--- title: HyperView Visual Safety emoji: 🛡️ colorFrom: yellow colorTo: gray sdk: docker app_port: 7860 pinned: false --- > **Archived live runtime — 2026-09-11.** This container is intentionally paused. [Open the active Static Space](https://spaces.hyper3labs.com/visual-safety/). The source and Git history are retained for reproducibility and can be restored by the owner. # HyperView Visual Safety — review-queue trade-off This demo answers one bounded operations question for a trust-and-safety review-operations lead: > On a curated 120-image Open Images label proxy, is Hyper3-CLIP's one > additional caught proxy-positive worth five additional false reviews and a > six-item larger queue versus CLIP? It is a fixed, auditable image-neighbour ledger—not a production content-policy classifier. Object labels are only proxies, and scores are neighbour-vote fractions, not calibrated safety probabilities. ## Workspace layout The full HyperView shell hosts three panels: 1. **Hyper3-CLIP · Review queue** — native Samples result panel over a prepared Hyper3 queue collection. 2. **CLIP ViT-B/32 · Review queue** — native Samples result panel over a prepared CLIP queue collection. 3. **Review-queue trade-off audit** — compact bottom extension panel with the operational scoreboard, fixed-threshold explanation, prepared queue-slice toggles, and four representative disagreement selectors. Native Samples owns queue browsing, media, and selection. The custom panel does not render a result grid or evidence image. Geometry/Scatter is intentionally omitted: the claim is about queue composition, not embedding layout, and this ledger has no trustworthy layout artifact. Prepared queue slices (full queues, disagreements, false reviews, misses) are materialized once at launch through public `session.ui.show_samples` and switched via documented panel props. There is no runtime mutation of durable workspace data. ## Evidence protocol - Dataset: 120 public Open Images V7 validation images, 60 proxy-positive and 60 proxy-negative. - Models: OpenAI CLIP ViT-B/32 and Hyper3-CLIP v1 image embeddings. - Scoring: leave-one-out vote among seven nearest neighbours in each persisted image-embedding space. - Operating point: queue at least five proxy-positive neighbours, applying the same fixed supermajority rule to both models without fitting a threshold. - Reproducibility: `benchmark.json` contains the complete 120-row prediction ledger, neighbour IDs, metrics, protocol, and content hash. - Audit cases: one candidate gain, one candidate-only false review, one residual miss, and one shared false review, selected from that same ledger. At this operating point CLIP queues 62/120 with 58 TP, 4 FP, 2 FN, and 56 TN. Hyper3-CLIP queues 68/120 with 59 TP, 9 FP, 1 FN, and 51 TN. CLIP is also stronger on this proxy's AUROC and average precision. The visible decision is therefore a real workload/recall trade-off, not a candidate-model victory. ## Run and export From the HyperView repository: ```bash HYPERVIEW_PORT=18248 uv run python \ hyperview-spaces/demos/visual-safety-content-clip-hyper3clip/demo.py ``` Then export the prepared workspace: ```bash uv run hyperview export visual-safety-review-queue-evidence-v3 \ --out dist/landing-demos/visual-safety-v3 ``` The static export retains the full HyperView shell and supports prepared queue browsing, slice switching, case selection, and sample selection. It does not expose inference, threshold tuning, policy actions, model recomputation, or Scatter geometry. ## Data and rights The checked-in `demo_assets/` subset preserves every image's Open Images source URL and CC BY 2.0 license in dataset metadata and the ledger. This bounded demonstration does not cover production prevalence, contextual policy, sexual content, hate, self-harm, jurisdiction, or seller metadata.