Apache Maka reconciliation

Type: agentic-system-analyses/types/agentic-system-reconciliation-report.md

Reconciliation

Shared boundary and evidence

The four reports use the whole-system boundary and SRC-1 fixed by the boundary. Their implementation findings are not observed runs or causal experiments. The memory and epistemic reports expand the runtime's initial inspection without changing its source identity. The workflow selects one foreground delegation family rather than asserting complete workflow coverage. No source reads were needed for this reconciliation; evidence below is supplied by the cited records. No correction answers or declined objections were supplied to this attempt.

Duplicate identity and distinct operative parts

Amendment: EPI-RTE-local-memory-use is superseded by MEM-RTE-local-memory. Both describe the same durable Local Memory mutation-to-prompt route in SRC-1 packages/runtime-host/src/server/memory-coordinator.ts and packages/core/src/local-memory.ts: pending/active lifecycle mutations, policy and session-scope selection, bounded prompt delivery and withdrawal from subsequent projection. The epistemic record explicitly includes the mutation branches as well as prompt consumption; it is not a separate downstream evaluator. Both retain unresolved outer caller identity and distinguish approval metadata from evidence-consuming semantic acceptance. This identity connects the epistemic ledger and bounded conclusion to the memory route; it does not replace either report's fields or strengthen their statuses. Both declarations remain available, including the epistemic account of the limited authority granted by approval metadata.

EPI-OBJ-local-memory is a valid bundle-level container, not an identical declaration of either MEM-OBJ-local-entries or MEM-OBJ-bundle-metadata. The epistemic object includes pending and active Markdown content and lifecycle/scope metadata; memory separates model-interpreted bodies from fields consumed by software admission and eligibility checks. Its claim that only selected active content reaches the prompt remains compatible with those parts. No split supersession is warranted merely because memory provides finer parts. None of these objects is RT-OBJ-run-composition, which records composition identity and hashes rather than entry content or eligibility.

EPI-OBJ-memory-item likewise spans transient proposal/canonicalization payloads and admitted writes, including text and indexing/evidence fields. MEM-OBJ-memory-item identifies retained item content, while MEM-OBJ-item-index identifies keys, lifecycle and extraction control. These are overlapping container/part scopes, not interchangeable whole referents. Neither memory part replaces the epistemic object's transient proposal scope. Its mixed description does not assert that all metadata reaches the later model; both reports explicitly describe omission at Recall. The container remains alongside the declared parts, without supersession.

Extraction, checks, disposition and recall

MEM-RTE-fact-extraction gives the broader trace-to-fact route and its triggers. EPI-RTE-memory-evidence, EPI-RTE-memory-canonicalize, EPI-RTE-memory-disposition and EPI-RTE-memory-commit distinguish consequential stages within extraction. Their targets and endpoints differ: lexical/source admission, model support judgment, write selection and durable commit. They are not duplicate routes, nor does the valid broader extraction account require supersession by its stages.

The reports converge on computational admission without per-item Markdown-queue approval. Memory's caution that raw-event provenance does not prove canonical entailment is compatible with epistemic's positive finding of a fallible, evidence-consuming acceptance decision. The latter licenses reuse as a supported user assertion, not independent external truth. It leaves actual output entailment indeterminate. These are different conclusions, not opposing findings.

MEM-RTE-recall-history supplies the later retrieval consumer named by the extraction and commit accounts. Both analysts identify the content/kind/time projection and omission of stored citations and canonicalizer rationale. Persistence of lineage is therefore not a claim that the model receives that lineage. Retrieval is distinct from the preceding acceptance and retention stages; no identity supersession is appropriate. The default active fact lookup described by memory adds inspection to the epistemic commit record's lifecycle uncertainty rather than contradicting it. Full edit/withdrawal caller authority remains unresolved in memory as well.

Compaction, reload and guidance

RT-RTE-context-projection remains the valid effective-history container. MEM-OBJ-text-checkpoint and MEM-OBJ-provider-checkpoint are explicitly declared alternative retained outputs, not competing descriptions of one encoding. Their different consumption channels prevent treating the encrypted payload as a readable summary. The memory annotation adds replay details without changing the runtime route's referent.

MEM-RTE-archive-read explicitly declares the reload branch of that route. It partly resolves runtime's archive-retrieval inspection gap through canonical event-address Read and ledger reconstruction. The epistemic report's archive uncertainty records its own inspection limit; it does not deny that branch. No conflict or supersession follows. The remaining legacy/backend coverage limits prevent universal reload or losslessness claims. All three accounts keep structural coverage checks separate from semantic fidelity.

MEM-BAP-local-guidance explicitly narrows RT-BAP-composed-prompt to workspace and skill delivery. Its budgets and discovery/loading controls supplement, rather than duplicate, the overall prompt path. Both distinguish interpreted guidance from executable permission authority in RT-BAP-tool-boundary. Static skill delivery and editable workspace instructions do not, in these reports, establish an evaluated self-revision loop.

Delegation, publication and delivery

The memory and workflow annotations independently connect child isolation and explicit output return to RT-RTE-child-graph. They agree that children do not implicitly inherit local memory or personalization, selected output crosses back, and durable child identity survives. Neither annotation declares a duplicate route.

WF-RTE-patch-publication explicitly identifies a terminal worktree handoff within that container. It adds quiescence, capture, publication and recovery details to a branch runtime had located but not fully traced. It does not supersede spawning, Graph admission or output lookup. Artifact publication is not parent application or semantic acceptance. Workflow's final delivery criterion still belongs to RT-RTE-cli-turn, after continuation through RT-RTE-tool-settlement. Its structural child-result checks are not an independent task-quality verifier.

Distinct evaluators and retained-state loops

EPI-RTE-goal-judge and EPI-RTE-goal-selection add an evaluator and consequential continuation control to runtime's expressly uninspected goal coverage. They are not RT-RTE-durable-wake: a goal judgment and its next-turn selection differ from scheduled-fire identity and recovery admission. Workflow's lack of a mandatory independent verifier concerns its selected child workflow, not every Maka route. Thus the goal evaluator is not a contradiction of that bounded finding.

MEM-RTE-matter-state, MEM-OBJ-matter-state and MEM-OBJ-matter-reviews add a separate objective-state, settlement-review and resumption loop. Its review is not goal continuation or generic scheduled admission. Intermediate checkpoint admission, final settlement and immediate rejection feedback retain different horizons; the retained review journal does not establish that every future wake consumes its narrative.

MEM-RTE-index-maintenance concerns authored index bodies and range/revision-controlled coverage, not generic child execution. MEM-RTE-history-extract instead produces independent extraction files without changing index content or coverage. Neither is identical to core fact extraction. MEM-OBJ-review-archive is a generated daily activity product with a later client query, not an established automatic behavior-revision route. These positive additions coexist with epistemic's stated noninspection of assistant and daily-review paths; that uncertainty is not an absence finding.

EPI-RTE-eval-verification and EPI-OBJ-verifier-result add external score import and operative reporting classification, not a foreground answer oracle or local proof of task truth. They qualify RT-CLM-task-runtime without duplicating that public claim. EPI-CLM-goal-judge is also distinct: its claimed prevention of premature completion concerns the separate goal judge, not the public event-authority and benchmark contract. Neither claim gains observed or causal support from source wiring.

Integration dispositions and remaining limits

No unresolved disagreement between the supplied reports was identified. The local-memory route identity above is the duplicate supersession; the other material overlaps are valid containers, distinct parts, annotations or independent convergence. No required missing part was established by these connections. There are no supplied declined objections to adjudicate, and this reconciliation makes no support judgment on an individual report.

Memory's integration issues remain local limits: outer human approval identity, fact withdrawal authorization, index revocation/deletion propagation, future Matter-review consumption, and uninspected retained-state loops. Epistemic adds bounded goal and Eval findings but does not close all Goal/Todo/WorkHub recovery and reuse questions. RT-RTE-plugin-admission supplies capability/configuration admission, not knowledge-quality admission or proof of self-improvement; alternate execution through RT-RTE-executor-plugin also prevents extending local tool guarantees universally. Across all reports, implementation, model compliance, successful operation and benefit remain separate. These unresolved inspection and execution limits prevent whole-system completeness, universal evaluator/security guarantees and exercised-improvement conclusions without creating a cross-report conflict.