How Judgment Pack compares to prior art
On this page
Reviewers usually arrive with an existing model in mind and ask, reasonably, "how is this different from the thing I already use?" This page answers that for the closest neighbors. The short version: Judgment Pack standardizes a portable, testable document describing the judgment behind one decision — not a rule language, not an orchestration language, and not a knowledge store.
Decision Model and Notation (DMN)
DMN is the closest analog and the most useful comparison. DMN standardizes decision logic — primarily decision tables — and a Decision Requirements Graph that links decisions together. Both DMN and Judgment Pack treat a decision as a first-class, portable artifact.
Where they differ:
- Evidence, uncertainty, and escalation are first-class in a pack. DMN centers on inputs and decision tables; a pack additionally declares what evidence is required, what happens when information is unknown or conflicting, and when a human must take over. In a pack, "unknown" is a first-class result, not an implicit false.
- Judgment Pack ships a smaller core and defers composition. DMN already standardizes the requirements graph; Judgment Pack deliberately ships only the single decision today and treats composition as an open proposal (the vision page explains why).
- Conformance is explicitly separated from truth and authority. A pack's specification takes care to say that a conforming document is not thereby correct, authorized, or safe.
A fair one-line framing: Judgment Pack aims to be to AI-era, evidence-driven judgment what DMN is to business decision tables — with uncertainty and escalation built in, and a smaller starting core.
Policy engines (OPA / Rego, Cedar)
Policy engines evaluate whether an action is allowed under a policy, and they are excellent at it. The difference is purpose. A policy engine answers an authorization question and is designed to be enforced. A pack represents judgment — the evidence and reasoning behind a decision that may be uncertain, may abstain, and may escalate to a human. A pack explicitly refuses to conflate evidence with authority; authorization stays in your own systems. The two are complementary: a policy engine might gate whether a pack's outcome is permitted to act.
Rule engines and json-rules-engine-style formats
General rule engines execute condition/action rules. A pack is not a general-purpose rule language (that is an explicit non-goal). It constrains itself to representing one decision's judgment so that the document stays small, testable, and portable, rather than becoming a Turing-complete program whose behavior no second tool can reproduce.
Workflow and process notation (BPMN)
BPMN models process — the sequence of tasks and gateways in a workflow. A pack models the judgment at a single decision point, not the flow around it. The common DMN/BPMN pairing is instructive: process notation calls out to decision artifacts. A pack sits at that decision point, carrying the evidence and escalation logic BPMN never intended to hold.
Knowledge graphs, vector stores, and retrieval
These answer what is known and what is relevant. They do not declare which decision the information serves, when a rule applies, which exception overrides it, or when to escalate. A knowledge source is an evidence input to a pack, not a parent of it. Why Judgment Pack? develops this at length.
LLM evaluation frameworks
Eval frameworks record inputs and expected outputs to measure a model. A pack represents the rules and reasoning being evaluated, in a form separate from any one model or prompt. Evaluation cases can be attached to a pack, but portable evaluation is a possible later profile, not a claim of the current draft.
Summary
| Model | Primary question | What a Judgment Pack adds |
|---|---|---|
| DMN | What is the decision logic? | Evidence, uncertainty, escalation as first-class; smaller core |
| Policy engine | Is this action allowed? | Judgment vs authority kept separate; abstention |
| Rule engine | Which rules fire? | A bounded, testable, portable single decision — not a language |
| BPMN | What is the process flow? | The judgment at a decision point, not the flow |
| Knowledge graph / RAG | What is known / relevant? | Which decision the knowledge serves, and when to stop |
| LLM eval | How good is the model? | The reasoning being evaluated, independent of the model |
For where each proposed capability would live if it graduates, see the architecture vision; for the recurring questions, see the FAQ.