Curation routes

Experimental

The curation subsystem is experimental in v0.6.0. It is real and tested, but backward compatibility across releases is not guaranteed until it graduates. It ships opt-in behind the curation compose profile.

The generated contract, contracts/openapi/curation.json, lists every route and is the source of truth for every path, parameter and schema. This page groups them by area. Request models reject unknown keys (extra='forbid'), so a stale or misspelled field is a 422, never a silent no-op. Consumers should read the mount prefix (OP_API_PREFIX, default /curation) from configuration rather than hardcode it. There is no /v1 twin and no alias.

Projects and scoping

A project is an isolated dataset workspace with its own indexes, class registry, exports, settings and job state. Every route that reads or writes data lives under /curation/projects/{project}/. There is no unscoped alias: an unscoped data path is a 404. Paths on this page omit that prefix unless they are global.

The routes outside a project:

Method	Path	Purpose
GET	/curation/health	Deployment facts only: status, Triton, OpenSearch, VLM, versions
GET	/curation/events	Server-sent events with project: null: project.*, combine.*, vlm.changed
GET	/curation/projects	List projects (?include_archived=true), shard capacity, slug limits, display labels
POST	/curation/projects	Create a project (slug, display_name, optional description, clone_settings_from, clone_axes). 201
GET	/curation/projects/{project}	One project: status, revision, counts, resources, flags (writable, deletable, ...)
PATCH	/curation/projects/{project}	Edit display_name / description; needs expected_revision
POST	/curation/projects/{project}/archive	Archive (read-only). /unarchive reverses it
POST	/curation/projects/{project}/clone_settings	Copy settings axes from another project
DELETE	/curation/projects/{project}	?dry_run=true reports what would go; a real delete needs ?confirm=<slug> and answers 202 with status deleting (async: the project answers 404 a few seconds later, and the slug is retired and cannot be reused)
GET	/curation/vlm/endpoints	The VLM endpoint registry (see VLM endpoints)
GET	/curation/vlm/catalog	The local VLM catalog

A project's status is building, active, archived, deleting, deleted or failed. Only an active project is writable; an archived one answers writes with 409 project_read_only. The default project is an ordinary project that can be archived but never deleted. Lifecycle mutations answer one envelope, {project, warnings}, so a client can adopt the returned summary without a re-read.

Errors from project, config and import routes share one body, {"detail": {"error": "<code>", "message": "...", ...}}, with code-specific extra fields. Read endpoints fail closed: a backend error is an HTTP error, never a silently empty result.

Health, discovery and settings

Method	Path	Purpose
GET	/curation/projects/{project}/health	status ok / degraded / down plus project, triton, opensearch, vlm, registry, active region_profile, mlflow_public_url
GET	/models/status	Every model the curation pipeline depends on; VLM rows are one per registered endpoint (kind: "vlm", with active)
DELETE	/models/{model_name}	Unload a promoted model. Models the active profile depends on are refused (403); other core models need ?force=true (409 without it)
GET	/models/{model_name}/class_mapping	How a model's classes map by name onto this project's classes
PUT	/models/{model_name}/sharing	Opt a promoted model into or out of use by other projects
GET	/methods	Capability discovery: axes and strategies per axis (cluster, score, sort, overlay, export, detection_profile, prompt_pack, vlm) and feature flags. Gate optional UI on this, never on a probing write
GET	/settings	Shared per-axis strategy defaults (defaults is an open map keyed by axis id)
PUT	/settings	Partial update. defaults.vlm activates an endpoint through the same gate as /vlm/endpoints/{name}/activate; defaults.prompt_pack and defaults.detection_profile delegate to the config store's activation
GET	/config/vocabulary	One place every model choice is served: the active VLM, endpoints, and a model_choices table
GET	/class_sources	Every class_source value this deployment can write, with role and label
GET	/stats	Project counts
GET	/stats/dataset	Dataset totals (labeled by human / VLM / classifier / import, regions, clusters, in-flight work)
GET	/stats/classes	Per-class counts
GET	/events	Server-sent events for this project (for example crop.region_verified, config.changed)
POST	/events/publish	Cross-process publish endpoint the workers use
GET	/events/stats	Hub stats for this project's stream
GET	/pause	Whether the project's pipeline is paused, and by what
POST	/pause	Workers skip this project's work. /resume clears it

Ingest

Ingest runs the configured item detector, crops each detection into an item, embeds it, rejects duplicates and indexes in bulk. Images are read from configured source roots or uploaded as bytes. To import an already-labeled dataset, use Dataset import.

Method	Path	Purpose
POST	/ingest/image	Ingest one image by path (under a configured source root)
POST	/ingest/batch	Ingest items by path; capped by OP_BATCH_MAX_ITEMS_PER_REQUEST (413 over it)
POST	/ingest/upload	Multipart upload of images; bytes are persisted content-addressed under the project's upload root; optional run_id
POST	/ingest/path_lookup	Which image paths (or upload identifiers) are already indexed
GET	/ingest/config	Served upload/batch limits and drain-poll parameters
GET	/ingest/status	Totals by source and day (optionally ?run_id=)
GET	/ingest/region_drain	Pending region work, a stability verdict (drained), per-dependency health and a stall_reason

Each image result carries status (success / duplicate / failed), image_id, n_crops, n_regions, and on failure an error_kind (empty, unservable_path, unsupported_type, decode_failed, detector_infer, bulk_index, ...). Byte-identical files are deduplicated within a batch as well as against the index.

Dataset import

Bring a labeled YOLO or COCO dataset, or an OpenProcessor export, into a project. Classes map by name. The import is chunked, resumable and undoable. See Dataset import.

Method	Path	Purpose
GET	/datasets/formats	The formats the importer reads
POST	/datasets/uploads	Upload a zip or tar archive (201, 413 upload_too_large)
POST	/datasets/preview	Scan a dataset without writing: classes, counts, issues, and suggestions for the class mapping
POST	/datasets/imports	Start an import. 202; reused when the same import_key already ran
GET	/datasets/imports	List imports
GET	/datasets/imports/{import_id}	Status, progress and the report
GET	/datasets/imports/{import_id}/issues	The scan issues, paged
GET	/datasets/imports/{import_id}/entries	Per-image ledger rows, paged
POST	/datasets/imports/{import_id}/cancel	Stop after the current chunk
POST	/datasets/imports/{import_id}/resume	Continue an interrupted or cancelled import (202)
POST	/datasets/imports/{import_id}/undo	Reverse the import

The mapping is a list of {dataset_class, action, ...} rows where action is map (class_id), create (new_class_name, new_class_group), skip or region. Every class that has boxes needs a decision (422 class_mapping_incomplete lists the unmapped names). A COCO layout (images/ next to annotations/*.json) is detected automatically.

Reprocess

One route re-runs pipeline stages over existing data. See Reprocess.

Method	Path	Purpose
POST	/reprocess	targets (exactly one of image_ids, crop_ids, filter), scopes (detect, region, vlm, embed), region_mode (redetect / reverify), dry_run (default true)
POST	/images/{image_id}/reprocess	The per-image button; applies by default
POST	/crops/{crop_id}/reprocess	The per-item button; applies by default
GET	/reprocess/jobs/{job_id}	Progress of a job (detect or embed beyond OP_REPROCESS_SYNC_MAX images runs as a file-backed job; one live job per project, 409 reprocess_busy)
POST	/reprocess/jobs/{job_id}/cancel	Stop a live job after its current chunk

Locked items and boxes are never written; each scope reports locked_skipped. The filter also selects by provenance: profile_not, profile_revision_below, import_id, dataset_split, region_status, reason, detector, missing_status, missing_provenance.

Combining projects

POST /curation/projects/combine builds a NEW project from 1 to 8 existing ones. It never writes to a source: the target is created building, filled, then made active (a failed run leaves it failed; deleting the target is a complete undo). These routes are global, not under a project. See Combine projects.

Method	Path	Purpose
POST	/curation/projects/combine/preview	Dry run (writes nothing): errors and warnings, suggested_mapping, per-source and target counts, duplicate and conflict counts, bytes to link, and a preview_sha
POST	/curation/projects/combine	Start with the preview body plus expected_preview_sha (409 preview_stale if a source changed, 422 combine_invalid with a report). 202 {job_id, target}
GET	/curation/projects/combine/{job_id}	status, phase, done / total images, report, next_steps; also published as combine.progress events
POST	/curation/projects/combine/{job_id}/cancel	Cancel a running job
POST	/curation/projects/combine/{job_id}/resume	Resume an interrupted or cancelled job from its last finished chunk

A class is its name. class_mapping has one list of map / create / skip / region rows per source, the same rows a dataset import takes; every source class with an included label needs a row (unmapped_class). The target numbers its own classes. Sources are listed in priority order. With dedup: content_hash, an image that is byte-identical in two sources is copied once, from the highest-priority source; boxes of the other copy with IoU of at least dedup_iou and the same target class are merged (the more trusted label wins: human, import, VLM, model), and a box with a different class keeps the priority label and is flagged combine_conflict (the combine_conflict filter on GET /review/{tab} lists them). Clusters are not copied: recluster the combined items. holdout: preserve_union keeps anything that was test in any source as test; recompute picks a new split and warns. A live combine marks its sources and target project_busy.

Items (crops) and labeling

Method	Path	Purpose
GET	/crops	Browse and filter items, including import_id, dataset_split, on_negative_frame, proposed_by_import. n_unembedded is returned only for order=diverse or core_first (null otherwise; use GET /stats/dataset embedding to watch an embed drain)
GET	/crops/{crop_id}	One item
GET	/crops/{crop_id}/thumbnail	Item crop thumbnail
GET	/crops/{crop_id}/image	The clean source image of the item's frame
GET	/crops/{crop_id}/region_thumbnail	One region box's crop; box_id is required
GET	/crops/{crop_id}/context	The item's source image and the other items on it
GET	/crops/{crop_id}/history	Label/edit history
PUT	/crops/{crop_id}/label	Set the class (class_id, label_source human or human_confirmed)
DELETE	/crops/{crop_id}/label	Clear a human label
PUT	/crops/batch_label	Label many; returns updated_ids and per-item conflicts
POST	/crops/{crop_id}/label/undo	Undo the last human class write
POST	/crops/label/undo_batch	Batch undo
POST	/crops/move	Move items to another cluster
POST	/crops/batch_exclude	Exclude from datasets (with a reason)
POST	/crops/batch_unexclude	Restore excluded items
POST	/crops/{crop_id}/discard	Discard an item
POST	/crops/discard_batch	Discard many
POST	/crops/flag_new_class	Flag items as a class the registry lacks
POST	/crops/{crop_id}/review_dismiss	Hide from every review queue
POST	/crops/{crop_id}/review_undismiss	Return an item to the queues
POST	/crops/{crop_id}/vlm_dismiss	Reject the VLM's class suggestion
POST	/crops/{crop_id}/vlm_dismiss/undo	Undo that

The lock rule. A human-set class, a validated imported label, an item in the frozen test holdout, and a human-owned or imported-and-settled region box are locked. No automated writer (worker, VLM batch, re-ingest, reprocess, import over an existing item) changes a locked value. A human label always writes class_source: "human"; a client cannot make a human write look machine-made. An item serves label_locked.

Classes

Class identity is the name. A project's registry assigns its own ids; a dataset, a combined source or a promoted model is matched to it by name.

Method	Path	Purpose
GET	/classes	List with counts, thresholds and reserved hotkeys
POST	/classes	Create (name must match ^[a-z0-9_]+$)
GET	/classes/{class_id}	One class
PUT	/classes/{class_id}	Rename, regroup, bind or clear a hotkey
GET	/classes/{class_id}/crops	Items of a class
POST	/classes/merge	Merge source_id into target_id; ?dry_run=true previews
POST	/classes/{class_id}/deprecate	Retire a class that has no data yet
POST	/classes/{class_id}/restore	Restore it
POST	/classes/sync_to_opensearch	Re-sync the JSON registry into the classes index

The registry is a JSON file in the project's data directory; the first POST /classes creates it with dense ids starting at 0. A hotkey that the project's keymap reserves answers 422 hotkey_reserved; one that another class holds answers 409 hotkey_taken.

Keymap

Each project stores its own keyboard map; a project with no stored map serves the built-in defaults (is_default: true). Writes use optimistic concurrency (expected_revision).

Method	Path	Purpose
GET	/keymap	The action registry, grammar and this project's bindings
PUT	/keymap	Replace the overrides. A clash with a class hotkey is 409 class_hotkey_conflict unless unbind_conflicting_class_hotkeys is set
POST	/keymap/validate	Dry run of a PUT; always 200 with a report
POST	/keymap/reset	Reset all or the listed actions

See Keymaps.

Clusters, scores, selection and search

Method	Path	Purpose
GET	/clusters	Cluster cards: cluster_kind, dominant class, purity (geometry), label_agreement (label share), validated count, sub-clusters, representatives
GET	/clusters/representatives	Representatives only
POST	/clusters/refine/{cluster_id}	Split one cluster into sub-clusters
POST	/clusters/auto_promote	Validate classifier-labelled items in high-purity clusters. Gated on the accuracy audit: 409 audit_required / 409 audit_precision_low (with classes) unless force=true; promote_min_precision defaults to 0.95; dry_run is never gated
POST	/cluster/umap/rebuild	Rebuild the residual reducer
GET	/viz/projection	Cached 2-D coordinates for plotting; never fits inline
POST	/viz/projection/rebuild	Start the projection fit job
GET	/viz/projection/status	Job status. /viz/projection/cancel stops it
POST	/scores/compute	Mistakenness, uniqueness and near-duplicate scores (OP_SCORES_ENABLED)
GET	/scores/status	Status. /scores/coverage reports field coverage and /scores/cancel stops a run
POST	/select/diverse	k-center-greedy diverse sampling (OP_SELECT_DIVERSE_ENABLED)
GET	/select/status	Status of a backgrounded selection. /select/cancel stops it
GET	/search/text	PE-Core text-to-image semantic search (OP_SEMANTIC_SEARCH_ENABLED)

Review queues and holdout

Method	Path	Purpose
GET	/review/tabs	The tab catalog with each tab's label, filters, defaults and empty-state hints
GET	/review/{tab}	One queue. Tabs: all, mismatches, vlm_low_conf, detector_disagreements, outliers, uncertainty, model_disagreements, regions, primary_low_conf, classifier_blind_spots, new_class_proposals, imported
GET	/review/{tab}/locate	Position of one item within a queue
GET	/review/new_class_proposals/summary	VLM-proposed class names, grouped
POST	/review/new_class_proposals/resolve	Accept or reject a proposed name for every item that proposes it
GET	/review/raw_label_clusters	Raw VLM answers that matched no class, clustered
GET	/review/unmatched_terms	The same answers, aggregated by term
POST	/test_holdout/freeze	Freeze a per-class test split before exporting
GET	/test_holdout/stats	Per-class test counts
POST	/audit/start	Draw a stratified accuracy-audit sample of machine-labelled crops (min_per_class, sample_size); 409 audit_no_candidates when none are eligible
GET	/audit/queue	Drawn crops still waiting for a human label (batch_id, page, page_size)
GET	/audit/report	Detector and VLM precision per class with Wilson 95% intervals, the confusion matrix and insufficient_sample flags (min_per_class)
POST	/probe/run	Active-learning probe predictions (feeds model_disagreements and uncertainty)
GET	/probe/status	Probe status. /probe/cancel stops it

Queue filters include on_negative_frame and combine_conflict.

Regions of interest

Regions are meaningful once a region profile is active. An item carries a list of boxes (region_boxes), each with a box_id and its own state, verdict, text, cluster and embedding; see Multi-box regions. A profile's max_regions_per_item caps the list (default 1).

Method	Path	Purpose
GET	/regions	Browse region rows: one row per matching box, or per item. total counts items, total_rows rows. Filters include box_state
GET	/regions/statuses	The status catalog, the per-box state vocabulary and which are human-writable
GET	/regions/vocabulary	Detector, segmenter and verifier vocabulary for the active profile, with limits.max_boxes_per_write
PUT	/crops/{crop_id}/regions	Set a crop's box list (frame source or parent). A box with only a box_id is left untouched; box_id: null adds a box. A stale expected_region_revision is 409 region_conflict; an empty list means "no region visible"
PUT	/crops/batch_regions	Add the same new boxes to many crops
PATCH	/crops/{crop_id}/regions/{box_id}	Patch one box's state or text without touching its siblings
POST	/regions/batch_box_state	Set one state on many {crop_id, box_id} targets
PATCH	/crops/{crop_id}/region_meta	Edit region_status and the rejection reason
POST	/regions/batch_status	Set a human-writable whole-item status on many items
POST	/crops/{crop_id}/region/undo	Undo a region write (restores the box list)
POST	/crops/region/undo_batch	Batch undo
POST	/regions/cluster	Cluster region boxes (background job; a cluster card's size is items, box_count boxes)
GET	/regions/cluster/status	Job status
GET	/regions/clusters	Region cluster cards
POST	/regions/clusters/refine/{cluster_id}	Refine one region cluster
POST	/regions/fp_centroids/build	Build false-positive sub-centroids from false-positive boxes
GET	/regions/fp_centroids/status	Build status
GET	/regions/suspected_false_positives	Non-false-positive boxes ranked by similarity to the false-positive sub-centroids, one row per box
GET	/regions/training_candidates	Curated cohorts for training the next region detector

Per-box states are proposed, accepted, rejected and false_positive. An item's region_status is derived from its boxes (accepted, then false positive, then proposed, then rejected, then empty).

Config store: prompt packs and region profiles

A prompt pack (the VLM's prompts) and a region profile (what to detect and how) are named, versioned documents stored per project. Every save is a new immutable revision. Exactly one pack and one profile are active at a time, and activation is checked against the other settings before it applies.

Method	Path	Purpose
GET	/prompt_packs	List stored packs and templates
POST	/prompt_packs	Create a pack
GET	/prompt_packs/schema	Field specs and placeholders, for a form
POST	/prompt_packs/validate	Validate a draft without saving
POST	/prompt_packs/test	Run one call of a pack, saved or draft, over stored crops and preview each item's write
GET	/prompt_packs/active	The active pack, with what each worker applied
POST	/prompt_packs/active/rollback	Re-activate the previous pack
GET	/prompt_packs/{name}	One pack
PUT	/prompt_packs/{name}	Save a new revision (needs expected_revision)
DELETE	/prompt_packs/{name}	Delete a pack
POST	/prompt_packs/{name}/activate	Activate a revision
POST	/prompt_packs/{name}/clone	Copy a pack, optionally from another project (read-only)
GET	/prompt_packs/{name}/revisions	Revision history; /revisions/{revision} returns one
GET	/region_profiles	List stored profiles and templates
POST	/region_profiles	Create a profile
GET	/region_profiles/schema	Field specs, for a form
POST	/region_profiles/validate	Validate a draft without saving
POST	/region_profiles/validate_segmenter_prompt	Check a segmenter text prompt
POST	/region_profiles/test	Run a profile's detector and segmenter legs over one stored crop, with every candidate and why it was kept or dropped
GET	/region_profiles/active	The active profile
GET	/region_profiles/active/impact	What activating or deactivating would affect, with a suggested_reprocess request
POST	/region_profiles/active/rollback	Re-activate the previous profile
POST	/region_profiles/deactivate	Turn region detection off for the project
GET	/region_profiles/{name}	One profile
PUT	/region_profiles/{name}	Save a new revision
DELETE	/region_profiles/{name}	Delete a profile
POST	/region_profiles/{name}/activate	Activate a revision
POST	/region_profiles/{name}/clone	Copy a profile
GET	/region_profiles/{name}/revisions	Revision history; /revisions/{revision} returns one

| GET | /open_vocab | List open-vocabulary prompt sets and templates | | POST | /open_vocab | Create a set | | GET | /open_vocab/schema | Field specs, ranges and defaults, for a form | | POST | /open_vocab/validate | Validate a draft without saving | | POST | /open_vocab/test | Try one unsaved target on one image; writes nothing | | GET | /open_vocab/active | The active set | | POST | /open_vocab/active/rollback | Re-activate the previous set | | POST | /open_vocab/deactivate | Turn the pass off for the project | | GET | /open_vocab/{name} | One set | | PUT | /open_vocab/{name} | Save a new revision | | DELETE | /open_vocab/{name} | Delete a set | | POST | /open_vocab/{name}/activate | Activate a revision | | POST | /open_vocab/{name}/clone | Copy a set or template, optionally from another project | | GET | /open_vocab/{name}/revisions | Revision history; /revisions/{revision} returns one |

The test routes read stored crops and write nothing. They are limited to 2 concurrent VLM runs and 4 concurrent segmenter calls (429 test_busy), a 60 s bound (504 test_timeout) and a crop cap (422 too_many_crops). See Test on a crop, Prompt packs, Open-vocabulary detection and Settings and config store.

VLM endpoints

The endpoint registry is deployment-wide; which endpoint a project runs is per project. The registry is global, so these routes are not under a project. Keys are never stored or served: an endpoint holds an api_key_ref such as secret:<slug>. See VLM selection.

Method	Path	Purpose
GET	/curation/vlm/endpoints	Every endpoint (the read-only env built-in first) with where each is active
POST	/curation/vlm/endpoints	Register an endpoint
GET	/curation/vlm/endpoints/schema	Form field specs
POST	/curation/vlm/endpoints/validate	Validate a draft; ?probe=true also tests it
GET	/curation/vlm/endpoints/{name}	One endpoint
PUT	/curation/vlm/endpoints/{name}	Save a new revision (needs expected_revision); changes nothing until activated
DELETE	/curation/vlm/endpoints/{name}	Delete; 409 in_use while any project runs it
POST	/curation/vlm/endpoints/{name}/clone	Copy an endpoint
POST	/curation/vlm/endpoints/{name}/probe	Test with synthetic images and record the served model, context length, image cap and JSON-mode support
GET	/curation/vlm/endpoints/{name}/revisions	Revision history
GET	/curation/vlm/catalog	The local model catalog with fits, serving and desired per entry
GET	/curation/vlm/local	The local VLM's served and desired model
POST	/curation/vlm/local/select	Record the desired local model (202). The API never restarts vLLM; the host applies it with openprocessor vlm apply
DELETE	/curation/vlm/local/select	Clear the desired model
GET	/curation/projects/{project}/vlm/endpoints/active	This project's active endpoint (active.name: null = off)
POST	/curation/projects/{project}/vlm/endpoints/{name}/activate	Activate name for the project, with optional revision, force, acknowledge_external
POST	/curation/projects/{project}/vlm/endpoints/active/rollback	Re-activate the previous endpoint
POST	/curation/projects/{project}/vlm/endpoints/deactivate	VLM off for the project

VLM and auto-label pipeline

Every VLM route takes an optional ?vlm=<name> or <name>@<revision> and ?acknowledge_external=; an unknown name is 422 unknown_vlm, and a project with the VLM off is 409 vlm_not_configured.

Method	Path	Purpose
POST	/vlm/label_batch	VLM class suggestions for up to 64 crop_ids
POST	/assist/runs	VLM suggestions for a crop, a selection, a cluster, a sub-cluster or a filter, as a background run (dry run by default)
GET	/assist/runs	The newest assist runs for a cluster or sub-cluster, each with scope_now
GET	/assist/runs/{run_id}	State, progress, result (with the cluster proposal) and scope_now of one assist run
POST	/assist/runs/{run_id}/cancel	Stop a live assist run after its current chunk
POST	/assist/runs/{run_id}/apply	Label the scope's current members with the proposed class, as a human would (dry run by default)
POST	/vlm/verify_region_batch	The same in batches per upstream call
POST	/vlm/region_visible_batch	"Is a region visible at all?" pre-filter
GET	/vlm/policy	The VLM scope policy: scope (all, uncertain, representatives, off), conf_max, per_cluster, max_crops_per_day, sample_frac, revision
PUT	/vlm/policy	Replace it (needs expected_revision; 409 revision_conflict when stale). Limits the VLM worker and the auto_label VLM stage; see VLM selection
POST	/pipeline/auto_label/start	Queue a job for the auto-label worker (run_vlm, vlm_scope, train_clusters, ...). Clusters only unless run_vlm=true
GET	/pipeline/auto_label/status	Current or last job
GET	/pipeline/auto_label/status/{job_id}	One job
POST	/pipeline/auto_label/cancel	Cancel
POST	/pipeline/auto_label	Run the pipeline in-process (synchronous)
GET	/pipeline/events	Pipeline progress events

A VLM reply with no usable answer is omitted, not written as a negative verdict, so the item can be retried.

Export

Method	Path	Purpose
POST	/export/yolo	Multi-class YOLO export (always every class); require_fully_labeled_images, seed, max_images, dedup_threshold
POST	/export/single_class	Single-class or class-subset export with background and hard-negative frames and frozen_test_sha
GET	/export/status	Last export
GET	/export/single_class/status	Last narrowed export for a profile
GET	/export/datasets	Every export (kind, dataset_sha, split counts, is_current)
GET	/export/registry/{artifact}	Download a frozen artifact of the current export (whitelisted names only)
GET	/training_cohorts	Per-class training cohorts (?class_id=)

Training and model comparison

Method	Path	Purpose
POST	/train/preflight	Blocking and warning checks for a spec (dataset, classes, GPUs, trainer reachability)
POST	/train/start	Submit one run (?force=true bypasses blocking checks)
POST	/train/start_campaign	Submit a progressive multi-size campaign
GET	/train/status	Latest run: epochs, eval block, comparison
GET	/train/status/{job_id}	A specific run
GET	/train/runs	Past runs
GET	/train/log/tail/{job_id}	Log tail
GET	/train/manifest/{job_id}	Lineage manifest
GET	/train/artifacts/{job_id}/{name}	A whitelisted run artifact
POST	/train/cancel/{job_id}	Cancel a run. /train/cancel_campaign/{campaign_id} cancels a campaign
GET	/train/profiles	Hyperparameter tables. /train/presets, /train/augmentation_presets and /train/gpus list presets and GPU options
POST	/train/promote/{job_id}	Promote a finished run into Triton, behind a quality gate. The model is named for the project ({slug}__<requested triton_name>, none for default); the final name is returned as triton_name in the result and the job status, and is the name /detect?model_name= and DELETE /models/{model_name} expect. Returns 202 with a promote_id: use this job and poll it for UIs and anything behind a reverse proxy (proxies time out near 120 s). ?wait=true blocks until loaded and warmed (2-3 minutes) and returns 200; it is for direct API scripting with a client timeout of at least 300 s. A dropped wait=true connection abandons the response, not the promote: the load still finishes. The gate-bypass force may be sent in the JSON body or as ?force=true
GET	/train/promote/{job_id}/jobs/{promote_id}	Phase of a background promote: queued, exporting, loading, building, warming, done or failed
POST	/train/reload_promoted	Reload promoted models after a Triton restart
GET	/bakeoff/eval_datasets	Eval sets: exports with a labelled test split and frozen external sets. /bakeoff/trained_models, /bakeoff/baseline_models and /bakeoff/profiles list the rest of what can be compared
POST	/bakeoff/run	Queue a comparison for the evaluator
GET	/bakeoff/runs	Comparison jobs
GET	/bakeoff/status/{job_id}	Job status
GET	/bakeoff/results/{job_id}	Ranked comparison on one dataset (schema v2)
GET	/bakeoff/matrix/{job_id}	The model by dataset matrix

POST /train/start does not merge profile defaults: copy the hyperparameters from GET /train/profiles into the request's own hyperparameters object. See Training, bake-off and promote.

Images

Method	Path	Purpose
GET	/images/serve	Serve a source image by path (confined to configured roots)
GET	/images/root/{alias}	Serve from a named extra root (OP_SOURCE_PATH_ALIASES)
GET	/images/cache/stats	Thumbnail cache stats