Pramaan re-checks recorded AI workflow runs against your declared business controls and issues a signed, tamper-evident verdict — that any auditor can independently re-perform.
OTel captures what the AI did. emit() adds who approved what. State probes check the system of record. The same run + same pack always produces the identical verdict.
OTel traces capture what the AI did — model calls, tool invocations, retrieved documents. emit() SDKs record the human layer: who approved, who escalated, what business decisions were made. State probes check against your system of record.
The same run against the same versioned control pack always produces an identical result — byte for byte. This is not monitoring. An auditor can re-run the check independently, without access to your production systems, and reach the same verdict.
Ed25519 signatures in in-toto/DSSE format. Each report is linked to the previous one via hash chain. Tampering at any point in the chain is visible. Export for audit, committee review, or CI/CD gate.
Designed to close the gap between what an AI workflow says it does and what an auditor can independently verify it did.
Versioned control packs for lending, KYC, collections, support and fraud triage. 17 rule types covering every key BFSI obligation. Packs live in version control alongside your models — every change is diffed, reviewed, and approved before it affects a verdict.
Lending pack complete and ready for validation. KYC, collections, support and fraud packs are drafts — awaiting bank review before production use.
The same run against the same pack always produces the identical result, byte for byte. An auditor can re-run the check without access to your systems.
Ed25519 signatures in in-toto/DSSE format. Each report is linked to the previous one. Tampering is visible at any point in the chain.
Four possible outcomes. No ambiguity. Missing evidence is never counted as a pass. This matters when preparing for regulatory examination.
Indeterminate = evidence present but inconclusive. Missing evidence never silently passes.
Pass, hold or block a release automatically via exit codes. Control checks run before deployment — not after the model is in production.
Inject violations — the same person acting as maker and checker — to prove your controls actually catch them. Run mutation tests in CI to prevent control regressions.
A control that passes on clean data but misses injected violations is not a control. Mutation testing is how you know the difference.
Re-run recorded workflows after a model update or prompt change. Receive a signed comparison report showing exactly which verdicts changed and why.
Current replay runs against a local simulator. Live LLM provider replay is on the roadmap.
Each report is linked to the previous via SHA-256. Modify any block and every subsequent link breaks. Click "Tamper block 2" to see it.
Every control, every approval and every conclusion is made by the right person — and the audit trail proves it.
Propose changes. Submit runs for verification. "My work" queue shows open items and pending attestations.
Pack changes require diff-first review. Nobody approves their own work. Oversee validation engagements and sign conclusions.
Re-perform any run without access to production systems. Read-only view of all verdicts, hash-chain intact.
Define population window. Record findings. Track remediation. Run re-tests. Issue signed conclusion. Each engagement produces a committee-ready PDF with PAdES legal digital signatures via DSC certificates.
When someone leaves, access ends immediately. Every verdict they signed continues to verify — the hash chain is intact, the signature remains valid. The audit trail outlives the employee.
OTel captures what the AI did. emit() adds who approved what. Both feed the same deterministic verification engine.
Records approvals, escalations and business decisions that OTel cannot see
"Tested with Langfuse 3 and 4 (self-hosted)" means exactly that. We have not tested against Langfuse Cloud or claimed Langfuse certification.
The verification engine makes zero outbound calls during a run. Verdicts cannot be influenced by external state.
Customer-held salt. PII is removed before reaching the verification engine. You control the redaction key.
Runs, packs and verdicts are isolated at the storage layer. No cross-tenant data access is architecturally possible.
CSP, HSTS and frame-ancestors restrictions enforced on all workspace endpoints.
Same run + same pack = identical verdict. Always. The verification function has no hidden state.
The verification engine's core logic is open source. Auditors can read exactly how verdicts are computed.
We show you the control pack, a clean PASS, inject a same-actor maker-checker violation to get BLOCK, demonstrate that tampering is detected in the hash chain, then re-perform the run from scratch.
Book a 12-minute demolending-v2.1 pack · show 14 rules · run clean workflowLooking for design partners — banks and NBFCs with a validation or audit team ready to run a pilot. Talk to us →