Compound Engineering Plugin reconciliation

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

Reconciliation

Frozen account and evidence limits

The four reports share the extension boundary and sole frozen source in boundary and its Source register. They distinguish executable support programs from model-facing workflow policies and from the excluded enclosing hosts. None supplies observed target execution or causal improvement evidence. The whole-system functional boundary is not an exhaustive behavioral audit: the runtime, memory and epistemic coverage accounts retain available but uninspected paths. Workflow deliberately selects the native LFG plan-to-PR route. Those limits are compatible, not evidence of absent mechanisms.

No analyst correction answers or declined objections are supplied to this attempt. The dispositions below compare the supplied reports themselves; they do not infer a correction history. No source reads were needed beyond those reports.

Duplicate identities

Amendment: EPI-OBJ-solutions is superseded by MEM-OBJ-solutions. Both identify project learning Markdown under <root>/solutions/, produced by capture and consumed by later learning research. Their assembly/template and researcher evidence identifies the same retained corpus; memory also describes derived pattern guidance. The epistemic inventory and capture/refresh ledger references therefore concern this corpus, not a second store. Its candidate-versus-accepted assessment remains an epistemic finding; this identity does not grant all retained prose acceptance.

Amendment: EPI-OBJ-pack-rules is superseded by MEM-OBJ-pack-rules. Both identify top-level Markdown rules in declared local or Git-backed pack roots, with title/applicability metadata and researcher consumption. Their shared researcher evidence separates rule content from resolver access metadata and workflow instructions. The epistemic acquisition and reliance findings concern these same rules. Memory's writable-path capture alternative and unknown third-party authorship remain beside epistemic's unknown imported warrant; neither replaces the other.

Amendment: EPI-OBJ-concepts is superseded by MEM-OBJ-concepts. Both identify project-root CONCEPTS.md, its model-authored definitions and behavior assertions, and later planning terminology consumption. Their assembly/research evidence identifies the same independently edited glossary. The epistemic prewrite defining-source check and Full/Lightweight distinctions qualify the memory object's admission account without changing its identity or asserting observed validation.

Amendment: EPI-RTE-pack-resolution is superseded by MEM-RTE-pack-resolution. Both trace declared sources through the same inspected capture resolver's URL/ref-keyed Git cache to absolute roots for later rule readers. Both retain the trusted-cache immediate return and lack of established refresh on that path. Epistemic's acquisition/import function and unknown source warrant apply to this same route, not another acquisition mechanism. Both implementation statuses are wired and operation statuses uninspected; no status selection is necessary. Their uninspected enumeration/parser branches and per-skill copies still prevent a universal resolver guarantee.

Amendment: EPI-RTE-learning-delivery is superseded by MEM-RTE-learning-retrieval. Both identify planning's glossary-informed researcher selection, relevance filtering, up-to-five distilled findings, scratch dossier and foreground planner consumption, supported by the same planning research and researcher passages. Memory additionally spells out inline Lightweight delivery. Epistemic explicitly permits an inline result and describes the researcher/planner consumer part of the broader admission container; it does not establish a separate planning consumer route. Its content-transformation uncertainty and operational-consumption ledger rows remain assessments of this route. Both implementation statuses are claimed and operation statuses uninspected; dossier delivery is not observed decision dependence.

These supersessions resolve duplicate referents, not competing evidence values. Both original declarations remain available. No other declared pair establishes identity merely by sharing source paths, producers or consumers.

Containers, material parts and distinct routes

RT-RTE-learning-admission remains a valid capture/maintenance container with prescribed later planning delivery. MEM-RTE-session-capture, EPI-RTE-capture-grounding and EPI-RTE-refresh-curation declare material parts rather than duplicate whole routes. Session enrichment selects and reduces archives; capture grounding checks and adjudicates written claims; refresh investigates and changes an existing corpus. Their different targets, timing and consumers rule out mutual supersession. The planning route supersession above does not supersede this container.

MEM-OBJ-session-skeleton and EPI-OBJ-learnings-dossier are distinct scratch derivatives. The former reduces prior host sessions for a capture historian; the latter distills stored learning for a planner. Neither is the durable solutions corpus. MEM-OBJ-handoff and MEM-RTE-handoff-resume supply selected continuity and orientation, not controller authorization or automatic action permission.

MEM-OBJ-experiment-log, MEM-OBJ-strategy-digest and MEM-RTE-experiment-digest are optimization evidence and next-batch guidance, not duplicates of RT-OBJ-run-manifest or RT-RTE-external-unit. The latter bind implementation dispatch/integration and recovery. A generic optimization target also need not be the plugin's own methodology, so its theory and reflection findings cannot be transferred from RT-RTE-skill-authoring.

MEM-RTE-review-retrieval, EPI-RTE-review-findings and WF-RTE-review-admission are connected but distinct. Memory traces historical solution/rule selection into review candidates or Known Pattern notes. Epistemic traces review mechanics and semantic disposition, including report-only default and separately authorized local apply. Workflow traces LFG's caller-owned mechanical-fix filter, tests, commits and residual delivery. Thus a rule contradiction is a candidate, confirmation grants no new apply authority, and LFG applies only its eligible subset. The workflow route remains a part of RT-RTE-lfg-shipping, not a replacement for the whole shipping pipeline.

EPI-OBJ-review-findings, EPI-OBJ-debug-findings and EPI-OBJ-work-evidence differ in production and use: diff-defect disposition, symptom/causal investigation, and canonical implementation acceptance. EPI-RTE-debug-investigation is not native implementation, and the work-evidence object is explicitly inside RT-RTE-native-work, not an external transaction receipt.

EPI-OBJ-eval-artifacts and EPI-RTE-eval-grading provide scenario evidence within repository-local authoring. They do not duplicate the method-revision decision in RT-RTE-skill-authoring. EPI-RTE-pi-conversion transforms host-facing instructions and is reused outside install; EPI-RTE-legacy-cleanup withdraws old resources. Neither is identical to RT-RTE-resource-install or RT-RTE-host-loading.

RT-OBJ-skill-files and RT-BAP-skill-instructions concern shipped instructions reaching models. MEM-BAP-project-evidence concerns retained project evidence reaching decision consumers. RT-BAP-controller-state concerns fixed symbolic transition checks. These are not interchangeable authority paths. The two model components, RT-CMP-host-model and RT-CMP-external-model, retain separate host/recipient identities and opaque provider mutability, not a uniform pinned-model claim.

Integration dispositions and independent convergence

Grounding is postwrite and mode-specific. Memory's qualification of RT-RTE-learning-admission is corroborated by EPI-RTE-capture-grounding: retention precedes grounding, flags are adjudicated, neither pass is an unconditional hard stop, and Lightweight omits semantic validation. Runtime's broad eligibility/grounding rejection account does not state an atomic prewrite gate. Its claimed status therefore need not be replaced by epistemic's wired status, which expressly belongs only to the mechanical script. The ledger separately labels semantic acceptance doctrine only. EPI-CLM-trusted-knowledge is a stronger source warrant claim being compared with those routes, not an analyst assertion that every persisted document is certified.

Withdrawal and uncertainty differ. Memory's stale-status issue and epistemic's EPI-RTE-refresh-curation agree that age or lack of corroboration does not establish falsehood. Deletion changes the live corpus; stale metadata does not establish automatic planning exclusion. MEM-RTE-learning-retrieval supplies the current-evidence conflict safeguard. These limits prevent equating stale marking with withdrawal, not reconciliation of the maintenance findings.

Cache and lineage limits stay local. The duplicate pack-resolution records independently converge on the trusted-cache return, unknown external pack warrant and mutable-ref freshness limit. MEM-RTE-session-capture retains lossy active-path/compacted-tail extraction and shortened summaries; epistemic does not independently audit those extractors. Its transformation uncertainty does not contradict that supported trace-fed write chain or establish that all reasons are preserved.

Child completion is not caller delivery. Workflow's annotations on RT-RTE-native-work and RT-RTE-lfg-shipping add a one-time evidence reconciliation path, caller-level require route-mismatch veto, and distinctions between watch-started and CI-decided, local-only delivery and failed shipping, and needs-human handoff and successful completion. Runtime permits native fallback for an unavailable require external engine; workflow states LFG then withholds shipping for the mismatched completed return. These findings concern different admission points, not incompatible fallback policies. The runtime's no-merge default and host-owned canonical verification remain compatible with workflow's more detailed delivery account. Neither a DONE marker nor worker success alone establishes full acceptance.

Knowledge, product and improvement claims remain separate. RT-CLM-verified-change concerns current implementation verification; EPI-CLM-compound-loop concerns future learning benefit; EPI-CLM-pack-enforcement concerns planning/review rule use. These are distinct source claims. Mechanical controller/grader wiring cannot establish adequate behavioral tests, semantic warrant, model compliance or causal benefit. Runtime authoring and memory optimization retain independently scoped theory-builder conditions, learning, reflection and autonomy assessments. Epistemic explicitly preserves these rather than assigning a single epistemic or improvement verdict.

Included unresolved alternatives remain limitations. Memory names retune archive/write-up details, dogfood, ideation, live-polish, pending-feedback and sweep state, plus additional glossary/solution consumers whose leaf behavior is uninspected. Runtime and epistemic likewise disclose unaudited recovery, host, shipping and release paths. These coverage limits prevent exhaustive admission/read-back classifications and system-complete negatives. They do not establish an undeclared required split of one cited combined record: the supplied reports do not assert complete classifications of those alternatives.

No unresolved conflict between the supplied reports was identified. The supersessions above address duplicate identities; the remaining integration issues qualify scope, timing, consumer authority or evidence strength without selecting a stronger finding or correcting another member.