An Ecore micro-model of collaborative decision analysis records: criteria hierarchies, alternatives, expert panels, judgments, and computed results. It is the “decision” floor of the micro-model tower, sitting directly above the lifecycle model (with the seal model and the IAM model beneath it). The model holds what was asked, who judged, what they said, and what came out; the math - AHP eigenvalues, weighted sums and products, best-worst scaling, message-passing over the criteria graph - is a capability resolved at run time, the same provider-free posture the seal model takes toward cryptography.

The model is one of a deliberate pair. Analysis without binding is a first-class use: a grassroots record of judgments needs no variation point. Binding without analysis is equally first-class one floor up: a fiat binding needs an alternative but no evaluations. The decision binding model commits an alternative into a variation point; this model is the record of why one alternative deserved it.

Competitive landscape

Saaty-lineage commercial MCDA. Expert Choice / Comparion, Decision Lens, 1000minds, TransparentChoice, D-Sight. Proprietary hierarchy editors behind a login, one methodology per analysis, weak visualization, and the analysis lives - and dies - inside the vendor’s walled garden. All sold top-down to the executive who buys for the whole org.

Decision-intelligence tools. Cloverpop, Quantellia, Loomio: decision tracking and consensus, not formal analysis. They compete for the same buyer attention without occupying this position.

Academic and open source. Decision Deck and the XMCDA exchange standard are the closest in spirit - federated, machine-readable, multi-method - but academic. R and Python AHP packages and Goepel’s BPMSG templates compute; they do not keep records, federate, or integrate.

The real competitor: Excel. Most enterprise MCDA actually happens in a spreadsheet, and Excel will always be cheaper for the math. The wedge is the consequences of having done the math: judgments that are committed, dated, and signed; divergence that stays legible; results that cite their basis. Excel for the math, the model for the record.

What a typed model adds

Judgments are records, not cells. Every judgment carries its expert, date, confidence, and rationale, and - because everything here extends the seal model’s SealedElement - is signable and sealable. “I provided my judgment, it is committed, dated, signed; if you disagree, commit yours.” This is what makes individual, bottom-up adoption possible: an analyst records judgments in their own repo without asking permission, the way Git and Notion spread.

Adoptable at any level. Managers want their bias, not somebody else’s: Hammurabi stood above the laws but wanted his judges bound by them. A criteria hierarchy with recorded weights constrains the layer below and defends against the layer above, so the model is adoptable by a team, a middle manager, or an executive - no incumbent tool occupies that position, because they all sell top-down.

Override-as-record. Judgment.overrides adds a parallel assertion with the overrider’s authority and date; the original stays. Overruling an expert is legitimate - erasing one is not. Sensitivity by authority (“how much would I have to disagree with this expert’s weights to flip the result”) is computable precisely because overridden judgments survive.

Delphi confidentiality is inherited, not invented. Experts not seeing their own ranking or each other’s judgments, facilitators seeing everything - the Delphi method (Linstone & Turoff, RAND) - is IAM access control plus seal encryption on judgment records, which every class here carries by inheritance. Group membership already is the key distribution graph; blinding an expert panel is instance data, not a feature.

Hybrid methods per node. Criterion.method selects how each node aggregates: AHP eigenvalue rigor where it is worth the cognitive cost, direct weights where intuition suffices, best-worst scaling where pairwise matrices would exhaust the panel, narrative elicitation where experts would rather write than fill matrices. No commercial tool mixes methods below the analysis level; real decisions do.

Narratives with honest elicitation. A free-form review contains the n-wise comparisons and evaluations a GenAI elicitor extracted from it, so every elicited judgment is traceable to source text. What the narrative does not support is absent, not hallucinated - the agent admits what it does not know.

Vector results with explicit convergence. The default output is a vector per alternative - benefit over time, risk, cost Opex/Capex - not a single ranking number: dimensionality reduction to where humans can decide, not a decision made for them. Cycles (ANP-style feedback) are opt-in; execution is fixed-point iteration; a Result says whether it converged and with what residual, never a silent last-iteration value.

Reusable hierarchies as artifacts. A vendor-selection criteria hierarchy authored once in a Catalog, versioned, shipped as a Maven artifact, and instantiated across decisions via Criterion.prototype. Excel cannot do this; Expert Choice does it only inside its walls.

Applications

NCDA. The capability this model exists for: federated collaborative decision analysis - drawio and YAML authoring surfaces compiled to the criteria graph, the graph message-passing engine executing it, ECharts multi-dimensional visualization of result vectors, GitHub Actions regenerating the report on push. Decision analysis without a vendor account.

Expert assignment as analysis. Experts are IAM Subjects (agentic experts are Principals like any other), and an Alternative may represent any element including a Subject - so “which expert for this criterion group” is the same shape as any other analysis, run on the same substrate.

Up-tower consumers. The decision binding model narrows embodiments with analyses over them and cites the analysis behind each binding. Remediation prioritization under a budget in the CISO model, treatment selection in governance, portfolio ordering in work - each is an analysis over comparands the tower already holds.

Study and demo material. A worked analysis - a real Nasdanika capability-sequencing decision, not Alternative A versus Alternative B - seeds the NCDA template repository and doubles as the writing subject: the record is the demo.

Model overview

Relation to other Nasdanika work

The floor below is the lifecycle model: decision elements have lifecycles. An analysis moves from drafted through judged to closed, a judgment can be retracted, a published criteria catalog is eventually retired, and Delphi rounds and approval gates are lifecycle data over these records rather than a workflow engine bolted on — Comparand extends Staged, so every record here occupies stages via dated sojourns. Beneath lifecycle sit the seal model and the IAM model, which contribute the other two properties this floor depends on: Delphi blinding as access control plus encryption, and judgments that are committed, dated, signed.

The decision binding model sits directly above: its variation-point alternatives extend Alternative, its embodiments extend Comparand, its bindings cite the Analysis that informed them, and its VariableElement extension point carries the chain upward so governance, work, architecture, and threat elements are all analyzable and can all carry decisions.