Evidence infrastructure for AI in finance
AI agents now approve refunds, flag transactions, and decide credit. SwornTrace seals every decision they make into a tamper-evident record, verifiable by auditors and regulators without trusting you.
Runs in your infrastructure. Records never leave your perimeter.
$ sworntrace verify --ledger decisions.db chain ok 1,284 records re-hashed, intact signatures ok verified against public key anchors ok external S3 + timestamp match $ sworntrace replay dec_4a29b83a refund-agent decided 'rejected' because 3 of 4 policy rules failed: refund-policy/section-1 · section-2 · section-4 confidence 90% · sealed #6 · retained to 2032
The problem
Your engineering logs were built for debugging. They miss the reasoning behind every decision, and because you can edit them, they prove nothing to a skeptical outsider. Three moments expose the gap.
Regulators have fined firms over $2B for business happening in unrecorded channels. An AI agent is the newest unrecorded channel, and EU AI Act logging obligations for high-risk systems are now enforceable.
GDPR Article 22 gives customers the right to an explanation of automated decisions. When the lawyer asks for proof of what your AI did, a debugging trace won't settle it. A sealed record will.
Annual audits ask for evidence, not architecture diagrams. Decision records with retention, integrity, and clause mapping built in turn a week of scrambling into a single export.
How it works
A lightweight SDK sits inside your agent. Everything else follows automatically.
Inputs, retrieved policies, rules evaluated, tool calls, alternatives considered, outcome, confidence. All recorded at the moment of decision.
Records join a hash chain, are signed, and anchored to storage you don't control. Any alteration, even by you, is detectable.
Each record is evaluated against specific clauses (SEC 17a-4, GDPR 22, MiFID II, EU AI Act 12), reporting evidence or gaps per decision.
Auditors replay decisions and verify integrity with an open verifier: no access to your systems, no trust in your software required.
The SDK
from sworntrace import SwornTrace st = SwornTrace(agent="refund-agent") with st.decision(action="approve_refund", risk_level="high") as d: d.add_input(request) d.add_context(policy, source="kb://refund-policy", role="policy") d.check_policy("refund-policy/section-1", passed=True) d.add_alternative("escalate", rejected_because="within limits") d.set_retention_years(6) d.set_outcome("approved", confidence=0.92) # sealed, chained, anchored, verifiable by anyone
Where SwornTrace sits
Observability tools answer "is my agent working?" for engineers. Governance platforms answer "do we have policies?" for boards. SwornTrace answers the question the other two can't: prove what this decision was, and why.
| Observability | GRC platforms | SwornTrace | |
|---|---|---|---|
| Built for | Engineers | Compliance programs | Auditors & regulators |
| Unit of record | Traces & spans | Policies & questionnaires | Complete decisions |
| Captures the reasoning | Partially | No | Yes |
| Tamper-evident | No, editable | No | Yes, sealed & anchored |
| Clause-level mapping | No | Program-level only | Per decision |
| Independent verification | No | No | Offline, open verifier |
Early access
We're running shadow-mode pilots with a small number of financial firms: our SDK against one live agent workflow, in your environment, producing sealed records and a compliance report from your own traffic.
Talk to us