The AI Trust Platform
Three planes working together: verified context for reads, deterministic enforcement for writes, and a cryptographically signed audit trail for everything in between.
A deterministic cage around probabilistic intelligence.
Book a DemoThree Planes. One Platform.
Praxom
"Is this action allowed?"
Intercept, validate, and enforce every agent action before it executes — even if the agent has been compromised. Enforcement lives outside the model.
- ✓ALLOW / DENY / REWRITE / WARN / ESCALATE — decided pre-execution
- ✓Ed25519-signed, hash-chained attestation store
- ✓Fail-closed under the production profile — a workspace with zero constraints is schema-mode and allows the call
Flight Recorder
"What happened, and can we prove it?"
Every agent action recorded, cryptographically signed, and exportable in OCSF format. Immutable audit trail for EU AI Act, DORA, and SOC 2.
- ✓Ed25519-signed action records
- ✓OCSF export for SIEM integration (Splunk, Sentinel)
- ✓Scored, PDF-rendered compliance reports
Epistom
"What does this data mean?"
Validated understanding of enterprise data, with calibrated abstention. Five trust tiers — deterministic paths answer first; the LLM only runs where it earns its place.
- ✓Virtual knowledge graph (OWL 2 QL) — steward-approved definitions
- ✓SQL validation gate before any query executes
- ✓Calibrated abstention — “I don’t know” beats a wrong answer
The gauntlet every action runs
Nine slots, in this order, on every proposed action. Any slot can stop it. The whole path is deterministic and fail-closed — a jailbroken model changes nothing, because the model isn't consulted.
Workspace hard-stop. Checked before identity, before anything — a killed workspace executes nothing.
Caller resolved to a registered agent and trust tier — crypto-verified via AWS IAM, Azure MSI, or SPIFFE workload identity (optional in dev, required under the production profile, where an unproven agent is denied).
Is this tool in the agent’s granted scope, under a valid task lease? Expired lease, no action — on by default in the production profile.
Per-agent cost ceilings and action-count budgets — in the production profile a metering failure denies rather than allows. Runaway loops suspend themselves.
Data-taint tracking blocks exfiltration paths — sensitive reads can’t flow to outside destinations. On by default in the production profile; unclassified data is treated as sensitive. Fails safe to max-tainted.
Business rules evaluated deterministically: JSON predicates (auto tier) and SHACL shapes (formal tier).
Where policy allows with changes: cap amounts, mask PII, set safe defaults — the original preserved in the record.
ALLOW / DENY / REWRITE / WARN / ESCALATE. Deterministic, fail-closed, no LLM anywhere in this path.
Ed25519-signed, hash-chained record of the decision — every path through the gate ends here, allowed or not.
How a question earns an answer
Five trust tiers, tried in order of certainty. Most questions never reach an LLM — and when nothing clears the confidence bar, the platform abstains instead of guessing.
How It Fits Together
Six standalone packages. Pip install only where live.
Core capabilities ship as independent, MIT-licensed packages — including the free MCP scanner and the verifier that lets your auditor check our work without trusting us. A pip install line is shown only when that package is marked live.
pip install pramiti-mcp-gatewayFree OSS security scanner for MCP servers — see what your agents can already do. Live on PyPI.
pip install pramiti-mcp-verifyNeutral, offline verifier for signed execution records (SEP-2828). Your auditor checks the chain without trusting us.
pip install pramiti-flight-recorderAgent action recorder with Ed25519 signing, hash chaining, and OCSF SIEM export.
pip install pramiti-agent-finopsPer-agent warehouse cost caps and tool-call budgets — the loop-stopper.
pip install pramiti-export-copilotDeliver semantic models to a webhook, GitHub or GitLab — hash-based change detection.
pip install pramiti-merger-reconcilerCross-ontology class matching for M&A data integration.
Ready to make your AI safe?
30-minute demo. See Epistom and Praxom on your own data.