Napkin reconciliation
Type: agentic-system-analyses/types/agentic-system-reconciliation-report.md
Reconciliation
Shared boundary and independent convergence
The four reports use the frozen source and evidence layers in the boundary's Source register. They connect filesystem capabilities, external-agent instructions and shipped benchmark callers without treating Napkin as the enclosing model runtime. Implementation, requested model behavior, observed operation and demonstrated benefit remain separate. No report supplies execution or causal evidence; this agreement does not establish that the described routes operated.
RT-RTE-vault-retrieval, its memory annotation and the epistemic ledger converge on request-driven delivery with ranking and freshness controls, not source-truth endorsement. MEM-OBJ-search-cache and MEM-OBJ-overview-cache identify distinct retained inputs to that route. Path/mtime fingerprints, reread snippets and overview fallback bound freshness and keyword-validation claims in all three accounts. RT-CLM-keyword-probes is best effort, not a universal top-three retrieval guarantee.
RT-RTE-distill-session, RT-RTE-tend-vault, the memory annotations and the workflow account describe model-mediated selection and upkeep as policies. The epistemic account's doctrine only classification of those semantic decisions does not contradict runtime's claimed instruction status. The command executor can be wired while reliable model compliance and candidate-level acceptance remain unestablished. EPI-CLM-distill-knowledge and EPI-CLM-tend-safety assess claims about those procedures, not replacement identities for the routes.
Record identities and overlap dispositions
No full-record identity is established and no supersession is made. The following overlaps retain valid containers or distinct operations; sharing files, passages or a consumer does not make them interchangeable.
- EPI-OBJ-vault-prose spans knowledge/context prose in MEM-OBJ-markdown-notes and MEM-OBJ-napkin. The memory records separately expose general notes/access metadata and the context-note selection path. Neither alone replaces the epistemic container; the container's finding does not become wrong when those parts are distinguished. EPI-OBJ-benchmark-notes instead identifies driver-imported instances across benchmarks. MEM-OBJ-session-summary isolates their LoCoMo derived-content part and opaque upstream lineage. These scopes overlap but are not identical.
- EPI-OBJ-answer-score isolates answer, reference and score parts of benchmark results. MEM-OBJ-evaluation-results additionally covers retrieval proxies, timing/counts, persistence and aggregation. Their valid container/part relationship requires no supersession. EPI-OBJ-base-values is returned query data, not the retained definitions in MEM-OBJ-base. EPI-OBJ-link-index is transient diagnostics, not either retrieval cache or stored note content. None is identical to RT-OBJ-vault-info, which locates stores and configuration.
- MEM-RTE-auxiliary-edits groups separately exposed daily, property/task, template-insert and base-create writes. EPI-RTE-property-edit and EPI-RTE-template-insert isolate operations within that coverage; neither replaces the wider route. MEM-RTE-structured-access includes base projection and canvas/bookmark operations. EPI-RTE-base-summary isolates numeric aggregation, not the complete structured-access route. These also differ from the ordinary CRUD and overview/search/read routes.
- MEM-RTE-benchmark-variants covers LoCoMo/HotpotQA ingestion, later answering and evaluation. EPI-RTE-dataset-import, EPI-RTE-answer-extract and EPI-RTE-token-score isolate different consequential functions in those drivers. No part replaces the valid workflow container or RT-RTE-benchmark-answer, which covers LongMemEval. EPI-RTE-longmem-judge and EPI-RTE-score-report expressly remain parts of that LongMemEval container: checking and consuming its result are separate functions.
- WF-RTE-distill-verification includes the post-write diagnostic, CLI serialization and agent repair handoff within the distill parent. EPI-RTE-link-check covers the structural diagnostic callable outside that workflow too, without the same repair progression. Shared link-index implementation is not identity between these routes. Both retain the wikilink-only, stripped-heading and non-semantic scope of the check.
Integration-issue dispositions
Background distillation remains distinct and unresolved in its external wiring. MEM-RTE-background-distill describes separate-model/extension doctrine; RT-RTE-distill-session describes the packaged agent procedure. Identity is ruled out on the supplied evidence, not assumed from their shared retention purpose. EPI-CLM-background-distill is a claim about the former, not another declaration of its route. All reports preserve the unavailable extension implementation. Timer/cursor operation, enforcement and automatic context injection remain limitations, not cross-report disagreements. MEM-OBJ-napkin establishes requested overview delivery, not a verified unsolicited push route.
The exposure scorer correction does not invalidate keyword selection. The memory integration issues retain the caller/signature mismatch in SRC-1, bench/overview-exposure.ts and src/core/search.ts. That finding prevents treating the scorer as verified executable validation. RT-CLM-keyword-probes concerns the separate overview implementation and already retains fallback exceptions. The reports do not assert conflicting findings about the scorer.
Withdrawal and write guards remain branch-scoped. RT-RTE-caller-mutation says ordinary trash deletion hides notes from listings. MEM-OBJ-markdown-notes adds exact-path access to retained trash bytes; this qualifies withdrawal without contradicting listing exclusion. Likewise, the existing-target create guard does not extend to every auxiliary writer. MEM-RTE-auxiliary-edits preserves baseCreate's unguarded write and daily/template transformations. Neither requires replacing the CRUD route or strengthening its guarantee.
The workflow adds instruction/executor mismatches, not a contrary runtime finding. Its mutation annotation supplies the distill example's folder-valued --path, the executable's complete-file-path interpretation and template replacement of supplied content. RT-RTE-caller-mutation already records caller paths and template replacement; MEM-OBJ-note-templates distinguishes copy from insertion/substitution. Literal execution of the example therefore cannot certify the intended titled, populated note. This delivery limitation remains alongside requested distillation; no report demonstrates successful completion that would contradict it.
Benchmark checks do not admit knowledge. EPI-RTE-longmem-judge adds substring counterexamples and continuous fallback; the runtime's broader normalized-match/model-judge/fallback account does not promise semantic correctness. EPI-RTE-answer-extract and EPI-RTE-token-score add preprocessing and overlap limits to the memory benchmark account. Results affect reporting, not an inspected score-to-note or policy update. EPI-CLM-benchmark-results scopes the no-summaries claim to LongMemEval, so the LoCoMo summary import is not a contradiction. Different inspection depth for aggregation, cleanup and auxiliary internals remains local uncertainty, not competing statuses.
No unresolved cross-report conflict is established by these pinned members. No correction answers or declined objections were supplied. The integration limits above remain for verification; this reconciliation neither judges individual record support nor settles missing execution, provenance or external-host evidence.