CASE WORKSPACE
Controlled demonstrationModel Registry & Performance
The backend selects models through adapters; the interface never hardcodes model-specific logic. Declared limitations are shown alongside every model.
Registered models
7
Module executions
0
Total compute
0.0 s
Median latency
2.3 s
Queue time
0.4 s
Processing class
Near real-time
< 15 s end-to-end
Model registry
| Model | Task | Version | Provider | I/O | Latency | Status |
|---|---|---|---|---|---|---|
| vision-fa | Frame & face forensic analysis Known limitations: Degrades below 480p; sensitive to heavy re-encoding; calibrated on demo corpus. | 2.4.1-demo | In-house adapter | Image / video frames → Score, confidence, region masks | ≈ 130 ms / frame | DEMO ADAPTER |
| audio-spec | Spectral & vocoder-signature analysis Known limitations: Assumes ≥16 kHz; narrowband telephony reduces recall; music beds cause false positives. | 1.8.0-demo | In-house adapter | PCM audio → Score, confidence, segment list | ≈ 90 ms / s audio | DEMO ADAPTER |
| temporal-flow | Optical-flow & frame-timing consistency Known limitations: Fast camera motion inflates residuals; variable-frame-rate sources need remux. | 1.2.3-demo | In-house adapter | Video → Score, anomaly segments | ≈ 70 ms / frame | DEMO ADAPTER |
| xmodal-align | Audio-visual alignment & viseme match Known limitations: Requires a visible mouth region; multi-speaker scenes unsupported. | 1.1.0-demo | In-house adapter | Video + audio → Alignment score, drift curve | ≈ 110 ms / s media | DEMO ADAPTER |
| meta-inspect | Container & metadata inspection Known limitations: Metadata is trivially forgeable — never authoritative for authenticity. | 0.9.4 | Deterministic parser | Any media container → Structured metadata, inconsistency flags | < 250 ms | PRODUCTION |
| prov-verify | Provenance manifest & hash verification Known limitations: Only as strong as the trust anchor; absence of a manifest is not evidence of manipulation. | 0.6.1 | Deterministic verifier | Asset + manifest → VERIFIED / UNVERIFIED / INCONSISTENT / UNAVAILABLE | < 300 ms | PRODUCTION |
| forensic-reasoner | Evidence-grounded narrative reasoning Known limitations: Never sees raw media; constrained to cite supplied evidence IDs; not a classifier. | 1.0.0-adapter | Server-side LLM adapter | Normalised evidence objects only → Assessment, reasoning, uncertainty, review actions | 1.5–4 s | ADAPTER READY |
Per-module processing time
Aggregated across demonstration runs
Vision Detector
0.00 s
Audio Detector
0.00 s
Temporal Detector
0.00 s
Metadata Analyzer
0.00 s
Provenance Verifier
0.00 s
Cross-Modal Analyzer
0.00 s
Measurement caveats
- — Scores shown across the console originate from demonstration adapters, not from validated production models.
- — No accuracy, precision or recall figures are published here, because none have been measured on an independent benchmark for this build.
- — Latency figures reflect the demonstration environment and will differ under real media loads.