Model catalogue
Reference data on someone else's cadence · model registry. On a deployment this row is a control; the demo shows the row with the reading its source supports.
Evidence
- Running
- 16 registered, 16 enabled (the models table, through the Control panel)
- Built
- platform default claude-sonnet-5 (core/assistant/config.py)
- Latest known
- no registered model_id retired; provider lists checked 3 days ago
- Source
- www-demo/app/emr_profiles.php models_all()
- Checked
- 3 days ago
- Cadence
- every 14 days
- Mode
- direct
- State
- current
- Backup unit
- The registry row's previous state.
- Recovery
- Restore the row.
The five steps
Plan, back up, apply, verify, recover — in direct mode. On a deployment each step runs, is verified and lands on the audit trail; every component on the demo is current, so every step reads done.
- ✓ done Plan The provider deprecation lists in the manifest are compared with every registered model_id; the plan names the IDs to retire and the slots they serve; the admin confirms.
- ✓ done Back up The updater records each affected registry row as it stands.
- ✓ done Apply The updater marks the retired ID disabled in the registry, audited; choosing the replacement stays the Control panel's register, enable, activate flow.
- ✓ done Verify Every slot still resolves to an enabled model or shows the degraded banner honestly; the healthcheck is green.
- ✓ done Recover Restore the row's previous state and verify again.
Every row behind this reading (16)
| Model | Kind | Provider | Model ID | Status |
|---|---|---|---|---|
| Ambient listener — faster-whisper | custom | Self-hosted (MIT) | small.en int8 | enabled |
| Claude Sonnet 5 | foundation | Anthropic API | claude-sonnet-5 | enabled |
| Coding integrity model | classifier | built-in | coding-integrity | enabled |
| Denial risk model | predictive | built-in | denial-risk | enabled |
| Imaging read assistant | vision | Anthropic (your BAA) | claude-sonnet-5 | enabled |
| Inbox triage ranker | deterministic | Built-in | urgency-rank v1 | enabled |
| Ingest QA reviewer | llm.analytic | organization-managed | byom-ingestqa-reviewer | enabled |
| No-show risk model | predictive | built-in | noshow-grad-boost | enabled |
| Patient instructions writer | llm.generative | organization-managed | byom-instructions-writer | enabled |
| PHI RAG | retrieval | built-in | enabled | |
| Prior auth evidence assembler | retrieval | built-in | pa-evidence | enabled |
| ROI requirements validator | classifier | built-in | roi-validator | enabled |
| Scheduling optimizer | optimization | built-in | sched-opt | enabled |
| Scripted fallback | deterministic | built-in | enabled | |
| Sensitivity classifier | classifier | built-in | sens-classify | enabled |
| Terminology mapper | mapper | built-in | term-map | enabled |
History
Applies and recoveries of this row, from the update journal. On a deployment every apply and every recovery lands here as it completes.
No apply or recovery recorded for this row. The demo's journal holds one completed update, of Running image.
Acknowledge: keep as is until <date> because <reason> — on a deployment this line is a form, audited as system.component_acknowledged; the demo shows the words.