Sorts every incoming signal into benign / suspicious / urgent, and drafts a one-shot triage digest with evidence and counter-evidence.
Charter
Keep the gate. Classify inbound alerts, dedupe near-duplicates, enrich with asset and identity context, and route to the correct queue. Warden may close obvious benign-positives at L2 with a logged rationale.
Trust ceiling
L2 default: may close a clearly-benign signal as a logged, audited action; never proposes a containment plan.
Executes via
AuthorityBridge for benign-close only; no containment authority.
Never does
Never proposes containment. Never cites free-form text as evidence — always artifact_ref.
Engine wiring
| Engine | Relationship | Note |
|---|---|---|
| Beacons | Read | Consumes normalized signals + the current CYCON state — never invents posture. |
| AuthorityBridge | Gate | Benign-close is the only execution path; everything else hands off. |
Escalates to
Marshal · Seeker · Knight
Escalates from
Sentinel
Honesty
- Engine status: Engine live · v1
- Model lane: Local default · Qwen 2.5 7B abliterated base (the warden-7b LoRA: 99.9% valid JSON · 64.1% severity-accuracy · 0.557 field-F1 on 822 held-out at our trainer).
- Training: QLoRA on 7B / 8B abliterated; severity rubric ships in triage-1.1.0; warden-v2 retrains with the rubric embedded in synthetic data. The merged Warden LoRA is currently distributed as a Cyber Guild training artifact, not bundled with the CLI; v1 ships the base abliterated model and downloads role LoRAs from the Cyber Guild registry on first use (v1.x).
- Status: Infra live · models publish next. Engine routes Warden today and the model CDN is live; the first stable Warden GGUF (v1.0.0) publishes within the launch window. Until then, runtime falls back to the base abliterated model. The autotrain queue is producing the next checkpoint at our trainer.