Optional AI features

No model is required for policy evaluation or the standard demonstrations.

Incident commentary

Set QINZORA_OLLAMA_URL and QINZORA_OLLAMA_MODEL, then call manager POST /v1/incidents/{id}/explain. Canonical explanation remains; optional text appears separately in model_explanation. Known categories, fixed meanings and counts are sent, not raw event identities or payloads. Inference timeout is 60 seconds, connection timeout 2 seconds. Commentary can be wrong and cannot activate policies.

Bounded policy administration

qinzora.administration lets a model propose a typed plan. A separate executor validates authority and performs actions. Schema compliance alone does not authorize execution. The model receives no database connection, executable functions or operating-system access.

Copy examples/management.json, configuring an existing database, provisioned account and installed model. Keep provider secrets in environment variables via api_key_env, not JSON or Git.

qinzora manage --config FILE.json
qinzora manage --config FILE.json --plan RETURNED_ID --execute

The example defaults to allow_enforce: false. Enforcement requires explicit service/route scopes, TTL and rate bounds. Plans contain at most five actions: stage_policy, enforce_policy, revoke_policy; experimental continuity actions require explicit profile grants. Empty plans are valid abstentions.

Before enforcement, the executor replays current history. Defaults require legitimate/abusive evidence, affected abuse, no affected unknown labels and zero observed false positives. Labels come from the host, not the model. Plans normally expire after five minutes; scope/policy/authority changes can invalidate them. Database actions commit atomically; repeated completed execution returns stored results. Rollback does not cover external operations.

qinzora watch --config FILE.json --execute persists checkpoints and a per-account lease, avoids repeated inference on unchanged evidence and backs off on model failure. It does not provide distributed HA or guarantee renewal of expired policies.

Providers and data

StructuredModel defines async generate(system, context, result_type). HTTPStructuredModel supports ollama and chat_completions with JSON Schema. Chat URLs include the provider prefix (e.g. /v1). Endpoints/models are trusted operator configuration, never model-selected. Invalid, incomplete or oversized responses raise ModelUnavailable. Response envelopes are bounded to 256 KiB; no implicit retries/fallback occur.

Administration sends configuration, authority and aggregate evidence, not raw identities, tenant values, bodies or queries. Service/route/policy names and IDs still disclose application metadata. Custom contexts are not automatically scrubbed. Review the provider destination and use TLS remotely.

Verification

Controlled provider tests use simulated responses and do not establish real inference quality. For a real local model with temporary storage and real HTTP:

python scripts/administration_smoke.py --model INSTALLED_MODEL
python scripts/administration_smoke.py --model INSTALLED_MODEL --watch

Abstention/rejected execution is not successful activation. Results depend on provider/model/environment. Real local Ollama inference completed the continuity scenario in three recorded post-clarification trials. See the evidence. The default demo planner remains explicitly scripted; model commentary may still misinterpret evidence.

The independent continuity_profile_ids grant permits continuity actions even when allow_stage and allow_enforce are false; those flags govern policy staging/enforcement only. A candidate profile does not provide live fallback until activated. Rehearsal eligibility describes cached-source evidence, while dependency health describes the primary dependency. The model is told this distinction; the executor continues checking profile IDs, scope, TTL and rehearsal freshness.

The monitor preserves the observation actually sent to the model, including events arriving during inference. Equivalent active policies are not duplicated; eligible candidate/expired policy IDs can be reused with their former configuration audited and replay recalculated. Stopping the monitor cannot undo committed external work. A custom repository must provide account scoping, serialized writes and atomic commit/rollback; the controller does not require an agent framework.