docs: separate Cadryl catalogues from the native PyTorch cascade

.gitattributes

@@ -33,3 +33,4 @@ saved_model/**/* filter=lfs diff=lfs merge=lfs -text
 *.zip filter=lfs diff=lfs merge=lfs -text
 *.zst filter=lfs diff=lfs merge=lfs -text
 *tfevents* filter=lfs diff=lfs merge=lfs -text
+docs/media/cascade-v25.png filter=lfs diff=lfs merge=lfs -text

CHANGELOG.md

@@ -2,7 +2,7 @@
 
 ## 10 October 2026 · English documentation and navigation
 
-Added an English quick card and shared navigation to task contracts, model/dataset relationships and variant guidance. Added the v2.5 branch diagram; it illustrates existing contracts rather than new execution.
+Added an English quick card and shared navigation to task contracts, model/dataset relationships and variant guidance. Added the v2.5 branch diagram; it illustrates existing contracts rather than new execution. The native PyTorch cascade and its NVIDIA L4 diagnostic are presented independently of the conversion archive.
 
 All newly edited presentation text, figures, guides and comparison captions are in English. Repository IDs, paths, original labels, technical contracts, weights, annotations, source rights and visibility are preserved. Original legal/attribution records retain their governing wording.
 

README.md

@@ -2,7 +2,7 @@
 
 **Visible fire/smoke research · documented datasets · geographic context.**
 
-8 public model repositories, 9 public datasets and one organization-card Space, reviewed on **10 October 2026**.
+This index presents seven native research-model repositories, nine public datasets and the organization-card Space, reviewed on **10 October 2026**.
 
 ## Quick card
 
@@ -11,34 +11,32 @@
 | Use | Find research models, documented corpora and geographic scene resources. |
 | Expected input | Task-specific images, ROIs or geographic packages with their exact contracts. |
 | Example | Start with a selected cascade component, follow its dataset and read the matching inference contract. |
-| Limits | Research resources; validation, external diagnostics, CPU parity and device qualification have distinct scopes. |
-| Contract | [Cascade contracts](https://huggingface.co/fireviewer/litert-models/blob/main/docs/CASCADE_V25.md) · [Dataset guide](docs/DATASET_V25.md) |
+| Limits | Research resources; selected validation, external diagnostics and geographic observations have distinct scopes. |
+| Contract | [Native cascade contracts](docs/CASCADE_V25.md) · [Dataset guide](docs/DATASET_V25.md) |
 
 [Choose a variant](https://huggingface.co/spaces/fireviewer/README/blob/main/docs/CHOOSING_VARIANTS.md) · [Resource navigation](https://huggingface.co/spaces/fireviewer/README/blob/main/docs/RESOURCE_INDEX.md) · [Update log](CHANGELOG.md)
 
-<img src="docs/media/cascade-v25.png" width="720" alt="FireViewer v2.5: full-image proposals, visible-flame ROIs, per-box point anchors and a separate smoke compatibility branch" />
+<img src="docs/media/cascade-v25.png" width="720" alt="Native FireViewer PyTorch cascade: proposals, visible-flame ROIs, visual anchors and a separate smoke compatibility branch" />
 
 ## Collections
 
 | Collection | Contents |
 | --- | --- |
-| [FireViewer v2.5 cascade: selected models](https://huggingface.co/collections/fireviewer/fireviewer-v25-cascade-selected-models-6ac47be5c15d70ee2a01d464) | Four selected PyTorch models and their four inference-only FP32 LiteRT exports. |
-| [FireViewer Calibration and Tests for Full Pipeline](https://huggingface.co/collections/fireviewer/fireviewer-calibration-and-tests-for-full-pipeline-6ac36983b2eaaf428a7577e1) | Measured geographic scenes and calibration resources; access restrictions remain applicable. |
-| [Training Datasets — v2.5](https://huggingface.co/collections/fireviewer/training-datasets-v25-6ac362a45c747ad9e99c9efb) | Three primary datasets and two point-annotation extensions. |
+| [FireViewer v2.5 · selected cascade models](https://huggingface.co/collections/fireviewer/fireviewer-v25-cascade-selected-models-6ac47be5c15d70ee2a01d464) | Four selected native PyTorch models: proposals, visible flames, visual anchors and smoke compatibility. |
+| [FireViewer · pipeline calibration and tests](https://huggingface.co/collections/fireviewer/fireviewer-calibration-and-tests-for-full-pipeline-6ac36983b2eaaf428a7577e1) | Measured geographic scenes and calibration resources; access restrictions remain applicable. |
+| [FireViewer v2.5 · datasets and point extensions](https://huggingface.co/collections/fireviewer/training-datasets-v25-6ac362a45c747ad9e99c9efb) | Three primary datasets and two point-annotation extensions. |
 | [(not good enough) Pilot and Experimental Models](https://huggingface.co/collections/fireviewer/not-good-enough-pilot-and-experimental-models-6ac362a31a779b0c63bb3f4a) | Retained pilot and experimental models, with their limitations. |
 | [FireViewer · historical corpora and negatives](https://huggingface.co/collections/fireviewer/fireviewer-corpus-historiques-et-negatifs-6a70a1f506d72626d7068e2a) | Historical V1/v2.5 detection corpora and difficult negatives. |
 
 ## Evidence and remaining limits
 
-The external **680-image cascade diagnostic** alerts on 220/249 positive images (88.35%) and 49/431 negatives (11.37%). It localizes 231/901 flame boxes (25.64%) and 178/788 small flame boxes (22.59%) at IoU ≥0.5. This is not the locked final TEST. Independent flame-base accuracy is unmeasured; PyTorch/L4 timings are not LiteRT phone timings. [Original report](https://huggingface.co/fireviewer/litert-models/blob/main/reports/cascade-v25-680/REPORT.md).
-
-The [four recent LiteRT exports](https://huggingface.co/fireviewer/litert-models) are inference-only and retain no Android-device qualification. Six historical LiteRT variants, including five trainable heads/adapters, keep their separate evidence. [All models and datasets](docs/RESOURCE_INDEX.md).
+The external **680-image cascade diagnostic** alerts on 220/249 positive images (88.35%) and 49/431 negatives (11.37%). It localizes 231/901 flame boxes (25.64%) and 178/788 small flame boxes (22.59%) at IoU ≥0.5. This is not the locked final TEST. Independent flame-base accuracy is unmeasured. Recorded timings concern native PyTorch on NVIDIA L4. [Diagnostic report](reports/cascade-v25-680/REPORT.md) · [All models and datasets](docs/RESOURCE_INDEX.md).
 
 ## Presentation and governance
 
 This README Space hosts documentation and the organization card; no running inference demo is present. FireViewer keeps observations, reconstructions and simulations distinct. Sensitive outputs require human review, and geographic scene packages are distinct from verified incident observations.
 
-FIRE-VIEWER provides administrative and financial stewardship for resources under its control. Technical governance is maintainer-led; upstream, source-specific and pre-association rights remain separate. [Unicorn Who Dev](https://huggingface.co/unicornwhodev) publishes Cadryl and other creative tools separately.
+FIRE-VIEWER provides administrative and financial stewardship for resources under its control. Technical governance is maintainer-led; upstream, source-specific and pre-association rights remain separate.
 
 [Association](https://fire-viewer.fr) · [Documentation](https://github.com/fireviewer/Fireviewer_doc/blob/main/docs/public/HUGGINGFACE.md) · contact@fire-viewer.fr
 

docs/CASCADE_V25.md

@@ -0,0 +1,63 @@
+# FireViewer v2.5 · native PyTorch cascade
+
+[Presentation](../README.md) · [Choose a component](CHOOSING_VARIANTS.md) · [Dataset guide](DATASET_V25.md) · [Update log](../CHANGELOG.md)
+
+Four selected native PyTorch checkpoints implement full-image proposals, visible-flame refinement, one visual anchor per accepted flame crop and a separate smoke-compatibility branch. The published cascade diagnostic used PyTorch on NVIDIA L4.
+
+| Stage | Native checkpoint | Input |
+| --- | --- | --- |
+| Generate single-class probable_fire candidate regions from full images. | [rfdetr-medium-probable-fire-v2.5](https://huggingface.co/fireviewer/rfdetr-medium-probable-fire-v2.5) | 1088 px |
+| Find visible flames inside probable-fire proposals. | [rfdetr-medium-visible-flame-v2.5](https://huggingface.co/fireviewer/rfdetr-medium-visible-flame-v2.5) | 1088 px |
+| Predict one visual flame anchor for each detected flame box. | [dinov2-small-visible-flame-monopoint-v2.5](https://huggingface.co/fireviewer/dinov2-small-visible-flame-monopoint-v2.5) | 448 px |
+| Estimate fire-associated smoke compatibility with an abstention interval. | [dinov2-small-fire-associated-smoke-v2.5](https://huggingface.co/fireviewer/dinov2-small-fire-associated-smoke-v2.5) | 448 px |
+
+![Native PyTorch cascade](media/cascade-v25.png)
+
+## Functional routing
+```mermaid
+flowchart TD
+    I[RGB whole image] --> P[RF-DETR M probable_fire · 1088]
+    P --> R[score .10 · NMS .75 · at most 6 ROI]
+    R --> V[30% context · min 128px · RF-DETR M visible_flame · 1088]
+    V --> F[Flame boxes · score .25 · reprojection · global NMS .60]
+    F --> K[Tight flame crops · DINOv2-S heatmap · 448]
+    K --> A[One flame_base_anchor per detected flame]
+    V --> G[ROI without a local confirmed flame]
+    G --> S[45% context · min 160px · DINOv2-S smoke classifier · 448]
+    S --> T[Calibrated compatibility · low / ambiguous / high]
+```
+
+Context is added on each side: expanded width and height are multiplied by `1 + 2 × context`, before the minimum side and image clipping. RF-DETR receives NCHW float32 RGB resized with half-pixel bilinear interpolation, **no antialias**, with RGB converted to float before resize; normalization is ImageNet mean `[.485,.456,.406]`, std `[.229,.224,.225]`. Point crops use Pillow bilinear; smoke crops use Pillow bicubic. All models stretch to the specified square size. The supplied NumPy/Pillow runner preserves these differences.
+
+Proposal score ≥0.10, NMS IoU 0.75, limit 6; visible score ≥0.25 and global image-space NMS 0.60, limit 100. Tight point crops use floor for the minimum and ceil for the maximum. The heatmap point is the argmax followed by a local 5×5 softmax expectation at heatmap pixel centers. It is an appearance anchor, **not an ignition location**, geographic position or independently calibrated source confidence.
+
+Smoke temperature is **1.1490650177001953**. Low compatibility ≤**0.30160276398937463**; high compatibility ≥**0.8199750333466715**; values between these are ambiguous and do not raise the benchmark image alarm. Smoke is a binary visual compatibility classifier. It supplies no mask, smoke-core, smoke point or origin. No learned fire-group association is performed.
+
+## Recorded external diagnostic · 680 images
+Frozen PyTorch FP32 pipeline on L4; no fitting or threshold selection on this panel. Case mix: 179 flame, 70 smoke, 431 negative images, 229 visual families. There may be several views of a family; families are not verified independent fire incidents.
+
+| Endpoint | Result | Numerator / denominator |
+|---|---:|---:|
+| Positive-image alarm recall | 88.35% | 220 / 249 |
+| Sampled image alarm precision | 81.78% | 220 / 269 |
+| Negative-image false-alarm rate | 11.37% | 49 / 431 |
+| Flame-image any-branch alarm recall | 99.44% | 178 / 179 |
+| Smoke-only any-branch alarm recall | 60.00% | 42 / 70 |
+| Flame bbox recall, IoU ≥ .5 | 25.64% | 231 / 901 |
+| Flame bbox precision, IoU ≥ .5 | 27.57% | 231 / 838 |
+| Tiny flame recall | 22.59% | 178 / 788 |
+| Multi-flame instance recall | 24.72% | 219 / 886 |
+| Expanded-proposal containment | 97.89% | 882 / 901 |
+| Point inside matched GT flame bbox | 97.84% | 226 / 231 |
+| Smoke-only image with false flame | 10.00% | 7 / 70 |
+
+An image alarm is any visible-flame output **or** high smoke compatibility; it is not a statement that every flame has a precise box. There were TP=220, FP=49, FN=29, TN=382 for image alarms; flame localization TP=231, FP=607, FN=670. Tiny means bbox area/image area ≤ `(32/1088)^2`, a protocol bin rather than original-pixel COCO-small. Proposal containment means at least 90% of the GT box lies within an expanded proposal. The 229-image one-representative-per-family panel still has tiny recall 101/399=25.31%.
+
+**Interpretation:** stage 1 generally provides coverage, while precise instance separation/localization of small visible flames remains the primary weakness. In the inspected worst-error gallery, several small annotated flames are merged into a broader detection. This is an observed error pattern, not a proven causal attribution for every miss. Good image suspicion and low box recall can coexist. The point-in-box statistic measures consistency only: no independent human flame-base GT is available, so PCK and base accuracy are unmeasured. Smoke-only labels establish visible smoke, not independently verified fire association.
+
+Mean end-to-end inference latency was 226.82 ms/image, median 172.44 ms, P95 458.01 ms on **L4 PyTorch**. All invoked stages, decoding, preprocessing, NMS and reprojection are included; model loading/warmup are excluded. Stage calls: proposal 680, visible 1636, point 838, smoke 855. These measurements describe the frozen native PyTorch cascade on NVIDIA L4.
+
+The image-level SHA/perceptual novelty audit covered all known historical splits including private sources and legacy FASDD/Pyro-SDIS manifests. Detected neighbors and suspected FLAME1 scenes were excluded before inference. No detected duplicate proves neither unseen events nor complete unknown provenance. Sources: [DIFD](https://zenodo.org/records/15860575) and [SAINetset v8](https://huggingface.co/datasets/SAINetset/SAINetset_v8.0), whose images may include synthetic/donated material. Only 208 retained images had a visual QA pass; there is no independent human annotation audit. Ratios describe this selected correlated panel, with no independent-event confidence intervals or deployment certification.
+## Sources
+
+[Retained scientific source](https://huggingface.co/fireviewer/litert-models/blob/261c16b47744b893261481da089a4d9d5de569aa/docs/CASCADE_V25.md) · [Original diagnostic report](https://huggingface.co/fireviewer/litert-models/blob/261c16b47744b893261481da089a4d9d5de569aa/reports/cascade-v25-680/REPORT.md) · [Original metric JSON](https://huggingface.co/fireviewer/litert-models/blob/261c16b47744b893261481da089a4d9d5de569aa/reports/cascade-v25-680/metrics.json) · [Frozen release manifest](https://huggingface.co/fireviewer/litert-models/blob/261c16b47744b893261481da089a4d9d5de569aa/cascade/v2.5/source-release-manifest.json). Source-specific rights and the original evidence dates remain applicable. This English presentation extracts native-model routing and diagnostic results from the retained records; it does not rerun the cascade.

docs/CHOOSING_VARIANTS.md

@@ -11,17 +11,17 @@
 | Predict one visual flame anchor for each detected flame box. | [dinov2-small-visible-flame-monopoint-v2.5](https://huggingface.co/fireviewer/dinov2-small-visible-flame-monopoint-v2.5) | Tight RGB flame crop, bilinear 448×448 and ImageNet normalization. | Trained on pseudo-points. No independent human flame-base accuracy or calibrated point confidence is established. |
 | Estimate fire-associated smoke compatibility with an abstention interval. | [dinov2-small-fire-associated-smoke-v2.5](https://huggingface.co/fireviewer/dinov2-small-fire-associated-smoke-v2.5) | ROI without confirmed flame; 45% context, minimum side 160, bicubic 448×448 and ImageNet normalization. | No smoke point, source location, mask or physical aerosol density is inferred. |
 
-For Android experimentation, use the corresponding **four FP32 inference-only exports** in [fireviewer/litert-models](https://huggingface.co/fireviewer/litert-models). Retain crop transforms, operating thresholds, smoke calibration and abstention. The point and smoke stages are distinct branches.
+The selected cascade uses these four native PyTorch checkpoints. Retain crop transforms, operating thresholds, smoke calibration and abstention. Point prediction and smoke compatibility are distinct branches. [Cascade contract](CASCADE_V25.md).
 
-## Historical alternatives and adaptation
+## Historical alternatives
 
-Six historical LiteRT variants include five trainable heads/output adapters on frozen backbones. Read each exact contract; adapting a detector cannot create proposals absent from its frozen detector. A historical SDK emulator pass does not qualify the current application or a physical ARM phone.
+The retained YOLO and RT-DETR reference detectors and the DINOv3 pointing pilot have their own tasks, source terms and evaluation protocols. Choose the required output semantics before comparing architectures.
 
 ## Read the right evidence
 
 - Selected v2.5 scores are validation/selection evidence with task-specific populations.
 - The external **680-image diagnostic** is a cascade diagnostic, not a per-model score or the locked final TEST.
 - Historical independent detector benchmarks and pilot validation have different protocols; do not rank them against v2.5 or pseudo-point agreement.
-- Source PyTorch/L4 timings, LiteRT CPU parity and Android latency are different measurements.
+- The recorded full-cascade timings concern PyTorch on NVIDIA L4.
 
 No new shared benchmark is run here. The retained reports, weak tiny-flame localization, unknown independent flame-base accuracy and source-specific restrictions remain visible.

docs/RESOURCE_INDEX.md

@@ -10,7 +10,6 @@
 | [fireviewer/rfdetr-medium-visible-flame-v2.5](https://huggingface.co/fireviewer/rfdetr-medium-visible-flame-v2.5) | Find visible flames inside probable-fire proposals. |
 | [fireviewer/dinov2-small-visible-flame-monopoint-v2.5](https://huggingface.co/fireviewer/dinov2-small-visible-flame-monopoint-v2.5) | Predict one visual flame anchor for each detected flame box. |
 | [fireviewer/dinov2-small-fire-associated-smoke-v2.5](https://huggingface.co/fireviewer/dinov2-small-fire-associated-smoke-v2.5) | Estimate fire-associated smoke compatibility with an abstention interval. |
-| [fireviewer/litert-models](https://huggingface.co/fireviewer/litert-models) | Ten LiteRT variants with distinct historical and v2.5 contracts. |
 | [fireviewer/dinov3-pointing-pilot-v1](https://huggingface.co/fireviewer/dinov3-pointing-pilot-v1) | Inspect the retained five-epoch multi-task research pilot. |
 | [fireviewer/fire-smoke-yolo11m-strict-v1](https://huggingface.co/fireviewer/fire-smoke-yolo11m-strict-v1) | Use the retained visible smoke/flame baseline as a preannotation helper. |
 | [fireviewer/rtdetr-v2-r50-fire-smoke](https://huggingface.co/fireviewer/rtdetr-v2-r50-fire-smoke) | Inspect the historical visible smoke/flame reference detector. |

reports/cascade-v25-680/REPORT.md

@@ -0,0 +1,33 @@
+# FireViewer v2.5 - external natural-scene cascade benchmark
+
+Status: **complete**. Fixed subset: 680/680 images. Hardware: NVIDIA L4. No training or threshold tuning was performed. All four model revisions and weight digests are in the release manifest and were verified before loading.
+
+The recipe is full-image probable-fire proposals, visible-flame refinement in proposal crops, flame-anchor heatmap in each detected flame crop, and fire-associated-smoke compatibility on proposals without a visible flame. Smoke is a compatibility output, without a precise smoke mask or source point.
+
+| Measurement | Result |
+|---|---:|
+| Image positive recall | 88.35% (220/249) |
+| Image precision at the sampled case mix | 81.78% (220/269) |
+| Negative-image false-alarm rate | 11.37% (49/431) |
+| Flame-box precision at IoU 0.50 | 27.57% (231/838) |
+| Flame-box recall at IoU 0.50 | 25.64% (231/901) |
+| Relative tiny-flame recall | 22.59% (178/788) |
+| Multiple-flame recall | 24.72% (219/886) |
+| GT flame coverage by expanded proposal crops | 97.89% (882/901) |
+| Point inside matched flame GT box - consistency only | 97.84% (226/231) |
+| Full cascade mean / p50 / p95 | 0.227 / 0.172 / 0.458 s |
+
+Ground-truth boxes are red for flame and blue for smoke; final flame predictions and their anchors are green. Orange rectangles show high smoke compatibility on a proposal crop, not smoke localization. Errors are sorted only for the post-benchmark gallery.
+
+The SHA and perceptual overlap audit covers the original raw corpus, all historical splits and private sources, plus both legacy FASDD and Pyro-SDIS manifests. Known neighbors and suspected FLAME1 snow/4K source scenes were excluded before inference. No detected duplicate does not prove event independence. DIFD is a separately distributed source with image-level novelty checks, but its event provenance is incomplete. SAINetset can include synthetic or donated images; individual capture provenance is not provided. The tenfold panel contains 229 visual families and may retain several distinct views of a family. The original 68-image panel and one representative per visual family are evaluated separately in metrics.json. Visual families are not verified independent events. Correlated video/camera scenes and this deliberately selected case mix limit external-validity claims. Ratios are descriptive; no independent-event confidence intervals are claimed.
+
+The archived original YOLO annotations and QA exclusions are retained. Smoke boxes supervise visible-smoke compatibility only; they do not independently establish fire association. Operational COCO AP uses the unchanged 0.25 flame cutoff and is explicitly score-truncated. Tiny means flame bbox area / image area <= (32/1088)^2, a protocol bin rather than original-pixel COCO small. No independent human flame-base anchors are available: point PCK, base accuracy and ignition accuracy are not measured. Historical TEST performance remains locked.
+
+Latency includes decoding, preprocessing, every invoked stage, NMS and coordinate remapping with CUDA synchronization, excluding load and warmup. Hardware, software versions, peak allocation, calls, predictions and all hashes are in the JSON receipts. A wall-clock cap prevents an unattended benchmark from continuing indefinitely; a capped run is reported as partial.
+
+Sources: [DIFD, authors' Zenodo record](https://zenodo.org/records/15860575), [SAINetset v8.0, AlterMundi](https://huggingface.co/datasets/SAINetset/SAINetset_v8.0).
+
+
+## Source and scope
+
+English mirror of the [original immutable report](https://huggingface.co/fireviewer/litert-models/blob/261c16b47744b893261481da089a4d9d5de569aa/reports/cascade-v25-680/REPORT.md). The recorded runtime is native PyTorch on NVIDIA L4. No new measurements were produced for this documentation update. [Metrics and machine-readable provenance](https://huggingface.co/fireviewer/litert-models/blob/261c16b47744b893261481da089a4d9d5de569aa/reports/cascade-v25-680/metrics.json).