093 — Memory comparisons keep evidence per value

Type: reference/types/adr.md · Status: accepted

Status: accepted Date: 2026-09-26

Context

An axis can contain an implemented automatic route and a merely afforded manual route. Giving the whole set its weakest evidence basis correctly limits claims about the complete profile, but loses the implemented route in membership counts. A comparison corpus needs both questions: which mechanisms have support, and which complete profiles can be compared.

Decision

Adopt option 2 of “Evidence granularity for memory-comparison counts”. Each controlled value carries its own evidence basis, canonical supporting records, and rationale. Axis coverage remains separate. known means complete coverage; partial preserves positive findings while naming unresolved included parts. Absence, inapplicability, uninspected and indeterminate cases stay explicit.

A value's strongest supported witness establishes its existence within the scope; weaker alternatives remain described in its records. It does not establish the same strength for every route. Count a system once per question. Implementation counts use code-grounded wired, observed or causally supported values. Complete-profile distributions require complete coverage and that strength for every member. Neither an omitted value nor weak evidence is an observed negative. Route-specific questions still require the full records.

Considered alternatives

Keeping only complete-profile statistics preserves a smaller contract but cannot answer supported mechanism-membership questions. Keeping those statistics alongside per-value counts preserves their useful meaning.

Narrowing every comparison boundary can isolate stronger routes, but multiplies profiles and risks counting a source twice. Scope remains an evidential boundary, not a means to discard inconvenient alternatives.

The chosen form uses a value list plus an evidence map whose keys must match exactly. It preserves direct set operations while making evidence explicit. Canonical records carry route distinctions; a second route taxonomy is not introduced. Structural checks cannot decide whether a quoted source warrants the authored classification.

Consequences

Analysis coordinators and memory specialists consume the type and producing instructions as binding authoring rules. Schemas and validators enforce shape, controlled values and declared references. Matrix and table emitters preserve per-value evidence; statistics and synthesis consumers apply separate positive and complete-profile counting rules. Per-value evidence adds authoring and validation work. Historical retained bytes are immutable; new analyses are required for admission under the current contract.

This decision governs the declared memory boundary of one analysis. It does not establish corpus representativeness, semantic correctness, observed benefit, or comparability between different scopes. Partial positives support lower-bound counts, not whole-population prevalence estimates.

The memory report type owns the authoring and counting contract; at the time of this decision the fields lived in the since-retired single-file analysis result type.