Offline incident-triage reference¶
This reference profiles an AI-assisted decision pipeline using synthetic signals:
Use a fresh or empty output directory. The example never loads cloud credentials,
executes a command, changes infrastructure or performs a rollback. Suggested
responses are inert structured values: execute and execution_permitted are
always false. This is not an incident platform, executor, rollback engine or
security sandbox.
Frozen signals and policy comparisons¶
Seven independent fixture outcomes cover normal operation, noisy signals, ambiguous evidence, severe incidents with and without a documented rollback candidate, a warning, and recovery. Inputs are qualitative synthetic flags rather than production hosts, cloud identifiers, secrets or operational runbooks.
| Configuration | Purpose |
|---|---|
baseline |
Local model-style fixtures, repeated severity analysis and irrelevant context |
hybrid |
A fixture-backed anomaly stage plus rules, preserving required verification |
cheap |
A latency-only shortcut that misses sustained-error incidents and over-escalates noise/warnings |
failing |
A declared timeout while producing a response proposal for one severe case |
Missing evidence yields an unknown anomaly/severity and a gather-evidence/escalation proposal. That is a labelled abstention, distinct from a failed execution with no output. A documented rollback candidate is a fixture signal, not proof that a rollback is operationally safe. Real response policies, authorization, safety checks and infrastructure execution remain outside this reference.
Observations and evidence¶
Detection, severity, response and disposition stages use generic workflow/stage metadata with declared dependencies, implementation versions, actual callback timing, outcomes and opaque input/output references. Required secondary checks remain distinct stages. Expected outcomes are not supplied to the decision functions; independent fields-scorer version 1.0 checks the full structured output.
Model-style calls use the shared deterministic fixture backend and its explicitly
synthetic fee/rate-card provenance. json-whitespace-fixture-v1 counts fixture
token units, not real LLM tokens. All traces and reports are marked synthetic.
Recorded local timing and normal replay gates remain visible, but short callback
noise is not presented as a stable real-model performance improvement.
Replay and paired studies retain missed incidents, over-escalation, unknown business states, failed steps and unavailable billing/quality. Lower fixture cost does not override the labelled quality checks. The output includes the frozen corpus, comparison inventory, native replay/quality/study reports and per-run trace, diagnosis and standalone HTML analysis.
Optional semantic investigation¶
Baseline traces investigate repeated severity analysis and irrelevant context
through the typed judgment contract. A deterministic local fixture judge reads
approved summaries by default; main(out, judge=my_backend) accepts another
explicit backend, with any external spend/cancellation owned by that host.
Required verification is marked for retention and cannot become a redundancy recommendation. Other supported cases remain low-confidence hypotheses with unavailable removal savings unless separately attributable. These findings say nothing about whether a response or rollback is safe to execute. Task quality is checked independently against frozen fixture outcomes.
AgentLoop observes and evaluates the application-owned triage policy. Adapting the example to permission-cleared incident data requires organization-owned expectations, implementation/scorer versions and honest provider usage; it does not turn profiling evidence into production authorization.