Faogf: Auditable Evidence Provenance ForAutonomous Financial Ai Agents On Opentelemetry
Autonomous AI agents now read regulatoryfilings, compute figures and distribute reports inside financial firms, but their telemetry cannot tell an auditor which filing versions a decision rested on, under which authority and policy version an action ran, or whether untrusted content was in the model’s context. We present FAOGF (Financial Agent Observability and Governance Framework), an OpenTelemetry-compatible audit layer. Its core architectural contribution is context-bound evidence provenance recorded on two paths: a gateway binds every modelcall and action to the version-exact, content-hashed evidence actually in the model’s context, and writes that binding synchronously to a hash-chained evidence store before exporting standard spans, so audit completeness no longer depends on what the tracing pipeline keeps. We evaluate nine audit queries. In 15 controlled experiments with a scripted agent (200 workflows × 5 seeds, and a real OpenTelemetry Collector on two machines), all nine queries were evaluated: FAOGF answered the seven answerability queries (Q1–Q7) exactly in every configuration, detected 88–97% of spans lost to export-queue overflow (Q8) and 100% of audit attributes stripped by a Collector processor (Q9). Standard OpenTelemetry tracing recovered only 9–12% of protected actions at 10% head sampling and was silently stripped by a common Collector configuration; full prompt logging could not answer authority or policy questions. In a live study on SEC EDGAR filings with gpt-4o-mini and gpt-4o on public Finance Agent Benchmark questions, wescored the two queries the live workload exercises, supporting filings (Q3) and untrusted-context exposure (Q7). FAOGF was exact on both, standard tracing misattributed 4–18% of supporting filings, and full prompt logging used 7–10× more storage. Neither live model followed any of the 78 injected instructions, so enforcement was evaluated on a simulated susceptible agent. There, escalating only the sends requested with untrusted content in context blocked every injected send with 37% fewer human approvals than requiring approval for every send. FAOGF adds about 0.2 ms median overhead per call.
Authors
- Sowmith Javvaji (ORCID: https://orcid.org/0009-0001-7473-8417)
Publication Details
- Journal
- International journal of data science and machine learning.
- Published
- 2026-11-16
- DOI
- https://doi.org/10.55640/ijdsml-06-02-16
- Primary Topic
- Software System Performance and Reliability
- Type
- article
- Field-Weighted Citation Impact
- 0.00