AgentFly reconciliation

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

Reconciliation

The four reports share the frozen boundary and its Source register. They characterize shipped implementation and attributed design claims, not an observed execution or causal experiment. AgentFly remains the target name despite the repository's Memento terminology. This reconciliation uses only the supplied members; no source reread was needed. No analyst correction answers or declined objections were supplied to this attempt.

Duplicate identity and retained containers

Amendment: EPI-OBJ-training-entry is superseded by MEM-OBJ-training-data. Both declare the same parametric driver's training_entry rows in training JSONL: current query, retrieved case and plan, source case label, and current answer's truth_label. Identity evidence is the memory record's writer quotation and the epistemic EPI-RTE-outcome-label quotation of that same writer loop, SRC-1, client/parametric_memory_cbr.py. Both identify subsequent consumption by memory/train_memory_retriever.py. The amendment connects the epistemic inventory and retention ledger to the memory artifact's more detailed account; it does not change either report's label-provenance finding or strengthen its status. Both IDs remain declared.

EPI-OBJ-case-entry covers both drivers' reduced case artifacts. MEM-OBJ-memory covers its parametric memory_entry part, and MEM-OBJ-dummy-memo covers its nonparametric mem_entry part. Their writer quotations, SRC-1, client/parametric_memory_cbr.py and client/no_parametric_cbr.py, establish this overlap. Neither memory record is identical to the combined epistemic referent: substrates, schemas and downstream selectors differ. The combined record remains valid alongside these parts, so no split supersession is warranted. Its common finding—latest-plan retention with outcome metadata rather than full evidence—survives the separation. Selector-specific findings remain with MEM-RTE-classifier-examples and MEM-RTE-embedding-examples, not inferred from the combined object.

RT-RTE-case-admission remains a route container, not an alias for its objects or subroutes. Its distinct parts include the two case banks, MEM-OBJ-query-record, the training pairs, both example selectors, EPI-OBJ-model-output, EPI-OBJ-judge-result, EPI-RTE-answer-check and EPI-RTE-outcome-label. The answer and judge result are parts of the richer result artifact, not replacements for that artifact. The check compares prediction with references; the disposition consumes the normalized judgment for scoring and labels. Neither is identical to persistence or selection. Their quoted judge criterion, normalization and writer branches in SRC-1, client/parametric_memory_cbr.py, support the functional separation.

MEM-OBJ-checkpoint is the persisted numerical output of RT-RTE-retriever-training, not the training route or RT-CMP-case-retriever component. EPI-RTE-case-format overlaps admission, retrieval and training through deterministic text preparation; it is not identical to any whole route. Its formatting quotations in SRC-1, memory/train_memory_retriever.py and memory/parametric_memory.py, identify a separate transformation. No supersession follows from these overlaps.

MEM-OBJ-workspace-files and MEM-RTE-workspace-read are retained-file and display parts of RT-RTE-terminal-execution, not duplicate declarations of shell execution. MEM-OBJ-tool-files describes separate document/media side files. MEM-RTE-retrieval-display describes a standalone CLI, not driver example delivery. EPI-RTE-search-import, EPI-RTE-media-answer and EPI-RTE-math-result are distinct returning operations, not aliases for crawl extraction, model components or terminal execution. Their declarations identify separate source paths and consumers. Epistemic claim records have no counterpart in the other members' empty Claims sections. Workflow declares no records and annotates the existing ordinary query route. No other duplicate identity is established.

Independent convergence and consequential distinctions

The workflow account independently reads the ordinary client and converges with RT-RTE-query-execution: Python serializes delegation, executor text terminates a task, and the planner receives final task summaries rather than full tool conversations. Workflow's quoted replanning and print branches, SRC-1, client/agent.py, clarify that the planner runs again after the whole plan, and normal delivery can print an error or exhausted plan. This agrees with the epistemic ledger's distinction between parsing/completion admission and semantic answer acceptance. Benchmark judgment is a separate shipped path, not a missing gate that these reports insert into the ordinary workflow.

Memory's reduced-case lineage and epistemic's answer/plan scope distinction converge without making outcome labels meaningless. EPI-RTE-answer-check assesses semantic agreement with dataset references; EPI-RTE-outcome-label accepts that judgment for bounded scoring and labeling. MEM-OBJ-query-record retains richer traces and rationale, while the case banks omit those reasons. Both signs are retained. The check therefore has a consequential consumer, but neither plan validity nor each retrieved case's causal utility follows from it. Copied whole-answer labels in MEM-OBJ-training-data support the same limitation for trainer evaluation.

RT-BAP-case-guidance, both memory selectors and the epistemic ledger distinguish user-role example guidance from numerical checkpoint authority. The trainer changes retriever parameters; case reload changes the text pool, not those weights. Their implementation findings coexist with uninspected operation and unestablished benefit. EPI-CLM-experience-learning, EPI-CLM-weight-updates and EPI-CLM-automated-retraining retain public claims as claims. The epistemic comparison with implemented routes is not a cross-report disagreement: the runtime likewise distinguishes inference from retriever training and automatic collection from operator-started training.

Integration-issue dispositions

Exceptional history carryover: Memory's annotation on RT-OBJ-shared-history, SRC-1, client/parametric_memory_cbr.py, qualifies the parametric driver's exceptional path: failure before normal-return clear can leave history for the next query. The declaring runtime record already limits its clearing assertion to normal return. Workflow's reset finding applies to the ordinary driver. These scopes are compatible; no correction or supersession is needed. The exception does not establish durable authored case memory, and a synthesis cannot extend ordinary reset semantics to every parametric continuation.

Training configuration: Memory's annotation on RT-RTE-retriever-training quotes the CLI's false-by-default --use_plan and README's explicit inclusion of that option, SRC-1, memory/train_memory_retriever.py and README.md. Runtime describes possible configuration divergence, not omission by the documented recipe. The annotation mitigates that warning for the recipe without contradicting it. Configuration alignment remains a condition, not a guaranteed property of every training invocation.

Withdrawal and recovery: MEM-RTE-classifier-examples adds the missing-file guard's stale-pool behavior, SRC-1, client/parametric_memory_cbr.py. RT-RTE-case-admission says files are externally editable and appends reload cases; it does not guarantee immediate withdrawal after deletion. The added finding is compatible, but prevents treating physical deletion as immediate removal from live selection. The result-file restart consumer in MEM-OBJ-query-record also confirms runtime's nontransactional-write warning: a persisted result can suppress retry after later admission failure.

Source uncertainty and coverage: Initial corpus genealogy, exact checkpoints, external model versions, file reuse and provider state remain unresolved at their named records. Epistemic's unassessed converter and alternative-tool paths are limits on its conclusions, not contradictions of memory's separately inspected file persistence. Architectural wiring, observed activation and causal gains remain separate. No report relies on a corrected-away finding in the supplied versions.

No unresolved cross-report conflict is established by these pinned members. No required undeclared part was identified in the connections above. This states compatibility of their bounded findings, not semantic acceptance of each report or resolution of their source uncertainties; those remain for verification.