103-Classify memory profiles after record verification
Type: reference/types/adr.md ยท Status: accepted
Status: accepted
Amended 2026-10-06: the profile loop permits two corrections, matching the record loop (commit on this date). Two runs stopped after one correction on a second-look finding of the same kind the first had named; the one-correction rule below records the prior decision. Date: 2026-10-04 Amends: ADR 093 for the profile carrier and author role. Amended by: ADR 107 for the profile representation and semantics. Classifier ownership, evidence boundaries and independent verification below remain in force.
Context
Memory analysis combines source tracing with classification into ten comparison axes. Classification adds fourteen output obligations and about 9.2 KB of type definitions to memory, reconciliation and record verification. Profile faults then consume memory correction and reconciliation rounds, even when the source-native account is sound. Separating these tasks reduces what each analyst must hold without weakening evidence or independent review.
Decision
After reconciliation and independent record verification close without blockers, run a profile classifier and independent profile verifier before synthesis. Allow one profile correction. Persistent blockers stop the run before publication; they never reopen the analysts or reconciliation.
The classifier writes memory-profile.md as the sixth manifest-pinned member
of a complete analysis set. It owns the unchanged memory-comparison mapping
and Comparison rationale under the profile type.
It cites existing canonical records, applying reconciliation amendments and
supersessions. It declares or annotates no records and adds no evidence.
Optional frozen-source reading resolves a named ambiguity in a record's cited
paths; it cannot substitute for a missing supporting record. Missing facts
retain warranted bounded assessments and prevented conclusions.
Memory analysis keeps the descriptive facts, records, annotations, write-side and read-back accounts, in source-native terms. Reconciliation and record verification check those records independently. The overview records separate record, profile and synthesis verification. Blocked and out-of-scope sets remain overview-only. Axis definitions, values, per-value evidence and counting rules from ADR 093 remain unchanged.
Considered alternatives
Keep classification in memory analysis. Avoids two jobs but retains the combined task and profile-driven analyst correction rounds.
Merge the classifier's mapping into the memory member. Saves a file but rewrites the analyst's accepted bytes. A separate pinned member preserves exact provenance and ownership.
Let the classifier declare evidence or return work to memory. Could recover missing facts, but creates another analyst and return loop. Two Luna replays of retained Dynamic Cheatsheet records reproduced all twenty axes' values, coverage and per-value strengths after one bounded correction, without source reads. No axis weakened because a needed fact was missing. This supports keeping the narrower classifier for now; those records were authored with the axes in view, so future sufficiency remains untested.
Consequences
The workflow supplies the profile type and job instructions by absolute read-first paths as binding worker inputs, and schedules independent verification before synthesis. Validators enforce the sixth member, identity, controlled shape and references against canonical set declarations. Publication pins its bytes with the record members. Matrix, landscape and taxonomy consumers read the profile member directly; there is no legacy-location fallback. The site publishes the current retained profile alongside its supporting records.
Memory loses fourteen output obligations; reconciliation and record verification lose their axis checks. Mandatory input falls by about 10 KB for each. Existing unrelated concerns and residual instruction conflicts do not fall further in this phase. The classifier and verifier add jobs and their own contracts; this redistributes work and does not establish better analytical outcomes. The first separately commissioned regenerated analysis is the outcome check. Historical sets and stopped runs remain unchanged and outside the current population.