fragility-grid memory profile

Type: agentic-system-analyses/types/agentic-system-memory-profile.md

Comparison rationale

The comparison boundary is accumulated measurement and reporting state, not every byte used by inference. The unchanged model parameters and static format instructions do not become learned memory merely because later calls reuse them. Conversely, opaque dependency caches remain unresolved alternatives rather than being classified from their path names or omitted to obtain complete coverage. Explicit outputs, their admission routes and these cache limits are established by RT-RTE-record-analysis, RT-CMP-engine and RT-CMP-roster-model.

Summary existence is the consequential retained part for later collection. The model key selects a matching marker, and availability alone then determines a hard skip. Identifier selection and the coarse existence predicate are separate operations in that chain. This differs from CPU analysis explicitly requesting the bit mappings and legend. Neither the CPU's item-ID intersections nor its downstream tiny-subset comparison is unsolicited model-context retrieval (RT-BAP-summary-skip, RT-RTE-sweep-driver, RT-RTE-record-analysis, MEM-OBJ-config-legend).

The summary chain supports wired trace learning in the profile's write-route sense: model-call outcomes become bits, automatic software produces a durable summary, and later shell operation depends on its retained existence. The learning force belongs to the update consumer; routing/enforcement belong to the later marker consumer. This does not establish a model parameter update, an improved capacity, or exercised self-improvement. The original inference execution feedback supports tool-traces even though this workflow has no agent tool registry. Imported gold is a separate non-trace input. No session-log, trajectory or environmental-event source is inferred merely from repeated calls (RT-RTE-generation-grid, RT-RTE-likelihood-grid, EPI-RTE-gold-comparison, RT-OBJ-summary, RT-BAP-summary-skip).

Aggregation reduces measurement content; the broader manufactured-gap interpretations add claims. Thus consolidation and synthesis coexist without treating all statistical arithmetic as new explanatory knowledge. The latter mapping rests on the accepted non-entailment finding, not on the presence of fluent text. Retained aggregate and figure outputs afford human reporting but have no established later model-guidance route (MEM-OBJ-analysis, MEM-OBJ-fragility-figures, EPI-RTE-result-labeling).

The reconciliation supersedes the duplicate aggregate declaration with MEM-OBJ-analysis; all aggregate classifications use that canonical identity. The legend remains a distinct durable part of RT-OBJ-config, and the figures remain distinct display outputs. No parent route is superseded by its narrower statistical or labeling parts.

Unknown file initialization/manual replacement prevents complete lineage and admission classification. Cache opacity prevents complete substrate, form, authority, maintenance and read-back classification. No accepted evidenced-absence declaration supports a local trace-learning no; missing qualifying chains therefore remain unresolved rather than becoming negative findings. A wired witness is retained despite those limits, and afforded reporting is not promoted to observed use. All bases describe accepted implementation or affordance, not verified activation or benefit (RT-RTE-record-analysis, RT-OBJ-records, RT-OBJ-summary, RT-CMP-engine, MEM-OBJ-analysis).

Source-read log: no additional source files were read. This profile uses only the supplied accepted members and their canonical records; it declares no records and adds no quotations or evidence.