Odux · Internal Affairs for AI
Reconstruct the truth. Verify the evidence. Govern the spend.
Your agents make thousands of autonomous decisions a day. Internal Affairs for AI turns every one of them into a signed, tamper-evident case file — so you can answer “why did it do that?” and “what did it cost?” with proof, not guesswork.
Tested against every major provider
Three answers, on the record.
Every run becomes a case file you can read, trust, and act on.
Know exactly what happened, in the order it happened.
Every span, tool call, and model choice — normalized into a case file with root-cause and anomaly detection.
- Normalized canonical traces from LangGraph, OTLP, or raw spans
- Root-cause detection flags the exact decision that went wrong
- Anomaly detection surfaces drift across runs
- Architecture reconstruction from observed behavior
The evidence log is tamper-evident — any alteration breaks the chain, visibly.
Each record is hash-chained and signed at ingest. When compliance asks for proof, you have it.
- Hash-chained records signed at ingest time
- Any edit breaks the chain — visible immediately
- Cryptographic proof of log integrity
- Signed evidence log, audit-ready
Spot shadow usage and audit whether a cheaper model would've done the job.
Policy verdicts, cost rollups, and reconciliation against the provider's real bill.
- Dollar-to-decision mapping across every run
- Reconciliation against the actual provider bill
- Shadow spend detection — usage the bill missed
- Policy verdicts — was this decision within budget?
Autonomous agents. Zero accountability.
Agents act in the gap between the prompt and the answer. Logs are scattered, mutable, and incomplete — so when a run goes wrong, there’s no authoritative record of what actually happened.
Can’t answer “why”
That decision was made a week ago. The trace is gone, overwritten, or split across five different tools.
Can’t prove integrity
Anyone with a database login can edit the log. “Trust us” is not a compliance answer.
Can’t govern spend
The bill arrives, but you can’t map a dollar to a decision — or prove a cheaper model would’ve done the job.
Observe without touching the live path.
Traces are shipped after the fact, out-of-band. Your production latency doesn’t move.
Instrument
LangGraph or OTLP spans, dropped in under an hour.
Normalize
Raw spans become one canonical trace.
Redact
Secrets & PII replaced before storage.
Sign
Hash-chained, tamper-evident evidence log.
Investigate & verdict
Case file, root cause, policy, and cost.
Everything you need to hold the line.
Built for teams that ship autonomous agents and can’t afford blind spots.
Multi-tenant workspaces
Isolate every customer with its own keys, roles, and data region.
Role-based access
viewer, investigator, admin — scoped per tenant.
Redact-first ingestion
PII and secrets become [REDACTED] before anything is persisted.
Residency & retention
Pin regions; auto-expire raw traces on a schedule.
Cost reconciliation
Metered usage vs the provider’s actual bill. Shadow spend surfaces.
Anomaly & root cause
Drift detection and architecture reconstruction, built in.
Built for the moment you have to prove it.
When an audit lands or a customer asks hard questions, you need an answer that holds up. That’s the whole point.
Exhibit F · Work with us
Put your agents on the record.
Tell us what you’re running. We’ll have you instrumented and your first case file open in under an hour.
samstern@odux.io