BiMA·gov
Governed meaning, before the agent

Stop arguing about the numbers. Start deciding with them.

Two people pull the same number and get two different answers. The definition that would settle it lives in someone's head, never written where a system can reach it. The way out has two halves: ask better, and keep a governed record of what was decided.

Start the conversation → Read the paper
01

Ask better

Clarify the rule a person holds until it's unambiguous.

02

Keep the record

Written down once, with its owner and full history. A historian, not a guess.

Owned Approved Versioned
03

Grounded, or it refuses

Cite the approved source, or say it won't guess. The refusal is the feature.

A refusal routes the missing rule back into capture.

Same loop for any rule that lives in a head, not just revenue.

The problem

It was never a tooling problem. It is a meaning problem.

Both numbers came out of competent BI. What's missing is the definition that reconciles them, and it lives in people's heads. Adding an AI agent just makes the wrong answer faster and more confident.

Today's argument
  • Definitions live in two people's heads, never written in the same place
  • The meeting argues whose number is right instead of deciding
  • Point an agent at it and it answers anyway, fluent and unsourced
Governed
  • Each rule is stated by its holder, approved by its owner
  • Two legitimate definitions coexist, scoped and attributed
  • Every answer cites its source, or it refuses to guess
Figure 1 · The same meeting, before and after a governed layer.
How it works

Capture from heads, not ingestion.

Most context tools ingest: connect everything, index everything, retrieve. That fails on the first hard question, because the rule that resolves the argument was never written down.

The governed loop A question gets a cited answer or a refusal. A refusal routes to the owner, who states and approves the missing rule, which makes the brain grow, ready for the next question. Every miss makes the brain bigger. every miss makes the brain bigger Question Answer, or refuse Owner states and approves the rule Brain grows ✓ cites the source → or says it won't guess
Figure 2 · The loop. A refusal is not a dead end, it is the next governed rule.
Where it's applied first

Applied first to mid-market financial services.

Firms with the tribal-knowledge problem in full and no data-engineering team to build the layer themselves. In a regulated firm, a number nobody can trace to a governed definition is a finding waiting for an examiner. Recording each definition, its owner, and its approval is the difference between an AI deployment compliance blocks and one it signs off.

What it does, measured

Grounded, or it refuses. Tested 120 times.

Tested against 120 labeled questions across three fully synthetic firms. Every firm, fact, and figure was fabricated for testing. No client data is involved.

0 / 120
Invented facts, across all 120 questions.
0%
Answerable questions wrongly refused.
45 / 45
Answered with the right cited source.
Grounded + cited
72 / 72
Refused instead of guessing.
Incl. 42 / 42 near-miss
3 / 3
Conflicts: both surfaced and flagged.
Strict grading
3
Synthetic domains: RIA, hospital, manufacturing.
Held-out
Figure 3 · Held-out synthetic environments. An answer scored only if the cited fact was correct, not merely the prose. Methodology available on request.
The conversation

Five design partners. Built with you, not sold to you.

BiMA·gov is onboarding a small number of mid-market financial services firms that recognize the two-answers meeting and want the layer in place before the agent. It starts with a structured consultation, not a software install, because the capture process is the product and it has to fit how your firm actually holds its knowledge.

Start the conversation → Read the paper