>

01 · Pramana

Pramana.

The verification layer between data and decision in AI.

Pramana ensures that financial figures are recalculated from source data, locked until verified, explicitly labels estimates, and produces an audit-ready release record.

Acceptance demo

When the model and the source disagree.

Model proposes

The assistant drafts a total of $48,200 in the composer. It sounds confident. It is not yet bookable.

Risk: A number that can leave without proof.

Source sums

Your CSV or sheet lines sum to $41,650. Pramana recalculates in an independent sandbox from that source. Figures are never minted from the LLM.

Result: Verify fails until the draft matches source, or an ESTIMATE is stamped.

How it works

Verify. Lock. Sign. Receipt.

01

Independent sandbox

Recalculation runs from customer source (CSV / sheet) in a sandbox the language model cannot invent. Figures come from code against your data.

02

Pre-release lock

A browser extension (Chrome / Edge) blocks copy and send of unsigned totals. The lock is per-tab. New tabs start locked.

03

Sign PASS

Sign PASS only when verify matches source. That total is bookable. Humans remain accountable for the release.

04

ESTIMATE path

Intentional markup or judgment (for example a 15.7% market adjustment) may leave, but is stamped [… ESTIMATE - not bookable] on outgoing composer, copy, and send.

05

Finance receipt

Signed packs export as PDF and JSON so finance and audit can retain the proof pack with the work.

Doors

SMB sidecar. Enterprise API.

Same doctrine. Two ways in.

SMB

Sidecar plus local loopback bridge. Data stays on the device. Built for teams that already live in ChatGPT, Copilot, or Claude.

Enterprise

POST /v1/verify with an API key for governed pipelines that need the same verify gate in their stack.

Spec 12 locality

Customer data stays on their install. No phone-home of workspace by default. Cloud is optional, not the architecture.

Shared kernel

Built on the Verified Model.

Pramana is an overlay on the VedAGI Verified Model, the same doctrine behind MyCelAI and Serlyx: Figures ≠ LLMs, Estimates labeled, Actions-only writes, markings fail closed, bring your own model.

  • Figures ≠ LLMs
  • ESTIMATE labeled
  • Actions-only writes
  • Local-first · BYOM

Engagement

We work with a few enterprise and public sector teams at a time. If unsigned totals leaving AI tools is the problem you are wrestling with, talk to us.

VedAGI. Intelligence for Humanity.