Investigation flow
Import evidence
Paste a VoiceIncidentEvidence bundle. The Normalizer splits transcript, tool trace, and workflow definition — no inference.
Competing theories
Claim Tracer and Backend Witness open with conflicting explanations. They challenge, counter, and withdraw until one survives.
Cause + Architecture rooms
Full pipeline runs Cause Room debate, then Architecture Room localization against your workflow graph. Cross-room challenges can force revision.
Audit memo
Completed investigations produce a structured incident finding — customer impact, system reality, failed theories, and fix target.
Access walls
Agents only see evidence their role permits. The Normalizer routes packets — it does not judge.
- Claim Tracer — transcript only
- Backend Witness — tool trace only
- Architecture Room — workflow definition + Cause finding
What Localization needs
Architecture Room cannot localize without your agent's workflow definition — the same graph the production voice agent executes.
- Register the source agent name under investigation.
- Store nodes, tool calls, and confirmation gates.
- Evidence bundle links Layer 1 transcript + Layer 2 execution log.
Cause Room agents
Claim Tracer
L1 · Conversation
What causal class explains customer belief from transcript alone?
Backend Witness
L2 · Execution
What causal class explains tool/API/state evidence?
Causal Judge
Bridge · Causality
After conflict — what bridge class survives cross-domain scrutiny?
Architecture Room agents
Control Flow Investigator
Graph · Execution path
What execution path could emit this behavior?
Policy Investigator
Policy · Prompts
What instruction or policy permits this behavior?
Guard Investigator
Contracts · Guards
What missing guard allows this behavior?
Mechanism Judge
Discovery · Mechanism
What implementation mechanism survives when every surface theory is incomplete?
Demo incidents
Seeded demos: Elena + Klaus. For Vapi, Retell, Bland, and other shapes use Import evidence → Try platform samples on the desk.