The Intent Pipeline

From desired outcome to a reviewed proposal.

This illustrative planning path includes admission, decomposition, Huddle, review, and simulation. Its final step submits a proposal; it does not grant authority to commit.

1
Admission
Can we do this? Enough context? Have budget? Permitted?
→ Feasible
2
Decompose
Break complex intent into an IntentNode tree
→ 3 sub-intents
3
Huddle
Multiple reasoners explore plans; unsuccessful attempts remain in history
→ 4 candidate plans
4
Adversarial
Five kinds of skepticism stress-test every plan
→ 2 plans survive
5
Simulate
Five dimensions: outcome, cost, policy, causal, operational
→ Proceed with caution
6
Propose
Submit a candidate; selection and binding authority remain separate
→ ProposedFact
Adversarial Review

A plan meets skeptics before it reaches the engine.

Five kinds of skepticism put candidate plans under pressure: assumptions, constraints, causal effects, economics, and operations. Rejected proposals and their evidence remain in history, even when they no longer belong in the current selection.

Five Kinds of Skepticism

Assumption
What are the unstated assumptions in this plan?
Constraint
Do declared constraints actually hold?
Causal
What are the second-order effects?
Economic
What does this really cost?
Operational
Can the organization actually execute this?

Findings

⚠ Assumption
Vendor availability assumed but not verified
ℹ Constraint
Budget constraint satisfied: €48K < €50K limit
⚠ Causal
Switching vendors may trigger contract penalties
⛔ Economic
Total cost including transition: €62K over 18 months
ℹ Operational
Team has capacity; procurement lead available
Organizational Learning

Outcomes improve later plans, not authority.

Compare expected and actual outcomes, then revise the shared model for later attempts. Recording an episode alone is not learning. New knowledge does not rewrite an existing commitment or erase an unsuccessful attempt.

Illustrative episodes

Evaluate vendor compliance
Predicted 0.85 → Actual 0.91
OutcomeBeatExpectation
Compliance pack selection was accurate
Plan outbound campaign
Predicted 0.72 → Actual 0.68
OutcomeMissedExpectation
Budget estimates need calibration for Q3
Approve expense report
Predicted 0.95 → Actual 0.94
OutcomeMatchedExpectation
Expense outcome prediction was close to the observed result
Score inbound lead
Predicted 0.80 → Actual 0.45
OutcomeMissedExpectation
Behavioral scoring overweights website visits

Revise the shared model

1 Record attempts and expectations
↓
2 Compare actual outcomes
↓
3 Revise planning priors
↓
4 Apply and assess in later attempts
↶
Prior calibration shift
+0.00
Learning changes future reasoning.
It does not rewrite commitments.

Zero, one, or several attempts may reach a favorable fixed point: a stable result that meets the declared outcome and evidence requirements. If none does, record what happened and try a changed budget, formation, or approach. Selection and authorized commitment are separate steps.

The Stack

Four layers. Organism is the intelligence layer.

Layer 4

Helm

Control surface

Where operators describe outcomes, inspect evidence, and exercise authority.

helms.zone
Layer 3

Axiom

Specification layer

Checks Truth specs without granting authority or proving the desired outcome.

axioms.zone
Layer 2

Organism

Intelligence layer

Organizes reasoning, alternative attempts, review, and learning through formations.

organism.zone
Layer 1

Converge

Convergence kernel

The low-level contract for shared context, facts, authority boundaries, and promotions.

converge.zone

Runtime-Runway supplies surrounding capabilities such as messaging and storage. The composed environment can be understood as an operating system for intent; Converge alone is its kernel.

Selecting a plan does not bind the organization without an explicit mandate. Converge enforces the authority boundary.

From intent to commitment: how the parts fit together