Why does identity come before age?

Because a video cuts. Box tracking followed a cut between an interviewer and a child and reported one person whose age was the median of three people. Every individual estimate was right and the identity was wrong, which is worse because it looks correct. The template groups by face across cuts first, then ages the person, and reports how many looks the estimate rests on.

What does zero missed minors cost?

On single-frame evidence, a cutoff of 18 catches 82% of minors and sends 4.8% of adults to review. A cutoff of 22 catches 94% and sends 23%. A cutoff of 30 catches 100% and sends 94.5% of adult footage to review. The number is yours to choose, and the page shows the curve so you choose it with the price attached.

Does face size limit this?

Less than expected. Shrinking labelled faces from 200px to 32px moved recall from 82.2% to 86.2%. Detection is the real limit: a face the detector never fires on is never aged, which is why anything undetected is routed to review rather than assumed adult.

How do I know the feedback loop is working and not just collecting?

Because the loop ends in an evaluation, not in a collection. Annotations and interactions are turned into a held-out set and a threshold change is measured as precision and recall moving on it. Captured signals that do not move a number are not a working loop.

Can I use your reference sets instead of supplying my own?

You can start from the marketplace Celebrity Face Scanner and Logo detection indexes to see the shape of the results. For a production decision the reference set should be yours, because it decides which marks and which people count, and it is the thing that decays if nobody owns it.

Safety and compliance

Template v1.0 · updated 2026-09-03

Video Moderation for Training Sets

Pre-filtering at the ingest stage, before a single frame reaches training. Reduce raw footage to the subset you can defend training on: every excluded clip carries its reason, every uncertain one goes to a person, and the thresholds are yours.

A cleared training set, filtered at ingest, with an audit trail for every clip that did not make it.

Deploy this template

Teams preparing video corpora for model training who have to prove what was excluded and why. The problem shows up at ingest, where the corpus is largest and the cost of getting it wrong compounds through every downstream run.

82%

of minors caught at a cutoff of 18

on 650 labelled faces, 320 minors and 330 adults, single frame; 4.8% of adults sent to review

100%

recall is reachable, and the page says what it costs

a cutoff of 30 catches every minor and sends 94.5% of adult footage to review; the whole curve is in the evaluation tab

0.70

face detector gate

cleared six phantom minors the detector found in non-faces while keeping 85% of genuine faces

39%

of detections routed to review as unaged

no usable face, under 24px, or detector confidence under 0.70. None of them is passed as adult.

What it looks like

Frame in, decisions fire, a verdict lands, a reviewer's call feeds back.

Simulated walkthrough · illustrative frames, scripted decisions

A close-up interview frame with one face in shot

faces -> age-estimation -> person-verdictsStill: Pexels / Ron Lachframe 1 of 4

Decision path

face detected 0.93
grouped across 3 looks
age estimate: borderline
route: review

within 6 years of the cutoff, and 3 looks

Running tally

cleared

review

excluded

Reviewer feedback

When the frames are done, a reviewer's call on the borderline case feeds back into the thresholds.

What goes in

source-video
Your footage, read from your own object storage. Nothing is copied out.
brand-mark-refs
Logo and mark variants you supply, one file per variant, with a hard-fail or contextual rule each.
person-refs
Public-figure reference imagery you supply and hold the rights to. Several poses per person.

Where the decisions fire

Illustrative frames; boxes show the decision path, not live model output

brand mark · match 0.88 · rule: contextualcleared

keyframes -> brand-mark-match · A keyframe embedding is matched against your brand-mark reference index. The rule attached to that mark decides hard fail or contextual.Still: Pexels / Mathias Reding

Two people sparring in a gym — keyframes -> violence-context · Violence is scored in context. Sparring, practice and play score staged; the real thing scores real. The margin between them is what you set.Still: Pexels / Franco Monsalvo

A reporter beside a van carrying a BREAKING NEWS graphic — onscreen-text · Words shown on screen and never spoken are read by the on-screen text pass, static overlays and scrolling bands alike, and land in their own searchable field. Nothing in the audio says BREAKING NEWS.Still: Pexels / cottonbro studio

What comes out

segment-verdicts
One row per segment: decision, the reason, and counts of people, minors, unaged and borderline faces.
review-queue
Everything routed to a human, ranked, with the decision chain attached to each item.
cleared-set-export
The segments you can train on, as a filterable set ready to hand to a training job.

How the namespace is wired

3 buckets, 8 collections, 2 clean views, 6 retrievers. The diagram generates the manifest below; they cannot drift apart.

SourceBucketCollectionClean viewRetrieverClick a node to inspect it

Reward signals

How reviewer decisions move the thresholds

Thresholds at ingest drift as the corpus changes. The reviewers working the queue are the ones who see where a threshold is wrong first, so this template routes their decisions back into the model that set it.

DocsAnnotations Interactions Reward signals Auto-tune Build a feedback loop Generate evaluations from interactions

Reviewers do not only clear a queue. Every call they make is a labelled example, and the template is built so those labels move the thresholds rather than sit in a spreadsheet.

Explicit signals

A reviewer marks a verdict right or wrong and attaches what they saw. That is an annotation, written with POST /v1/annotations, versioned, and stored in the namespace's own mxp_document_annotations collection.

Implicit signals

What a reviewer opens, skips, or acts on is an interaction, written with POST /v1/retrievers/interactions into mxp_retriever_interactions. Nobody fills in a form for these.

Where they land

mxp_document_annotationsmxp_retriever_interactionsmxp_retriever_evaluations

System collections in your namespace, on the same vector store as the rest of the template. They are yours to query.

How the loop closes

POST /v1/retrievers/{id}/evaluations/generate-from-interactions turns those signals into an evaluation set, so a threshold change is judged by precision and recall moving on held-out data rather than by whether signals were captured. moderation-search is configured with learned fusion in shadow mode: it learns from the signals and reports what it would have ranked, and you flip it on with a retriever PATCH when the numbers justify it.

DocsNamespaces Pipeline configuration Retrievers

One file spins up the namespace. Generated from the diagram above. Also served at /templates/video-moderation.namespace.yaml.

# video-moderation: one manifest spins up the namespace.
# Platform manifest schema (GET /v1/discovery/schema). Validate with POST /v1/manifest/validate,
# apply with POST /v1/manifest/apply or the Deploy button. Wiring comes from the flow diagram:
# edges are bucket -> collection sources, collection -> retriever scope, retriever -> view.
version: '1.0'
metadata:
  name: video-moderation
  description: "Namespace template video-moderation. Generated from the flow diagram on mixpeek.com/templates/video-moderation."
namespaces:
  - name: video-moderation
    description: "Everything below lives in this namespace."
    feature_extractors:
      - name: multimodal_extractor
        version: v2
      - name: image_extractor
        version: v1
      - name: face_identity_extractor
        version: v1
      - name: scrolling_text_extractor
        version: v1

# Data sources. A storage connection carries credentials, so it is created in Studio
# (or POST /v1/organizations/storage-connections) and synced into the bucket named here.
#   source-video: s3, continuous, your footage, in your account -> bucket source-video
#   brand-marks: manual, one-shot, customer-supplied logo and mark variants -> bucket brand-mark-refs
#   person-references: manual, one-shot, customer-supplied public-figure reference imagery -> bucket person-refs
buckets:
  - name: source-video
    namespace: video-moderation
    description: "Fed by source-video (s3, continuous)."
    schema:
      properties:
        content:
          type: video
  - name: brand-mark-refs
    namespace: video-moderation
    description: "Fed by brand-marks (manual, one-shot)."
    schema:
      properties:
        content:
          type: image
  - name: person-refs
    namespace: video-moderation
    description: "Fed by person-references (manual, one-shot)."
    schema:
      properties:
        content:
          type: image
collections:
  - name: video-segments
    namespace: video-moderation
    description: "multimodal_extractor@v2 over bucket source-video. Feeds faces, keyframes, onscreen-text, roll-up-by-segment."
    source:
      type: bucket
      bucket: source-video
    feature_extractor:
      name: multimodal_extractor
      version: v2
      input_mappings:
        video: content
      field_passthrough:
        - source_path: file_location
        - source_path: segment_id
        - source_path: start_time
        - source_path: end_time
    enabled: true
  - name: brand-mark-index
    namespace: video-moderation
    description: "image_extractor@v1 over bucket brand-mark-refs. Feeds brand-mark-match."
    source:
      type: bucket
      bucket: brand-mark-refs
    feature_extractor:
      name: image_extractor
      version: v1
      input_mappings:
        image: content
      field_passthrough:
        - source_path: mark_name
        - source_path: rule
    enabled: true
  - name: person-ref-index
    namespace: video-moderation
    description: "face_identity_extractor@v1 over bucket person-refs. Feeds public-figure-match."
    source:
      type: bucket
      bucket: person-refs
    feature_extractor:
      name: face_identity_extractor
      version: v1
      input_mappings:
        image: content
      field_passthrough:
        - source_path: person_name
        - source_path: rights_note
    enabled: true
  - name: faces
    namespace: video-moderation
    description: "face_identity_extractor@v1 over collection video-segments. Feeds age-estimation, person-verdicts."
    source:
      type: collection
      collection: video-segments
    feature_extractor:
      name: face_identity_extractor
      version: v1
      input_mappings:
        video: content
      field_passthrough:
        - source_path: segment_id
        - source_path: bbox
        - source_path: detection_score
        - source_path: quality_score
    enabled: true
  - name: keyframes
    namespace: video-moderation
    description: "image_extractor@v1 over collection video-segments. Feeds violence-context."
    source:
      type: collection
      collection: video-segments
    feature_extractor:
      name: image_extractor
      version: v1
      input_mappings:
        image: content
      field_passthrough:
        - source_path: segment_id
        - source_path: frame_ts
    enabled: true
  - name: onscreen-text
    namespace: video-moderation
    description: "scrolling_text_extractor@v1 over collection video-segments. Feeds roll-up-by-segment."
    source:
      type: collection
      collection: video-segments
    feature_extractor:
      name: scrolling_text_extractor
      version: v1
      input_mappings:
        video: content
      field_passthrough:
        - source_path: segment_id
    enabled: true
  - name: person-verdicts
    namespace: video-moderation
    description: "Derived from faces, age-estimation. One document per person_id. Writes back age_median, aged, n_looks. Declared as a passthrough collection because the manifest schema has no resource for a retriever-materialised collection."
    source:
      type: collection
      collection: faces
    feature_extractor:
      name: passthrough_extractor
      version: v1
      field_passthrough: []
    enabled: true
  - name: segment-verdicts
    namespace: video-moderation
    description: "Materialised by retriever roll-up-by-segment (group_by segment_id) over person-verdicts, violence-context, onscreen-text, video-segments. One document per segment_id. Writes back decision, reason, n_person, n_minor, n_unaged, n_borderline. Declared as a passthrough collection because the manifest schema has no resource for a retriever-materialised collection."
    source:
      type: collection
      collection: person-verdicts
    feature_extractor:
      name: passthrough_extractor
      version: v1
      field_passthrough: []
    enabled: true

# Template-provided extractors. These collections run on age_estimator and context_scorer,
# plugins that ship with this template rather than hosted extractors, so apply would fail on
# a namespace that does not have them yet. Install the extractors, then move this block above.
#   - name: age-estimation
#     namespace: video-moderation
#     description: "age_estimator@v1 over collection faces. Feeds person-verdicts. age_estimator is a template-provided extractor (a plugin that ships with this template, not a hosted one): install it, then apply this block."
#     source:
#       type: collection
#       collection: faces
#     feature_extractor:
#       name: age_estimator
#       version: v1
#       input_mappings:
#         face: content
#       field_passthrough:
#         - source_path: person_id
#         - source_path: age_median
#         - source_path: aged
#         - source_path: n_looks
#     enabled: true
#   - name: violence-context
#     namespace: video-moderation
#     description: "context_scorer@v1 over collection keyframes. Feeds roll-up-by-segment. context_scorer is a template-provided extractor (a plugin that ships with this template, not a hosted one): install it, then apply this block."
#     source:
#       type: collection
#       collection: keyframes
#     feature_extractor:
#       name: context_scorer
#       version: v1
#       input_mappings:
#         image: content
#       field_passthrough:
#         - source_path: segment_id
#         - source_path: staged_score
#         - source_path: real_score
#         - source_path: margin
#     enabled: true
retrievers:
  - name: brand-mark-match
    namespace: video-moderation
    description: "Searches brand-mark-index across 1 feature index."
    collections:
      - brand-mark-index
    input_schema:
      query:
        type: text
        required: true
        description: "What to look for; searched across every index below"
    stages:
      - stage_name: search
        stage_id: feature_search
        parameters:
          searches:
            - feature_uri: "mixpeek://image_extractor@v1/google_siglip_base_v1"
              query:
                input_mode: text
                value: "{{INPUT.query}}"
              top_k: 5
          fusion: rrf
          final_top_k: 5
    tags:
      - template:video-moderation
  - name: public-figure-match
    namespace: video-moderation
    description: "Searches person-ref-index across 1 feature index."
    collections:
      - person-ref-index
    input_schema:
      query:
        type: text
        required: true
        description: "What to look for; searched across every index below"
    stages:
      - stage_name: search
        stage_id: feature_search
        parameters:
          searches:
            - feature_uri: "mixpeek://face_identity_extractor@v1/insightface__arcface"
              query:
                input_mode: text
                value: "{{INPUT.query}}"
              top_k: 5
          fusion: rrf
          final_top_k: 5
    tags:
      - template:video-moderation
  - name: roll-up-by-segment
    namespace: video-moderation
    description: "Pipeline retriever over person-verdicts, onscreen-text, video-segments. Groups documents by segment_id. Materialises into segment-verdicts."
    collections:
      - person-verdicts
      - onscreen-text
      - video-segments
      # - violence-context  (add once context_scorer is installed and its block above is applied)
    input_schema:
      limit:
        type: text
        required: false
        description: "Optional cap on documents grouped per run"
    stages:
      - stage_name: group
        stage_id: group_by
        parameters:
          group_by_field: segment_id
    tags:
      - template:video-moderation
  - name: review-queue
    namespace: video-moderation
    description: "Searches segment-verdicts across 1 feature index."
    collections:
      - segment-verdicts
    input_schema:
      query:
        type: text
        required: true
        description: "What to look for; searched across every index below"
    stages:
      - stage_name: search
        stage_id: feature_search
        parameters:
          searches:
            - feature_uri: "mixpeek://multimodal_extractor@v2/gemini-embedding-2"
              query:
                input_mode: text
                value: "{{INPUT.query}}"
              top_k: 100
          fusion: rrf
          final_top_k: 100
    tags:
      - template:video-moderation
  - name: moderation-search
    namespace: video-moderation
    description: "Searches segment-verdicts across 3 feature indexes with learned fusion."
    collections:
      - segment-verdicts
    input_schema:
      query:
        type: text
        required: true
        description: "What to look for; searched across every index below"
      session_id:
        type: text
        required: false
        description: "Session identifier; the learned fusion uses it as personal context"
    stages:
      - stage_name: search
        stage_id: feature_search
        parameters:
          searches:
            - feature_uri: "mixpeek://multimodal_extractor@v2/gemini-embedding-2"
              query:
                input_mode: text
                value: "{{INPUT.query}}"
              top_k: 25
            - feature_uri: "mixpeek://multimodal_extractor@v2/multilingual_e5_large_instruct_v1"
              query:
                input_mode: text
                value: "{{INPUT.query}}"
              top_k: 25
            - feature_uri: "mixpeek://multimodal_extractor@v2/multilingual_e5_large_instruct_ocr_v1"
              query:
                input_mode: text
                value: "{{INPUT.query}}"
              top_k: 25
          fusion: learned
          final_top_k: 25
          # Shadow first: weights are learned and logged, static fusion is served, until the
          # signals justify a rollout. Flip rollout_pct through the retriever update path.
          learning_config:
            shadow_mode: true
            rollout_pct: 0.0
            context_features:
              - INPUT.session_id
    tags:
      - template:video-moderation
  - name: cleared-set-export
    namespace: video-moderation
    description: "Searches segment-verdicts across 1 feature index."
    collections:
      - segment-verdicts
    input_schema:
      query:
        type: text
        required: true
        description: "What to look for; searched across every index below"
    stages:
      - stage_name: search
        stage_id: feature_search
        parameters:
          searches:
            - feature_uri: "mixpeek://multimodal_extractor@v2/gemini-embedding-2"
              query:
                input_mode: text
                value: "{{INPUT.query}}"
              top_k: 1000
          fusion: rrf
          final_top_k: 1000
    tags:
      - template:video-moderation