Platform overview

A deterministic cage around probabilistic intelligence.

Three planes working together: verified context for reads, deterministic enforcement for writes, and a cryptographically signed audit trail for everything in between.

Request a demo
0
Wrong answers (eval)
6 eval runs · 2 model families · LLM fully down — every failure became an abstention
5,000+
Automated tests
In the platform repo, run on every change
5
Compliance frameworks
SOC 2 · DORA · ISO 42001 · NIST AI RMF · EU AI Act
22
Industry starter kits
Pre-built knowledge models, SaaS to healthcare

Internal cross-model gauntlet (2026-07-21): with the LLM hard-down, throttled, or quota-capped — 0 wrong answers across 6 runs and 2 model families; every failure became an explicit abstention. Measured on our eval suite, not a customer SLA. Compliance frameworks are scored report capabilities generated from the attestation record — not certifications of Pramiti Labs.

Where enforcement belongs

Every agent action passes one checkpoint

INTENT
TOOL SELECTION
ARGUMENTS
THE ACTION LAYERthe last deterministic checkpoint
EXECUTION
EFFECT

Intent and arguments are probabilistic. The action layer is where determinism takes over — before an effect reaches the world.

Products

Three planes. One platform.

Control plane

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: no policy, no action
Explore Praxom
Audit plane

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, hash-chained records — independently verifiable offline
  • OCSF export for SIEM integration (Splunk, Sentinel)
  • Scored, PDF-rendered compliance reports
Explore the evidence plane
Context plane

Epistom

“What does this data mean?”

Keeps the agent's understanding of enterprise data correct. Five trust tiers — deterministic paths answer first; the LLM only runs where it earns its place.

  • Business knowledge model — steward-approved definitions
  • SQL validation gate before any query executes
  • Calibrated abstention — “I don't know” beats a wrong answer
Explore Epistom
The write plane, slot by slot

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.

  1. Kill switch
    Workspace hard-stop. Checked before identity, before anything — a killed workspace executes nothing.
  2. Identity
    Caller resolved to a registered agent and trust tier — AWS IAM, Azure MSI, or SPIFFE workload identity.
  3. Scope
    Is this tool inside the agent's granted task scope? Out-of-scope proposals stop here.
  4. Leases
    A valid, unexpired task lease must cover the action. Expired lease, no action.
  5. Budgets
    Per-agent USD and action-count ceilings. Budget-velocity spikes auto-suspend the agent.
  6. Tool pin
    The tool schema must match the version that was reviewed — swapped or mutated tools are refused.
  7. Policy + taint
    Business rules evaluated deterministically — JSON predicates and formal SHACL shapes — with data-taint state feeding the decision so read-sensitive-then-exfiltrate paths are blocked.
  8. Rewrite
    Where policy allows with changes: cap amounts, mask fields, set safe defaults — the original preserved in the record.
  9. Attest
    Ed25519-signed, hash-chained record of the decision. Every path through the gate ends here — allowed or not.

Two engineering invariants hold the pipeline together: the cheap checks run before any expensive reasoning — a deny never pays for reasoning — and every slot is synchronous and IO-free by design invariant.

The read plane, tier by tier

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.

01
KnowledgeDETERMINISTIC
Definitions, acronyms, discovery — answered straight from the steward-approved knowledge model. Zero LLM calls.
02
MetricDETERMINISTIC
Governed metrics compile to SQL from their approved definition — the number means what the steward said it means.
03
VerifiedDETERMINISTIC
Known question shapes match certified query templates and run over the virtual knowledge graph. No generation, no guessing.
04
VocabularyLLM-BACKED
Plain-English synthesis — the LLM writes prose, but only over verified context it is handed. It never invents definitions.
05
AnalyticalLLM-BACKED
Novel questions: LLM-generated SQL, then every table and column is checked against the live schema before execution.
BELOW THRESHOLDWhen confidence is too low for any tier, the answer is “I don't know” — an explicit, logged abstention. Never a guess.
Architecture

How it fits together

pramiti-platform-architecture
# Data flow: agent request to audited execution

AI Agent (Claude, GPT, Cursor — anything MCP)
   |
   +--> Epistom (Context Plane)
   |      "What does this data mean?"
   |      Verified definitions + SQL validation gate
   |
   +--> Praxom (Control Plane)
   |      "Is this action allowed?"
   |      ALLOW / DENY / REWRITE / WARN / ESCALATE
   |
   +--> Flight Recorder (Audit Plane)
   |      Ed25519-signed + OCSF export
   |      Immutable audit trail
   |
   v
Database (read-only)   |   Action Targets (Salesforce, Slack, APIs)
Standalone packages

Seven pip packages. Zero platform dependencies.

Core capabilities ship as independent packages — including the free MCP scanner and the verifier that lets your auditor check our work without trusting us.

Free OSS security scanner for MCP servers — see what your agents can already do.

live on PyPI

Neutral, offline verifier for signed execution records (SEP-2828). Your auditor checks the chain without trusting us.

live on PyPI
pramiti-agentguard

Install-time MCP guard — 130 real invocations, zero wrong verdicts in its ship-pass matrix.

ships with engagements
pramiti-flight-recorder

Agent action recorder with Ed25519 signing, hash chaining, and OCSF SIEM export.

ships with engagements
pramiti-agent-finops

Per-agent warehouse cost caps and tool-call budgets — the loop-stopper.

ships with engagements
pramiti-export-copilot

Push semantic models to GitHub, GitLab, or any webhook ingestion endpoint — hash-based change detection.

ships with engagements
pramiti-merger-reconciler

Reconciles two companies' knowledge models for M&A data integration.

ships with engagements

Plus the epistom-agents SDK — guard decorators for LangGraph, AutoGen, CrewAI.

Ready to govern the next action?

30-minute demo. See Epistom and Praxom on your own data.

Request a demoSee pricing