Model feedback from authoring the reference examples

Two fictional reference impact models (examples/intermediary and examples/market-adoption) were written against the model to find what it could not say, or said awkwardly. This page records each finding and what was decided, so the reasoning outlives the conversation that produced it.

Decisions are done (implemented and tested), documented (a convention, now written down), planned (assigned to a roadmap item), or declined (with the reason).

Done

Finding	Decision
A network-wide roll-up or common-layer indicator has no single outcome to measure, so it was attached to an arbitrary one	An indicator's measures may name a pathway. Validation accepts it, and a pathway-filtered view includes pathway-level indicators only for that pathway
Ordinal levels cannot be summed or averaged, so a "median level" lived only in formula text	Added median and mode aggregations
No actor type for a government office; no external-factor kind for an opportunity (risks already had one)	Added the government actor type and the opportunity factor kind
What may a restricted evidence record contain, and what do renderers show?	A restricted record (or one with no sensitivity set) may carry a title, type, strength, and source. Its excerpt, summary, and URL are withheld by every renderer, since a URL can reveal where a confidential document lives
A group's target is one outcome; nothing said what a nested group's target is	A nested group targets the same outcome as the group containing it. Validation enforces it (VAL-GRP-004)
The coverage rule wanted an indicator on every owned outcome, which inflates indicator counts	IMP-MEA-001 accepts coverage through another indicator's proxyFor

Documented

Finding	Decision
direction: increase is vague for a qualitative indicator whose rise is bad news	direction states the desired direction. If a rising level is bad news, the direction is decrease; say why in the indicator's definition
Negative-polarity links and precondition logic	A negative link is an inhibitor, not a precondition: it does not count toward all, any, or at_least logic or toward k. How an inhibitor lowers a pathway's status is decided with the roll-up algorithm (RMI-THEORYOFCHANGESPEC-018)
A competitor's momentum has no element type	Model it as an outcome in the interest sphere with category respond, linked to the outcomes it inhibits. A first-class competitor-response element is deferred until a second example needs it
Capability relation dependsOn is camelCase while every other enum is snake_case	It comes from prism-core's shared CapabilityRef and is intentionally left alone so the PRISM modules stay interchangeable
In YAML flow mappings ({a: b, c: d}), an unquoted comma inside a value silently splits the mapping	Schema validation catches it. The authoring guide recommends block style, which fmt always produces
provenance.decidedBy is free text, so nothing stops a personal name	By convention it names a group or role. Detecting personal names is not something code can do reliably; the authoring guide and the review checklist say so
Responsibility (own, shared, partner) and a problem tree are not modelled	Both are in RMI-THEORYOFCHANGESPEC-044. Until then the ceiling is a level and root causes are text

Planned

Finding	Roadmap item
A weighted_mean roll-up cannot be computed if the weight metric has no indicator	Lint rule IMP-MET-003, in RMI-THEORYOFCHANGESPEC-014
pathway.outcomeIds and indicators' pathwayIds are two lists kept consistent by hand	Lint rule IMP-PTH-003, in RMI-THEORYOFCHANGESPEC-014
A group's members should each have a link into the group's target	Already covered by IMP-STR-007, in RMI-THEORYOFCHANGESPEC-013
Assumption appliesTo and a link's assumptionIds can disagree	A consistency lint rule, in RMI-THEORYOFCHANGESPEC-014

Declined (for now)

Finding	Reason
A window field on metrics ("trailing twelve months")	Useful, but it changes how observations and targets are compared, so it deserves its own design. It is recorded for a later minor version; until then the window is stated in the formula
One indicator measuring both an outcome and an assumption	Two indicators are clearer and cost little. Revisit if authors keep hitting it
A "key outcome" or "monitored" flag on outcomes	proxyFor already expresses "this indicator also stands in for that outcome" without a second mechanism
A layer label on outcomes for two-layer impact	Layer one versus layer two is already visible through sphere, the indicator claim, and the ceiling of accountability. A fourth signal would add overlap, not information
changeType has no standardization	That value describes what an intervention does, not the kind of change an outcome describes; the outcome's adoption or policy change type fits