MerchantBench reconciliation

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

Reconciliation

Duplicate identity and retained containers

Amendment: EPI-OBJ-memory-doc is superseded by MEM-OBJ-memory-doc. Both identify the latest UTF-8 Markdown file selected by the same run-local, sanitized-agent path. Both retain its mixed claims/plans, opaque persistence and lack of semantic endorsement. Their identical path quotations establish identity, not merely similar function (SRC-1, env/tools/tools.py). Epistemic inventory, ledger and bounded-conclusion references to the scratchpad therefore concern the memory-declared latest document. Both IDs remain declared; this connection changes neither report's findings. The separate archive is MEM-OBJ-memory-history, not another latest-document identity.

EPI-RTE-memory-document remains a valid container alongside MEM-RTE-document-write and MEM-RTE-document-read. Its scope explicitly includes admission, retention and recall; the memory routes separate overwrite/archive production from requested latest-body delivery. Both reports distinguish byte admission from truth acceptance and separate filesystem overwrite from history append (SRC-1, env/tools/tools.py, env/tools/dispatch.py). No split supersession is warranted: separating these functions does not make the combined record's findings wrong. Neither memory route alone replaces the combined referent, and neither is a duplicate of the enclosing RT-RTE-act-gate. The declared containment of all three in that HTTP route is compatible.

MEM-OBJ-history is a state part of RT-RTE-react-loop, not another loop. MEM-OBJ-step-traces is a distinct persisted output spanning that loop and the action gate, not a replacement for either. MEM-RTE-trace-inspection describes subsequent human-facing access rather than foreground model accumulation (SRC-1, agent/baselines/react_160k_compact_30k.py, env/storage/agent_log.py, env/web/routes_agent.py). No supersession applies to these overlaps.

EPI-RTE-observation-render and EPI-RTE-wealth-metric retain their declared parts of the ReAct loop and hosted evaluator respectively. EPI-OBJ-store-snapshot is changing store evidence, not RT-OBJ-system-brief. EPI-OBJ-eval-metrics is a result object, not the enclosing RT-RTE-hosted-eval. These are containment or input/output connections, not duplicate identities (SRC-1, env/tools/observation.py, eval/scoring.py, eval/run_eval.py).

Independent convergence and delivery boundaries

Runtime, memory and workflow agree that RT-RTE-context-trim removes old messages and requests, but does not require, a successful save. Memory's annotations supply the suffix-selection and retry details; epistemic's ledger explicitly limits its reshaping classification to code deletion rather than optional model rewriting. The epistemic ledger associates this route with the scratchpad object, while runtime names local history as its input and memory declares that history separately. Read with the epistemic row's own limit, this is an association with optional preservation, not a finding that trimming edits the file; no incompatible mechanism is asserted. The shared conclusion cannot be strengthened to successful semantic retention or automatic document reinjection (SRC-1, agent/baselines/react_160k_compact_30k.py).

RT-RTE-act-gate, RT-CLM-http-controls and workflow's annotations converge on software validity checks with conditional HTTP coverage. Epistemic distinguishes that execution permission from epistemic warrant. RT-BAP-handler-binding is enforcing dispatch authority; RT-BAP-brief-guidance and MEM-BAP-document-guidance are distinct guidance paths, with different producers and delivery channels. None supersedes another (SRC-1, env/tools/dispatch.py, env/tools/registry.py, agent/baselines/react_160k_compact_30k.py).

WF-RTE-batch-runs connects queued configuration to run creation and lifecycle delivery. It is neither the inner ReAct loop, RT-RTE-other-clients, nor hosted Docker evaluation. Batch success is finished lifecycle status after the stopped/error check; hosted evaluation calculates wealth and serializes metrics. EPI-RTE-wealth-metric therefore supplies no hidden performance gate for batch completion (SRC-1, scripts/run_batch.py, eval/run_eval.py, eval/scoring.py).

EPI-CLM-horizon-evaluation concerns public benchmark purpose, not the HTTP-control claim. Its treatment of the README's “rather than complete a bounded task” wording retains both the source rhetoric and the bounded shipped scenario. This is not a disagreement with runtime or workflow's bounded-task account (SRC-1, README.md).

Integration-issue dispositions and limits

Memory reports no required runtime correction. Its sanitized-filename qualification limits universal agent isolation without contradicting the runtime's run-local document capability: MEM-OBJ-memory-doc permits aliasing for unusual IDs and does not promote ordinary-ID tests to an invariant (SRC-1, env/tools/tools.py, tests/test_memory_doc_tools.py).

Memory's unexercised save/recall chain remains an evidence limit shared with EPI-RTE-memory-document, not a dispute about wiring. Its older-version limitation concerns only the searched history consumers; no other report asserts an automatic history restore. MEM-RTE-run-withdrawal separately removes run-local artifacts, while MEM-CLM-state-boundary distinguishes simulator recovery and dashboard replay from agent summaries. These do not supersede context eviction or document overwrite (SRC-1, env/web/runner.py, env/storage/replay.py, env/tools/tools.py).

RT-RTE-hermes-launch, memory's annotation on it, MEM-RTE-trace-inspection and MEM-CLM-telemetry-limits agree on the external boundary. Profile preparation, displayed active/compacted messages and evolution counters do not establish external summary admission, skill learning or actual model consumption. The memory report's Hermes-specific denial of document tools does not reinstate the documentation's default-scenario denial rejected by the boundary: the configurations differ (SRC-1, env/web/runner.py, env/web/routes_agent.py, env/scenarios/agents/hermes.yaml, env/scenarios/default.yaml).

No unresolved cross-report disagreement is established by these pinned members. No analyst correction answers or declined objections were supplied to this attempt. Unknown external behavior, uninspected diagnostic equations and absent execution/comparison evidence remain local limitations, not conflicts or system-wide negatives. The route records' implementation and operation fields remain separate; epistemic architectural status does not replace either. The Source register supplies implementation and doctrine/design only. No fresh source read was needed for reconciliation.