What If Your Business Decisions Could Converge Like Software Tests?

Every organization runs on decisions. Hire this person. Acquire that company. Renew this vendor contract. Launch that pricing change.

Most of these decisions follow a pattern that nobody has formalized:

1. Someone declares what should become true ("we need a go/no-go on this acquisition") 2. Multiple people gather evidence from different angles 3. Contradictions surface ("source A says revenue is $42M, source B says $67M") 4. Someone synthesizes a recommendation 5. Someone with authority approves or blocks it

This pattern repeats across every function – sales, finance, procurement, partnerships, product. The actors change, the evidence sources change, but the structure is identical.

We have been building a system that treats this pattern as a first-class primitive. We call them Business Truths.

And they close a loop we opened months ago.

The loop

In January, we described the intent codec: the idea that the last piece of software is not whoever generates the prettiest output from a prompt, but whoever builds the decoder that makes reconstruction trustworthy. We introduced the concept of an intent packet – a structured specification of jobs, constraints, authority, contracts, budgets, and traceability – and argued that without one, you are shipping vibes and hoping the decoder guesses correctly.

In February, we showed what executable invariants look like: five real business cases expressed as Gherkin-like truths that are human-readable, executable, metered, and auditable. We showed what businesses actually drift from and why making leadership intent explicit is the precondition for any meaningful automation.

Business Truths are what happens when those ideas stop being separate articles and become a single running system.

A Business Truth is the intent packet made concrete. The Gherkin invariants are its guardrails. The convergence engine is its decoder. And the governance gates are what keep reconstruction honest.

A Truth is a job to be done, not a task to complete

A Business Truth describes what should be true: the desired outcome and its constraints, rather than a workflow. Evidence requirements belong in or alongside its versioned Truth spec. A commitment is separate: it binds the organization to a specific version under explicit authority. Describing an outcome does not establish that it has happened.

A concrete example:

Truth: Evaluate acquisition target

When: Our board identifies a potential acquisition

Job: Converge multi-source evidence into a go/no-go recommendation with traceable evidence before the LOI deadline

Success criteria:

• Recommendation produced with confidence >= 0.7
• All material contradictions surfaced and documented
• Each dimension (market, tech, financials) cites independent sources

Guardrails:

• No recommendation without adversarial review
• Contradictions must be surfaced, never resolved silently
• Human approval required before recommendation leaves draft

Notice what this does not specify: which databases to query, which APIs to call, which people to involve, or what order to do things in. The Truth declares the outcome. The system figures out how to get there.

This is the intent packet from the codec article – no longer theoretical. Jobs-to-be-done, constraints, authority, contracts, budgets, traceability. All present. All enforceable.

Convergence, not workflow

Traditional business process tools model work as a flowchart: step 1, then step 2, then a decision diamond, then step 3. This breaks down the moment reality gets complicated – when evidence contradicts itself, when a gap requires going back to research, when the team realizes they need expertise they did not plan for.

We took inspiration from a different domain: convergence in distributed systems. Instead of a linear workflow, we run an adaptive loop:

1. Seed initial research strategies (search wide for market context, search deep for financials) 2. Gather evidence from multiple sources in parallel 3. Extract factual claims with confidence scores and source citations 4. Detect contradictions across sources in real time 5. Identify gaps and propose follow-up research 6. Synthesize when – and only when – the evidence stabilizes

A loop can reach a fixed point, exhaust its budget, await authority, or stop because an invariant blocks progress. A fixed point means stability under that loop's rules; it does not by itself satisfy the Business Truth.

Assess the stable result against the declared outcome and evidence requirements. Formation can organize several attempts around the same problem, including parallel Competition and Huddle formations with different budgets or resources. Zero, one, or several attempts may produce acceptable results. Selection then chooses a candidate; only explicit authority can turn that choice into a commitment.

The governance gate: honest stopping

Here is the part that matters for real organizations: the system cannot override human authority.

When the convergence loop detects contradictory claims – say, two credible sources disagree on a company's revenue – it does not pick one and move on. It surfaces both claims with their sources, flags the contradiction, and blocks the recommendation until a human reviews it.

This is not a notification. It is a hard gate. The system cannot produce a recommendation while material contradictions remain unreviewed. The human does not rubber-stamp; they see the exact claims, the exact sources, and the significance of the disagreement.

This is what governed intelligence means in practice. The system does the work of gathering, organizing, and synthesizing evidence. Authority stays with humans. And the system is honest about what it knows and what it does not know.

The formation: right team for the problem

Different decisions need different collaboration structures. A pricing change needs a huddle – cross-functional, collaborative, everyone contributes. An acquisition decision needs a panel – formal review with designated critics, domain experts, and a synthesizer.

The system derives the appropriate team structure from the nature of the decision:

• Irreversible + high authority (like an acquisition) → Panel formation with curated experts
• Reversible + cross-domain (like a pricing experiment) → Huddle with capability-matched participants
• Moderate complexity (like a vendor renewal) → Discussion group with domain leads

Formation choice can be revised using evidence from earlier attempts. Keep unsuccessful attempts as well as selected results. Recording an episode is not enough: learning occurs when comparison of expected and actual outcomes changes the shared model used for future choices. Changing a formation or budget must be recorded as part of a new or resumed attempt.

Evidence provenance: every fact traces to a source

Every claim in the system carries provenance:

• Where did this come from? (which search result, which API, which document)
• How confident are we? (0.9 for primary sources like company filings, 0.5 for inferred claims)
• Does anything contradict it? (flagged in real time, not post-hoc)
• Who promoted it? (which agent extracted it, which human approved it)

When the final recommendation says "the target company has $42M in revenue," you can click through to the exact source that supports that claim. When two sources disagree, both are preserved. The system does not hide disagreement; it documents it.

This matters for regulated industries, board-level decisions, and any context where "trust me, the AI said so" is not good enough.

Integrity checks and the limits of replay

A Merkle root can identify the recorded fact set and support integrity checks. Matching roots under the same encoding provide evidence of matching recorded content, not proof that the claims are true or that fresh reasoning will reproduce them. Exact replay depends on complete recorded inputs, component and policy versions, and the execution conditions of the relevant component. Repeating an intent with fresh LLM calls is a different operation.

The practical value is provenance: which evidence supported this recommendation, which version was selected, and what changed afterward. Later corrections are appended with their reasons; inconvenient proposals are not erased.

The learning loop: outcomes inform later decisions

After each decision plays out, the system captures a learning episode:

• What did we predict would happen?
• What actually happened?
• Where was the prediction error?
• What should we do differently next time?

These lessons can revise planning priors, including evidence gathering, formation choice, and resource allocation. They do not create authority. Selection may be delegated under an explicit mandate; a formation cannot grant itself permission to bind the organization.

The adaptive cycle is Coordinate → Commit → Update Priors. Learning must influence later reasoning or choices while preserving the original promise and its outcome.

Five Business Truths we are building

We started with one (acquisition due diligence) and designed four more:

1. Evaluate acquisition target – converge multi-source evidence into a go/no-go recommendation 2. Validate pricing change – simulate impact across existing customers, revenue, and competitive positioning 3. Assess vendor renewal – converge internal usage data with market alternatives and compliance requirements 4. Qualify partnership opportunity – determine strategic fit, revenue potential, and integration complexity 5. Detect churn risk – converge behavioral signals with competitive context into an intervention plan

Each one follows the same pattern: declare the outcome, let the system figure out evidence gathering and team formation, gate on governance, learn from results.

Closing the loop

The traditional approach to business decisions is either:

a) Ad-hoc – someone does research in a spreadsheet, writes a memo, sends it around for comments, and the boss decides. Evidence is scattered, contradictions are hidden, and institutional learning is zero.

b) Over-engineered – a BPM tool encodes a rigid workflow with approval matrices and SLA timers. It works until the process changes, which is constantly. The tool becomes a constraint rather than an enabler.

Business Truths offer another approach: describe what should be true, examine the evidence, authorize commitments, and learn from outcomes. The structure can adapt to the problem while the history of each attempt remains inspectable.

In codec terms, the intent packet is the Business Truth specification. The surrounding system develops candidate results, checks their evidence, and supports selection. Governance controls what may be promoted or authorized; it does not turn an assertion into an accomplished business outcome.

The Business Truth specification connects declared intent to a governed recommendation.

Kenneth Pernyér is the founder of Reflective Labs, building Converge – an open platform for governed intelligence.