLARIDEN: 2026-10-01T00:00Z → 2026-11-01T00:00Z, 43 labels
found correct
Displacement 2578/3091 (83 %) 2578/2601 (99 %)
No data 6/20 (30 %) 6/12 (50 %)
Obstruction 1/5 (20 %) 1/5 (20 %)
Target issue 4068/4068 (100 %) 4068/4069 (100 %)
Thermal displacement 40/42 (95 %) 40/264 (15 %)
Visibility 3/87 (3 %) 3/3 (100 %)
FLORIDA: 2026-10-01T00:00Z → 2026-11-01T00:00Z, 64 labels
found correct
Displacement 0/41 (0 %) 0/14 (0 %)
No data 4/21 (19 %) 4/4 (100 %)
Obstruction 0/2 (0 %) 0/13 (0 %)
Target issue 2913/2913 (100 %) 2913/2914 (100 %)
Thermal displacement 815/1047 (78 %) 815/1338 (61 %)
Visibility 434/598 (73 %) 434/439 (99 %)
JUNGFRAU: 2026-09-30T00:00Z → 2026-11-01T00:00Z, 99 labels
found correct
Displacement — 0/19 (0 %)
No data 4/29 (14 %) 4/4 (100 %)
Obstruction 1/1 (100 %) 1/158 (1 %)
Target issue 3005/3005 (100 %) 3005/3006 (100 %)
Thermal displacement 2300/2557 (90 %) 2300/2654 (87 %)
Visibility 427/669 (64 %) 427/450 (95 %)
LENZSPITZE: 2026-10-16T00:00Z → 2026-11-01T00:00Z, 67 labels
found correct
Displacement — 0/4640 (0 %)
No data 12/615 (2 %) 12/16 (75 %)
Obstruction 19/47 (40 %) 19/84 (23 %)
Thermal displacement 1767/1925 (92 %) 1767/4253 (42 %)
Visibility 934/1886 (50 %) 934/1116 (84 %)
MICHIGAN: 2026-10-03T00:00Z → 2026-11-01T00:00Z, 15 labels
found correct
Displacement 130/130 (100 %) 130/171 (76 %)
No data 2/2 (100 %) 2/2 (100 %)
Target issue 132/147 (90 %) 132/132 (100 %)
Thermal displacement 19/25 (76 %) 19/46 (41 %)
NEBRASKA: 2026-10-03T00:00Z → 2026-11-01T00:00Z, 11 labels
found correct
Displacement 165/165 (100 %) 165/171 (96 %)
No data 2/2 (100 %) 2/2 (100 %)
Target issue 132/132 (100 %) 132/132 (100 %)
Thermal displacement 10/20 (50 %) 10/17 (59 %)
NIEDERHORN: 2026-10-01T00:00Z → 2026-11-01T00:00Z, 139 labels
found correct
Displacement 3/48 (6 %) 3/10 (30 %)
No data 2/26 (8 %) 2/2 (100 %)
Obstruction 66/156 (42 %) 66/134 (49 %)
Target issue 719/719 (100 %) 719/720 (100 %)
Thermal displacement 216/267 (81 %) 216/236 (92 %)
Visibility 28/104 (27 %) 28/32 (88 %)
NORDEND: 2026-08-10T00:00Z → 2026-11-08T00:00Z, 243 labels
found correct
Displacement 13/13 (100 %) 13/8549 (0 %)
No data 10/58 (17 %) 10/10 (100 %)
Obstruction 638/2754 (23 %) 638/2365 (27 %)
Target issue 3911/7441 (53 %) 3911/4290 (91 %)
Thermal displacement 5066/7667 (66 %) 5066/5411 (94 %)
Visibility 655/1950 (34 %) 655/1428 (46 %)
SNCF01: 2026-10-04T00:00Z → 2026-11-01T00:00Z, 97 labels
found correct
Displacement 2111/2111 (100 %) 2111/2112 (100 %)
No data 10/30 (33 %) 10/16 (62 %)
Obstruction 20/22 (91 %) 20/431 (5 %)
Target issue 2111/2111 (100 %) 2111/2112 (100 %)
Thermal displacement 1254/1726 (73 %) 1254/1334 (94 %)
Visibility 23/761 (3 %) 23/24 (96 %)
WILDSTRUBEL: 2026-09-01T00:00Z → 2026-11-01T00:00Z, 201 labels
found correct
Displacement 579/579 (100 %) 579/1986 (29 %)
No data 21/31 (68 %) 21/30 (70 %)
Obstruction 253/1026 (25 %) 253/482 (52 %)
Target issue 2089/2091 (100 %) 2089/2092 (100 %)
Thermal displacement 495/845 (59 %) 495/512 (97 %)
Visibility 908/1152 (79 %) 908/925 (98 %)
Total
found correct
Displacement 5747/6346 (91 %) 5747/20444 (28 %)
No data 75/836 (9 %) 75/100 (75 %)
Obstruction 1044/4061 (26 %) 1044/3718 (28 %)
Target issue 19080/22627 (84 %) 19080/19467 (98 %)
Thermal displacement 11986/16142 (74 %) 11986/16069 (75 %)
Visibility 3412/7207 (47 %) 3412/4442 (77 %)
@@ -0,0 +1,101 @@
BREITHORN: 2026-10-03T00:00Z → 2026-11-01T00:00Z, 27 labels
found correct
Displacement 168/168 (100 %) 168/171 (98 %)
No data 2/2 (100 %) 2/2 (100 %)
Obstruction 46/48 (96 %) 46/46 (100 %)
Thermal displacement 4/21 (19 %) 4/4 (100 %)
Visibility — 0/25 (0 %)
CLARIDEN: 2026-10-01T00:00Z → 2026-11-01T00:00Z, 43 labels
found correct
Displacement 2578/3091 (83 %) 2578/2601 (99 %)
No data 6/20 (30 %) 6/12 (50 %)
Obstruction 1/5 (20 %) 1/5 (20 %)
Target issue 4068/4068 (100 %) 4068/4069 (100 %)
Thermal displacement 40/42 (95 %) 40/264 (15 %)
Visibility 3/87 (3 %) 3/3 (100 %)
FLORIDA: 2026-10-01T00:00Z → 2026-11-01T00:00Z, 64 labels
found correct
Displacement 0/41 (0 %) 0/14 (0 %)
No data 4/21 (19 %) 4/4 (100 %)
Obstruction 0/2 (0 %) 0/13 (0 %)
Target issue 2913/2913 (100 %) 2913/2914 (100 %)
Thermal displacement 815/1047 (78 %) 815/1338 (61 %)
Visibility 434/598 (73 %) 434/439 (99 %)
JUNGFRAU: 2026-09-30T00:00Z → 2026-11-01T00:00Z, 99 labels
found correct
Displacement — 0/19 (0 %)
No data 4/29 (14 %) 4/4 (100 %)
Obstruction 1/1 (100 %) 1/158 (1 %)
Target issue 3005/3005 (100 %) 3005/3006 (100 %)
Thermal displacement 2300/2557 (90 %) 2300/2654 (87 %)
Visibility 427/669 (64 %) 427/450 (95 %)
LENZSPITZE: 2026-10-16T00:00Z → 2026-11-01T00:00Z, 67 labels
found correct
Displacement — 0/4640 (0 %)
No data 12/615 (2 %) 12/16 (75 %)
Obstruction 19/47 (40 %) 19/84 (23 %)
Thermal displacement 1767/1925 (92 %) 1767/4253 (42 %)
Visibility 934/1886 (50 %) 934/1116 (84 %)
MICHIGAN: 2026-10-03T00:00Z → 2026-11-01T00:00Z, 15 labels
found correct
Displacement 130/130 (100 %) 130/171 (76 %)
No data 2/2 (100 %) 2/2 (100 %)
Target issue 132/147 (90 %) 132/132 (100 %)
Thermal displacement 19/25 (76 %) 19/46 (41 %)
NEBRASKA: 2026-10-03T00:00Z → 2026-11-01T00:00Z, 11 labels
found correct
Displacement 165/165 (100 %) 165/171 (96 %)
No data 2/2 (100 %) 2/2 (100 %)
Target issue 132/132 (100 %) 132/132 (100 %)
Thermal displacement 10/20 (50 %) 10/17 (59 %)
NIEDERHORN: 2026-10-01T00:00Z → 2026-11-01T00:00Z, 139 labels
found correct
Displacement 3/48 (6 %) 3/10 (30 %)
No data 2/26 (8 %) 2/2 (100 %)
Obstruction 66/156 (42 %) 66/134 (49 %)
Target issue 719/719 (100 %) 719/720 (100 %)
Thermal displacement 216/267 (81 %) 216/236 (92 %)
Visibility 28/104 (27 %) 28/32 (88 %)
NORDEND: 2026-08-10T00:00Z → 2026-11-08T00:00Z, 243 labels
found correct
Displacement 13/13 (100 %) 13/8549 (0 %)
No data 10/58 (17 %) 10/10 (100 %)
Obstruction 638/2754 (23 %) 638/2365 (27 %)
Target issue 3911/7441 (53 %) 3911/4290 (91 %)
Thermal displacement 5066/7667 (66 %) 5066/5411 (94 %)
Visibility 655/1950 (34 %) 655/1428 (46 %)
SNCF01: 2026-10-04T00:00Z → 2026-11-01T00:00Z, 97 labels
found correct
Displacement 2111/2111 (100 %) 2111/2112 (100 %)
No data 10/30 (33 %) 10/16 (62 %)
Obstruction 20/22 (91 %) 20/431 (5 %)
Target issue 2111/2111 (100 %) 2111/2112 (100 %)
Thermal displacement 1254/1726 (73 %) 1254/1334 (94 %)
Visibility 23/761 (3 %) 23/24 (96 %)
WILDSTRUBEL: 2026-09-01T00:00Z → 2026-11-01T00:00Z, 201 labels
found correct
Displacement 579/579 (100 %) 579/1986 (29 %)
No data 21/31 (68 %) 21/30 (70 %)
Obstruction 253/1026 (25 %) 253/482 (52 %)
Target issue 2089/2091 (100 %) 2089/2092 (100 %)
Thermal displacement 495/845 (59 %) 495/512 (97 %)
Visibility 908/1152 (79 %) 908/925 (98 %)
Total
found correct
Displacement 5747/6346 (91 %) 5747/20444 (28 %)
No data 75/836 (9 %) 75/100 (75 %)
Obstruction 1044/4061 (26 %) 1044/3718 (28 %)
Target issue 19080/22627 (84 %) 19080/19467 (98 %)
Thermal displacement 11986/16142 (74 %) 11986/16069 (75 %)
Visibility 3412/7207 (47 %) 3412/4442 (77 %)
@@ -0,0 +1,101 @@
BREITHORN: 2026-10-03T00:00Z → 2026-11-01T00:00Z, 27 labels
found correct
Displacement 168/168 (100 %) 168/171 (98 %)
No data 2/2 (100 %) 2/2 (100 %)
Obstruction 46/48 (96 %) 46/46 (100 %)
Thermal displacement 4/21 (19 %) 4/4 (100 %)
Visibility — 0/25 (0 %)
CLARIDEN: 2026-10-01T00:00Z → 2026-11-01T00:00Z, 43 labels
found correct
Displacement 2578/3091 (83 %) 2578/2601 (99 %)
No data 6/20 (30 %) 6/12 (50 %)
Obstruction 1/5 (20 %) 1/5 (20 %)
Target issue 4068/4068 (100 %) 4068/4069 (100 %)
Thermal displacement 40/42 (95 %) 40/264 (15 %)
Visibility 3/87 (3 %) 3/3 (100 %)
FLORIDA: 2026-10-01T00:00Z → 2026-11-01T00:00Z, 64 labels
found correct
Displacement 0/41 (0 %) 0/14 (0 %)
No data 4/21 (19 %) 4/4 (100 %)
Obstruction 0/2 (0 %) 0/13 (0 %)
Target issue 2913/2913 (100 %) 2913/2914 (100 %)
Thermal displacement 815/1047 (78 %) 815/1338 (61 %)
Visibility 434/598 (73 %) 434/439 (99 %)
JUNGFRAU: 2026-09-30T00:00Z → 2026-11-01T00:00Z, 99 labels
found correct
Displacement — 0/19 (0 %)
No data 4/29 (14 %) 4/4 (100 %)
Obstruction 1/1 (100 %) 1/158 (1 %)
Target issue 3005/3005 (100 %) 3005/3006 (100 %)
Thermal displacement 1544/2557 (60 %) 1544/1786 (86 %)
Visibility 427/669 (64 %) 427/450 (95 %)
LENZSPITZE: 2026-10-16T00:00Z → 2026-11-01T00:00Z, 67 labels
found correct
Displacement — 0/4640 (0 %)
No data 12/615 (2 %) 12/16 (75 %)
Obstruction 19/47 (40 %) 19/84 (23 %)
Thermal displacement 1767/1925 (92 %) 1767/4253 (42 %)
Visibility 934/1886 (50 %) 934/1116 (84 %)
MICHIGAN: 2026-10-03T00:00Z → 2026-11-01T00:00Z, 15 labels
found correct
Displacement 130/130 (100 %) 130/171 (76 %)
No data 2/2 (100 %) 2/2 (100 %)
Target issue 132/147 (90 %) 132/132 (100 %)
Thermal displacement 19/25 (76 %) 19/46 (41 %)
NEBRASKA: 2026-10-03T00:00Z → 2026-11-01T00:00Z, 11 labels
found correct
Displacement 165/165 (100 %) 165/171 (96 %)
No data 2/2 (100 %) 2/2 (100 %)
Target issue 132/132 (100 %) 132/132 (100 %)
Thermal displacement 10/20 (50 %) 10/17 (59 %)
NIEDERHORN: 2026-10-01T00:00Z → 2026-11-01T00:00Z, 139 labels
found correct
Displacement 3/48 (6 %) 3/10 (30 %)
No data 2/26 (8 %) 2/2 (100 %)
Obstruction 66/156 (42 %) 66/134 (49 %)
Target issue 719/719 (100 %) 719/720 (100 %)
Thermal displacement 216/267 (81 %) 216/236 (92 %)
Visibility 28/104 (27 %) 28/32 (88 %)
NORDEND: 2026-08-10T00:00Z → 2026-11-08T00:00Z, 243 labels
found correct
Displacement 13/13 (100 %) 13/8549 (0 %)
No data 10/58 (17 %) 10/10 (100 %)
Obstruction 638/2754 (23 %) 638/2365 (27 %)
Target issue 3911/7441 (53 %) 3911/4290 (91 %)
Thermal displacement 4769/7667 (62 %) 4769/5097 (94 %)
Visibility 655/1950 (34 %) 655/1428 (46 %)
SNCF01: 2026-10-04T00:00Z → 2026-11-01T00:00Z, 97 labels
found correct
Displacement 2111/2111 (100 %) 2111/2112 (100 %)
No data 10/30 (33 %) 10/16 (62 %)
Obstruction 20/22 (91 %) 20/431 (5 %)
Target issue 2111/2111 (100 %) 2111/2112 (100 %)
Thermal displacement 1218/1726 (71 %) 1218/1275 (96 %)
Visibility 23/761 (3 %) 23/24 (96 %)
WILDSTRUBEL: 2026-09-01T00:00Z → 2026-11-01T00:00Z, 201 labels
found correct
Displacement 579/579 (100 %) 579/1986 (29 %)
No data 21/31 (68 %) 21/30 (70 %)
Obstruction 253/1026 (25 %) 253/482 (52 %)
Target issue 2089/2091 (100 %) 2089/2092 (100 %)
Thermal displacement 495/845 (59 %) 495/512 (97 %)
Visibility 908/1152 (79 %) 908/925 (98 %)
Total
found correct
Displacement 5747/6346 (91 %) 5747/20444 (28 %)
No data 75/836 (9 %) 75/100 (75 %)
Obstruction 1044/4061 (26 %) 1044/3718 (28 %)
Target issue 19080/22627 (84 %) 19080/19467 (98 %)
Thermal displacement 10897/16142 (68 %) 10897/14828 (73 %)
Visibility 3412/7207 (47 %) 3412/4442 (77 %)
@@ -0,0 +1,81 @@
# Modification
**Verdict**: changed
## Diagnosis
NORDEND, 2026-11-05T00:00Z → 2026-11-06T00:00Z, cadence 15 min, 19 targets and 3 references (CTR02, CTR03, CTR04; CTR01 is a local reference read as a target). Labels touching the range: NORDEND-229, 230, 231, 232, 233, 234, 235, 236, 243.
| Family | labelled | found | flagged | wrongly |
|---|---|---|---|---|
| Displacement | 1 | 0 | 3 | 3 |
| Obstruction | 32 | 28 | 60 | 32 |
| Target issue | 96 | 0 | 0 | 0 |
| Thermal displacement | 95 | 53 | 53 | 0 |
| Visibility | 0 | 0 | 43 | 43 |
- **NORDEND-229 / NORDEND-234 (Thermal displacement, camera cycle, 00:00 → 24:00)**: wrong bounds, too short. The detectors only had the wave from 07:00 to 20:00 (Camera thermal expansion 07:00 → 14:00, contraction 14:00 → 20:00). The return stopped as soon as the hourly gap profile fell under `EDGE` (0.1 of the amplitude), around 20:00. But the profile keeps a slow tail all night, which the labels count as part of the cycle ("sans repos jusqu'au départ du matin suivant"). This is the main defect, and it comes back every night. Fixed in part by S1. The early night of the 05/11 (01:00 → 06:30, the end of the 04/11 cycle) is still missed.
- **NORDEND-230 (Target issue, CTR01 rejected most of the time since 16/10)**: wrong family on the range. With only the 3 days of history the range command reads, CTR01 cannot reach the `WEEK` (7 days) a Degraded target needs. It comes out as a Target obscured (Obstruction) of CTR01 from 11-04 07:15 to 11-05 14:45. The families document allows this for a target hindered less than a week. Over the benchmark window (the whole label range), the detectors find it as a Degraded target from 10-17 to 11-07. This is a limit of the range window, not of the detector. The same accidental Target obscured gives most of the "found" Obstruction instants of the range (28/32).
- **NORDEND-231 (00:45 → 01:30) and NORDEND-233 (03:15 → 03:45), Obstruction, CTR03 rejected so the base jumps**: wrong family (Visibility). The jumps back of 01:00 and 03:30 (7 and 17 targets) are absorbed into a Poor visibility "reduced" from 00:45 to 09:30. CTR03 is at 0.7 to 1.4 times its threshold, rejected 1 to 2 measurements. The families document puts a reference at its threshold, with nothing seen in front of it, in Visibility or Displacement. Only the thumbnail shows the cobweb thread: a limit (see Limits). Over the full benchmark window, 03:30 is already a Brief obstruction of CTR03.
- **Poor visibility 00:45 → 09:30 and 13:30 → 15:00 (43 instants without a Visibility label)**: extra. At no instant are more than 3 of the 22 sensors high: CTR02, PR01, PR09 go back and forth all day (the context says "sans effet sur les courbes"), and the side-effects of CTR03 appear on PR04 and PR11. The bounds come from CTR01, which is rejected almost all the time. Four general rules were tried; each one makes another station go back (see Not fixed).
- **NORDEND-232 (03:00 → 09:45) and NORDEND-235 (15:00), Obstruction, PR09**: correct family in part. PR09 has Brief obstructions at 07:15 to 07:45 and 09:30, and its other rejections are inside the Poor visibility above. Wrong bounds, from the same cause.
- **NORDEND-236 (23:00 → 11-06 10:00), Obstruction, CTR03 comb**: wrong family on the range. The first two CTR03 rejections (23:15, 23:45) come out as "Jump, not back yet" (Displacement, 19 targets), because the range ends at midnight and the return cannot be seen yet. Over the benchmark window they are Brief obstructions of CTR03, the label's family. This is an edge effect of the window.
- **NORDEND-243 (12:00, Displacement, camera jolt)**: absorbed into the camera thermal wave. NORDEND-234 itself describes the 12:00 jolt as part of the thermal cycle ("à 12:00 où toute la scène … fait un à-coup"). The labels disagree with each other here, and the detectors keep 234's reading: alternative reading.
## What changed
- **S1** (commit `3ea6b3e`), `scripts/thermal_displacement.py`: new constant `RETURN = 0.05`. The return of the wave (contraction) now goes on until the hourly gap profile falls under `RETURN` times the amplitude, after midnight if needed, at most to the next day's start. Before, it stopped at `EDGE` = 0.1. `EDGE` (0.1) now only bounds the start of the wave. Docstring (point 5) and `thresholds.md` (`EDGE` line reworded, `RETURN` line added) are updated.
## Why
- S1: a camera or structure heated during the day warms up quickly and cools slowly. The tail of its return stays a few percent of the amplitude away from the night level for hours, and labels count this tail as the thermal cycle. One threshold for both ends cut the contraction in the evening. Lowering `EDGE` itself also moved the start earlier and overflowed into the mornings, so only the return is read lower.
## Before / after on the range
Events command, changed lines:
```
before: 2026-11-05T14:00Z → 2026-11-05T20:00Z Thermal displacement kept | Camera thermal contraction | 22 sensors
after: 2026-11-04T14:00Z → 2026-11-05T01:00Z Thermal displacement kept | Camera thermal contraction | 22 sensors
2026-11-05T14:00Z → 2026-11-05T23:45Z Thermal displacement kept | Camera thermal contraction | 22 sensors
```
No `skipped` line. Range score: Thermal displacement found 53/95 (56 %) → 73/95 (77 %), correct 53/53 → 73/73 (100 %). The other families are unchanged (Displacement 0/1 and 0/3, Obstruction 28/32 and 28/60, Target issue 0/96, Visibility 0/43 correct).
## Benchmark
| Station | Family | found | correct |
|---|---|---|---|
| JUNGFRAU | Thermal displacement | 1544/2557 (60 %) → 2300/2557 (90 %) | 1544/1786 (86 %) → 2300/2654 (87 %) |
| NORDEND | Thermal displacement | 4769/7667 (62 %) → 5066/7667 (66 %) | 4769/5097 (94 %) → 5066/5411 (94 %) |
| SNCF01 | Thermal displacement | 1218/1726 (71 %) → 1254/1726 (73 %) | 1218/1275 (95.5 %) → 1254/1334 (94.0 %) |
`Total`:
| Family | found | correct |
|---|---|---|
| Thermal displacement | 10897/16142 (68 %) → 11986/16142 (74 %) | 10897/14828 (73 %) → 11986/16069 (75 %) |
No other line changed. SNCF01 "correct" drops 1.5 points, under the 2-point limit. The output is in `benchmark-S1.txt`.
## For the user to decide
- NORDEND-243 (NORDEND, 2026-11-05 12:00, Displacement) against NORDEND-234 (Thermal displacement), whose description includes the same 12:00 jolt. The detectors absorb the one-measurement jump into the camera thermal wave, as 234 says. With this reading, the 1 Displacement instant of the range is never found.
- NORDEND-231 and NORDEND-233 (NORDEND, 2026-11-05 01:00 to 01:15 and 03:30, Obstruction). CTR03 is rejected just above its threshold, and the families document gives Visibility or Displacement for a reference at its threshold with nothing seen in front of it. The detectors read Visibility (absorbed into the Poor visibility). The labels are right because of the thumbnail.
## Not fixed
- **The false Poor visibility 00:45 → 09:30 and 13:30 → 15:00, and over the benchmark window 11-05 13:15 → 11-07 15:15.** It comes from 1 to 3 sensors high at a time, one reference among them, bounded by CTR01's chronic rejections. Four rules were tried in a copy, and each makes another station go back:
- (A) no Poor visibility when no instant has more than `FEW` sensors high: JUNGFRAU Visibility found 64 → 51 %, LENZSPITZE Obstruction correct 23 → 6 %;
- (C) the rejections of a target rejected more than half the time do not bound a Poor visibility: JUNGFRAU found 64 → 51 %, CLARIDEN 3 → 0 %;
- (T) the same, for the targets in a Target issue episode: JUNGFRAU, CLARIDEN and NORDEND Visibility found down;
- (S) heat shimmer agitation without the steps at a reference rejection, against the night heat shimmer on the CTR03 comb: WILDSTRUBEL Visibility 79 → 13 %.
On JUNGFRAU, nights with two references at their threshold are labelled Visibility, and the measurements alone give no general line between those and one reference hindered by a thread.
- **The early night of the 05/11 (01:00 → 06:30, end of NORDEND-229)** stays without the thermal wave. The return of the 04/11 contraction reaches `RETURN` at about 01:00, and the gap profile is no longer monotone at night (21:00 to 23:00 rise again).
- **NORDEND-232 (PR09)**: bounds only in part (Brief obstructions at 07:15 to 07:45 and 09:30); the rest of its rejections are inside the Poor visibility above.
- **The CTR03 rejection at 03:30 on the range** is not a Brief obstruction. PR04 and PR11 rise with it (the compensation runs on two references), so it fails the "fewer than `BRIEF_MIN` high sensors" guard. Over the benchmark window it already is one, so the gain would only be on the range.
## Limits
- NORDEND-231 (2026-11-05T00:45Z → 01:30Z, Obstruction), NORDEND-233 (03:15Z → 03:45Z, Obstruction), NORDEND-236 (23:00Z → 11-06 10:00Z, Obstruction): only the thumbnail shows that a cobweb thread lies on CTR03. From the measurements alone, it is a reference at its threshold that makes the base jump. Seeing the cause would take the image (a bright blurred band on the reference's spot).
- NORDEND-232 (03:00Z → 09:45Z) and NORDEND-235 (14:45Z → 15:15Z), Obstruction: by day, PR09's thread is only deduced from the night thumbnail.
- NORDEND-230 (Target issue, CTR01): the range command reads 3 days of history, too short for the `WEEK` (7 days) of a Degraded target. Over the whole label range it is found.
## Tools
- `seances/NORDEND_2026-11-05T0600_2026-11-05T1200_20261116T1028Z/tools/raw_episodes.py <detectors> <station> FROM TO [detectors]`: the raw episodes of each detector (before classification) on a range plus 3 days of history. Worth keeping.
- `.../tools/high_sensors.py <detectors> <station> FROM TO SHOW_FROM SHOW_TO`: per instant, the sensors visibility.py calls high, the rejected ones, and the median MAD gap (point 8). Worth keeping for Visibility diagnoses.
- `.../tools/sensor_values.py <detectors> <station> SENSORS FROM TO [MEDIAN_FROM]`: error_value (×median, ×threshold), corr_x and corr_y of a few sensors. Worth keeping.
- `.../tools/thermal_gaps.py <detectors> <station> FROM TO`: the hourly gap profile used for the thermal bounds, on corr and on raw positions. Useful for thermal bounds.
- `.../tools/events_in_window.py <detectors> <station> FROM TO SHOW_FROM SHOW_TO`: classifies a long window, as the benchmark does, and shows the events of a sub-window. Worth keeping: range runs and benchmark runs differ (history, medians).
- `.../tools/bench_diff.py BEFORE AFTER`: the changed lines between two benchmark outputs, with regressions over 2 points flagged. Worth keeping as a workflow tool.
@@ -0,0 +1,7 @@
NORDEND: 2026-11-05T00:00Z → 2026-11-06T00:00Z, 9 labels of labeling/auto/NORDEND.jsonl, per family and per instant
found correct
Displacement 0/1 (0 %) 0/3 (0 %)
Obstruction 28/32 (88 %) 28/60 (47 %)
Target issue 0/96 (0 %) —
Thermal displacement 73/95 (77 %) 73/73 (100 %)
Visibility — 0/43 (0 %)
@@ -0,0 +1,7 @@
NORDEND: 2026-11-05T00:00Z → 2026-11-06T00:00Z, 9 labels of labeling/auto/NORDEND.jsonl, per family and per instant
found correct
Displacement 0/1 (0 %) 0/3 (0 %)
Obstruction 28/32 (88 %) 28/60 (47 %)
Target issue 0/96 (0 %) —
Thermal displacement 73/95 (77 %) 73/73 (100 %)
Visibility — 0/43 (0 %)
@@ -0,0 +1,7 @@
NORDEND: 2026-11-05T00:00Z → 2026-11-06T00:00Z, 9 labels of labeling/auto/NORDEND.jsonl, per family and per instant
found correct
Displacement 0/1 (0 %) 0/3 (0 %)
Obstruction 28/32 (88 %) 28/60 (47 %)
Target issue 0/96 (0 %) —
Thermal displacement 53/95 (56 %) 53/53 (100 %)
Visibility — 0/43 (0 %)
@@ -0,0 +1,72 @@
# improve: NORDEND, 2026-11-05 → 2026-11-06
**Result**: modification kept
## The range
NORDEND, 2026-11-05T00:00Z → 2026-11-06T00:00Z, 9 labels; before → end (found, correct)
- Displacement: 0/1 (0 %) → 0/1 (0 %); 0/3 (0 %) → 0/3 (0 %)
- Obstruction: 28/32 (88 %) → 28/32 (88 %); 28/60 (47 %) → 28/60 (47 %)
- Target issue: 0/96 (0 %) → 0/96 (0 %); —
- Thermal displacement: 53/95 (56 %) → 73/95 (77 %); 53/53 (100 %) → 73/73 (100 %)
- Visibility: —; 0/43 (0 %) → 0/43 (0 %)
## Elsewhere
`Total`, before → end (found, correct)
- Displacement: 5747/6346 (91 %); 5747/20444 (28 %) — unchanged
- No data: 75/836 (9 %); 75/100 (75 %) — unchanged
- Obstruction: 1044/4061 (26 %); 1044/3718 (28 %) — unchanged
- Target issue: 19080/22627 (84 %); 19080/19467 (98 %) — unchanged
- Thermal displacement: 10897/16142 (68 %) → 11986/16142 (74 %); 10897/14828 (73 %) → 11986/16069 (75 %)
- Visibility: 3412/7207 (47 %); 3412/4442 (77 %) — unchanged
Other changed lines:
- JUNGFRAU, Thermal displacement: 1544/2557 (60 %) → 2300/2557 (90 %); 1544/1786 (86 %) → 2300/2654 (87 %)
- NORDEND, Thermal displacement: 4769/7667 (62 %) → 5066/7667 (66 %); 4769/5097 (94 %) → 5066/5411 (94 %)
- SNCF01, Thermal displacement: 1218/1726 (71 %) → 1254/1726 (73 %); 1218/1275 (96 %) → 1254/1334 (94 %)
## What changed
- S1 (`3ea6b3e`), `scripts/thermal_displacement.py`: new constant `RETURN = 0.05`; the return of the thermal wave (contraction) now goes on until the hourly gap profile falls under `RETURN` times the amplitude, after midnight if needed, at most to the next day's start, instead of stopping at `EDGE` = 0.1, which now only bounds the start of the wave.
Upgrader: **changed**. Validator: **keep**.
## For the user to decide
From the modification, "Limits":
- NORDEND-231 (2026-11-05T00:45Z → 01:30Z, Obstruction), NORDEND-233 (03:15Z → 03:45Z, Obstruction), NORDEND-236 (23:00Z → 11-06 10:00Z, Obstruction): only the thumbnail shows that a cobweb thread lies on CTR03. From the measurements alone, it is a reference at its threshold that makes the base jump. Seeing the cause would take the image (a bright blurred band on the reference's spot).
- NORDEND-232 (03:00Z → 09:45Z) and NORDEND-235 (14:45Z → 15:15Z), Obstruction: by day, PR09's thread is only deduced from the night thumbnail.
- NORDEND-230 (Target issue, CTR01): the range command reads 3 days of history, too short for the `WEEK` (7 days) of a Degraded target. Over the whole label range it is found.
From the modification, "For the user to decide":
- NORDEND-243 (NORDEND, 2026-11-05 12:00, Displacement) against NORDEND-234 (Thermal displacement), whose description includes the same 12:00 jolt. The detectors absorb the one-measurement jump into the camera thermal wave, as 234 says. With this reading, the 1 Displacement instant of the range is never found.
- NORDEND-231 and NORDEND-233 (NORDEND, 2026-11-05 01:00 to 01:15 and 03:30, Obstruction). CTR03 is rejected just above its threshold, and the families document gives Visibility or Displacement for a reference at its threshold with nothing seen in front of it. The detectors read Visibility (absorbed into the Poor visibility). The labels are right because of the thumbnail.
From the validation, "Reservations": none.
## Not fixed
- **The false Poor visibility 00:45 → 09:30 and 13:30 → 15:00, and over the benchmark window 11-05 13:15 → 11-07 15:15.** It comes from 1 to 3 sensors high at a time, one reference among them, bounded by CTR01's chronic rejections. Four rules were tried in a copy, and each makes another station go back:
- (A) no Poor visibility when no instant has more than `FEW` sensors high: JUNGFRAU Visibility found 64 → 51 %, LENZSPITZE Obstruction correct 23 → 6 %;
- (C) the rejections of a target rejected more than half the time do not bound a Poor visibility: JUNGFRAU found 64 → 51 %, CLARIDEN 3 → 0 %;
- (T) the same, for the targets in a Target issue episode: JUNGFRAU, CLARIDEN and NORDEND Visibility found down;
- (S) heat shimmer agitation without the steps at a reference rejection, against the night heat shimmer on the CTR03 comb: WILDSTRUBEL Visibility 79 → 13 %.
On JUNGFRAU, nights with two references at their threshold are labelled Visibility, and the measurements alone give no general line between those and one reference hindered by a thread.
- **The early night of the 05/11 (01:00 → 06:30, end of NORDEND-229)** stays without the thermal wave. The return of the 04/11 contraction reaches `RETURN` at about 01:00, and the gap profile is no longer monotone at night (21:00 to 23:00 rise again).
- **NORDEND-232 (PR09)**: bounds only in part (Brief obstructions at 07:15 to 07:45 and 09:30); the rest of its rejections are inside the Poor visibility above.
- **The CTR03 rejection at 03:30 on the range** is not a Brief obstruction. PR04 and PR11 rise with it (the compensation runs on two references), so it fails the "fewer than `BRIEF_MIN` high sensors" guard. Over the benchmark window it already is one, so the gain would only be on the range.
## Tools
From the modification:
- `seances/NORDEND_2026-11-05T0600_2026-11-05T1200_20261116T1028Z/tools/raw_episodes.py <detectors> <station> FROM TO [detectors]`: the raw episodes of each detector (before classification) on a range plus 3 days of history. Worth keeping.
- `.../tools/high_sensors.py <detectors> <station> FROM TO SHOW_FROM SHOW_TO`: per instant, the sensors visibility.py calls high, the rejected ones, and the median MAD gap (point 8). Worth keeping for Visibility diagnoses.
- `.../tools/sensor_values.py <detectors> <station> SENSORS FROM TO [MEDIAN_FROM]`: error_value (×median, ×threshold), corr_x and corr_y of a few sensors. Worth keeping.
- `.../tools/thermal_gaps.py <detectors> <station> FROM TO`: the hourly gap profile used for the thermal bounds, on corr and on raw positions. Useful for thermal bounds.
- `.../tools/events_in_window.py <detectors> <station> FROM TO SHOW_FROM SHOW_TO`: classifies a long window, as the benchmark does, and shows the events of a sub-window. Worth keeping: range runs and benchmark runs differ (history, medians).
- `.../tools/bench_diff.py BEFORE AFTER`: the changed lines between two benchmark outputs, with regressions over 2 points flagged. Worth keeping as a workflow tool.
From the validation: none.
## Errors
none. Files already modified: none. Base: `e7c03e73041c81a7d338fb08ef16dad8eb8e9a84` (branch `main`).
```
3ea6b3e improve: S1
```
@@ -0,0 +1,30 @@
"""Compare two outputs of bench/regression.py: print the station/family lines that changed, flagging a regression (found or correct down by more than 2 points). Usage: python3 bench_diff.py BEFORE.txt AFTER.txt"""
import re
import sys
def parse(path):
out, station = {}, None
for line in open(path, encoding="utf-8"):
if line and not line.startswith(" ") and line.strip():
station = line.split(":")[0].strip()
continue
m = re.match(r"\s+(\S.*?)\s{2,}(\S+(?: \(\d+ %\))?|—)\s+(\S+(?: \(\d+ %\))?|—)\s*$", line)
if m and m.group(1) != "found":
out[(station, m.group(1))] = (m.group(2), m.group(3))
return out
def pct(x):
m = re.match(r"(\d+)/(\d+)", x)
return 100 * int(m.group(1)) / int(m.group(2)) if m and int(m.group(2)) else None
a, b = parse(sys.argv[1]), parse(sys.argv[2])
for k in sorted(set(a) | set(b)):
va, vb = a.get(k, ("—", "—")), b.get(k, ("—", "—"))
if va != vb:
reg = [n for n, x, y in zip(("found", "correct"), va, vb)
if pct(x) is not None and pct(y) is not None and pct(y) < pct(x) - 2]
print(f"{k[0]:<12} {k[1]:<22} found {va[0]} → {vb[0]} | correct {va[1]} → {vb[1]}",
"REGRESSION " + ",".join(reg) if reg else "")
@@ -0,0 +1,18 @@
"""Classify a long window (as the benchmark does: the labels' whole range, 3 days of history) and print only the events touching a shorter sub-window. Usage: python3 events_in_window.py <detectors dir> <station folder> FROM TO SHOW_FROM SHOW_TO"""
import sys
from datetime import timedelta
sys.path.insert(0, sys.argv[1] + "/scripts")
import categories # noqa: E402
from data import iso, load, parse_time # noqa: E402
folder, a, b, sa, sb = sys.argv[2:7]
ms = load(folder)
h, e = parse_time(a) - timedelta(days=3), parse_time(b)
chunk = {s: k for s, x in ms.items() if (k := [m for m in x if h <= m["t"] < e])}
sa, sb = iso(parse_time(sa)), iso(parse_time(sb))
for v in categories.events(chunk, folder=folder):
if v["end"] >= sa and v["start"] < sb:
print(v["start"], "→", v["end"], f"{v['category']:<20}", f"| {v['episodes'][0]['label']}",
f"| {len(v['sensors'])}:", ",".join(v["sensors"][:6]), "| absorbed:",
", ".join(f"{x['label']} {x['start'][5:16]}" for x in v["episodes"][1:])[:200] or "—")
@@ -0,0 +1,34 @@
"""Per instant of a window: the sensors visibility.py calls high (point 1), the rejected ones (R), and the median MAD gap of point 8, on the range plus 3 days of history. Usage: python3 high_sensors.py <detectors dir> <station folder> FROM TO SHOW_FROM SHOW_TO"""
import statistics
import sys
from datetime import timedelta
sys.path.insert(0, sys.argv[1] + "/scripts")
import visibility as v # noqa: E402
from data import iso, load, parse_time # noqa: E402
folder, a, b, sa, sb = sys.argv[2:7]
ms = load(folder)
h, e = parse_time(a) - timedelta(days=3), parse_time(b)
chunk = {s: k for s, x in ms.items() if (k := [m for m in x if h <= m["t"] < e])}
refs = {s for s, x in chunk.items() if x[0]["pattern_type"] == "global_ref"}
rows = {}
for s, x in chunk.items():
ev = [m["error_value"] for m in x if m["error_value"] is not None]
med = statistics.median(ev)
mad = statistics.median(abs(y - med) for y in ev) or 1e-9
for m in x:
y = m["error_value"]
if y is None or not (sa <= iso(m["t"]) <= sb):
continue
rej = m["error_threshold"] and y > m["error_threshold"]
hi = y > med + v.K_MAD * mad and (y > v.FACTOR * med or rej and y > v.REJECT_FACTOR * med)
r = rows.setdefault(iso(m["t"]), {"hi": [], "rej": [], "z": []})
r["z"].append((y - med) / mad)
if hi:
r["hi"].append(s + ("*" if s in refs else ""))
if rej:
r["rej"].append(s + ("*" if s in refs else ""))
for t in sorted(rows):
r = rows[t]
print(t, f"z={statistics.median(r['z']):5.1f}", "high:", ",".join(sorted(r["hi"])), "| R:", ",".join(sorted(r["rej"])))
@@ -0,0 +1,26 @@
"""Print the raw episodes of every detector (before classification) touching a range, computed on the range plus its 3 days of history (as score.py). Usage: python3 raw_episodes.py <detectors dir> <station folder> FROM TO [detector names, comma-separated]"""
import inspect
import sys
from datetime import timedelta
sys.path.insert(0, sys.argv[1] + "/scripts")
import categories # noqa: E402
from data import iso, load, parse_time # noqa: E402
folder, start, end = sys.argv[2], iso(parse_time(sys.argv[3])), iso(parse_time(sys.argv[4]))
only = sys.argv[5].split(",") if len(sys.argv) > 5 else None
ms = load(folder)
h, b = parse_time(start) - timedelta(days=3), parse_time(end)
chunk = {s: k for s, v in ms.items() if (k := [m for m in v if h <= m["t"] < b])}
for mod in categories.detectors():
name = mod.__name__.removeprefix("detector_")
if only and name not in only:
continue
params = inspect.signature(mod.detect).parameters
kw = {k: categories._context(folder, k) for k in categories.CONTEXT if k in params}
if "end" in params:
kw["end"] = None
for e in mod.detect(chunk, **kw):
if e["end"] >= start and e["start"] < end:
print(name, e["start"], "→", e["end"], e["kind"], "|", e["label"], e.get("level", ""), "|", len(e["sensors"]),
",".join(e["sensors"][:6]), "|", e["text"][:90])
@@ -0,0 +1,20 @@
"""Print error_value (with its ratio to the sensor's median over the window read and to its threshold) and corr_x, corr_y of given sensors on a window. Usage: python3 sensor_values.py <detectors dir> <station folder> SENSORS(comma) FROM TO [MEDIAN_FROM]"""
import statistics
import sys
sys.path.insert(0, sys.argv[1] + "/scripts")
from data import iso, load # noqa: E402
folder, sensors, a, b = sys.argv[2], sys.argv[3].split(","), sys.argv[4], sys.argv[5]
m0 = sys.argv[6] if len(sys.argv) > 6 else "0000"
ms = load(folder)
for s in sensors:
ev = [m["error_value"] for m in ms[s] if m["error_value"] is not None and m0 <= iso(m["t"]) < b]
med = statistics.median(ev)
print(f"== {s} median {med:.3f}")
for m in ms[s]:
if a <= iso(m["t"]) <= b:
e, th = m["error_value"], m["error_threshold"]
f = lambda x: " - " if x is None else f"{x:7.3f}" # noqa: E731
print(iso(m["t"]), f(e), f"x{e / med:4.2f}" if e else "", f"/th {e / th:4.2f}" if e and th else "",
f(m["corr_x"]), f(m["corr_y"]))
@@ -0,0 +1,17 @@
"""Print the hourly gap profile (share of amplitude from the night level) that thermal_displacement.py uses for its bounds, for corr and raw positions, on a range plus 3 days of history. Usage: python3 thermal_gaps.py <detectors dir> <station folder> FROM TO"""
import sys
from datetime import timedelta
sys.path.insert(0, sys.argv[1] + "/scripts")
import thermal_displacement as td # noqa: E402
from data import load, parse_time # noqa: E402
folder, a, b = sys.argv[2:5]
ms = load(folder)
h, e = parse_time(a) - timedelta(days=3), parse_time(b)
chunk = {s: k for s, x in ms.items() if (k := [m for m in x if h <= m["t"] < e])}
for cols in (("corr_x", "corr_y"), ("x", "y")):
concerned, n = td.cycle(chunk, cols)
print(cols, f"{len({c[0] for c in concerned})}/{n} targets")
if concerned:
print(" ".join(f"{h:02d}:{g:.2f}" for h, g in enumerate(td.gaps(concerned))))
@@ -0,0 +1,26 @@
# Validation
**Verdict**: keep
## Answers
1. Detectors load: no `skipped` line in benchmark-before, -S1 or -after; rerunning the benchmark at HEAD gives output identical to benchmark-after.
2. No station goes back: only Thermal displacement changes, on JUNGFRAU, NORDEND and SNCF01. "Found" rises on all three. The worst "correct" drop is SNCF01, 95.5 → 94.0 (-1.5 points), which is under 2 points.
3. Simple and proportionate: one constant (`RETURN`) and one changed comparison in `_episodes`. The rule is generic: the cooling return ends at a lower share of the amplitude than the start does. The range gains +20 instants (56 → 77 %) and the whole benchmark gains +1089 found instants (68 → 74 %) while "correct" rises (73 → 75 %). The constant is not cut to the range's bounds.
4. Rules: still six detectors, one per family. Contract and episode keys are unchanged. No station-, sensor- or date-specific rule (station names appear only in provenance comments). Standard library only. `thresholds.md` has the new `RETURN` row and an updated `EDGE` row, matching the constants in the code.
## The range
NORDEND 2026-11-05T00:00Z → 2026-11-06T00:00Z
- Thermal displacement: found 53/95 (56 %) → 73/95 (77 %); correct 53/53 (100 %) → 73/73 (100 %)
- Displacement, Obstruction, Target issue, Visibility: unchanged
## Non-regression
- JUNGFRAU, Thermal displacement: found 1544/2557 (60 %) → 2300/2557 (90 %); correct 1544/1786 (86.4 %) → 2300/2654 (86.7 %)
- NORDEND, Thermal displacement: found 4769/7667 (62 %) → 5066/7667 (66 %); correct 4769/5097 (93.6 %) → 5066/5411 (93.6 %)
- SNCF01, Thermal displacement: found 1218/1726 (71 %) → 1254/1726 (73 %); correct 1218/1275 (95.5 %) → 1254/1334 (94.0 %)
- Total, Thermal displacement: found 10897/16142 (68 %) → 11986/16142 (74 %); correct 10897/14828 (73 %) → 11986/16069 (75 %)
## Reservations
none
## Tools
none