SAGE reconciliation

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

Reconciliation

Boundary and connections

The members share the frozen boundary and Source register. Runtime traces training, evaluation, expert generation and SFT. Memory adds retention, selectors and context controls. Epistemic separates transformations, checks, admission and authority within these routes. Workflow selects normal evaluation; it does not combine the training and expert routes into a single workflow. All distinguish implementation from observed operation and causal benefit.

No duplicate identity requiring supersession is established in these pinned records. The comparisons below retain valid containers, parts and distinct products. No Amendment is proposed. No cross-report disagreement requiring an Unresolved conflict: paragraph was identified. No analyst correction answers or declined objections were supplied; the runtime's revised reachability finding is assessed from its current member, not from an inferred correction rationale.

Training reachability and dependent findings

RT-RTE-paired-rollout and RT-RTE-policy-update now explicitly distinguish internal wiring from reachability through the shipped two-node launcher. The ordinary-invocation account traces AppWorld_SAGE through project-name dispatch to the Parquet reader, which is incompatible with the supplied scenario-text input. This is a stated implementation mismatch before training, not merely missing run evidence. Correcting loader selection is necessary for this launch-to-loop progression and does not demonstrate a successful run. The four-node launcher remains uninspected rather than inheriting that finding.

Memory's policy-update annotation describes the internal outcome-to-optimization-to-later-inference chain and checkpoint retention. Epistemic's authority-route ledger classifies adaptation as implemented and its bounded conclusion describes an improvement-directed architectural pathway, while withholding exercised self-improvement and learning benefit. These findings require the internal capacity that the runtime still declares; they do not assert that the shipped two-node configuration reaches it. Their omission of the dispatch mismatch is therefore not a contrary reachability finding. The runtime's qualification must remain attached whenever these connections are used: implemented internal adaptation is not an operational end-to-end launcher claim. The boundary's initial launcher-to-trainer description identifies shipped code connections, not successful progression through dataloader construction.

Identity comparisons

• RT-OBJ-skill-libraries, RT-OBJ-skill-library-functions and MEM-OBJ-rollout-functions are different objects. They respectively have paired-rollout, continuing normal-evaluation and expert-scenario lifecycles. Common Python payloads or similarly named variables do not identify their storage, admission or expiry as one referent.
• MEM-OBJ-skill-embeddings, MEM-OBJ-query-list and MEM-OBJ-query-embeddings are access parts of the evaluation container, not duplicate executable-source declarations. They also differ from RT-CMP-retrieval-model: vectors and query associations are retained products, whereas the component encodes inputs. The execution-facing container remains valid alongside these parts.
• EPI-RTE-function-extraction isolates source slicing within RT-RTE-skill-install. It does not replace dictionary admission, installation or prompt delivery. Memory's annotation on the paired object adds slicing details without declaring another extraction route. The similar evaluation helper does not make training and evaluation disposition identical.
• WF-RTE-failed-code-filter is a next-step filtering part of RT-RTE-evaluation-loop, not a duplicate task loop. Memory discusses the same mechanism in an annotation rather than declaring a competing record.
• MEM-OBJ-rollout-functions and MEM-OBJ-expert-dataset are products inside RT-RTE-expert-admission and RT-RTE-sft-admission. Their cross-kind containment does not identify objects with producing routes. Raw logs, selected messages, functions and weights remain different products.
• MEM-OBJ-llm-calls and MEM-RTE-response-cache share a label but identify response envelopes and their optional memoized return route respectively. Neither duplicates the expert corpus, RT-CMP-language-model, or the task loop. Normal evaluation disables the exposed cache branch.
• MEM-RTE-api-context covers API-context selection and foreground/retry context in the registered full-code alternative. EPI-RTE-api-prediction isolates prediction, filtering and documentation selection; EPI-RTE-code-retrial isolates feedback-fed code generation. Shared passages connect stages but do not establish identical route boundaries. The broader memory account remains valid, so split supersession is not warranted. EPI-OBJ-predicted-apis is current-task namespace data, not retained functions or query history.
• EPI-CLM-self-evolution concerns improvement framing; RT-CLM-function-rules concerns requested function properties. They are not duplicate claims. RT-BAP-function-execution, RT-BAP-function-guidance and RT-BAP-policy-weights likewise retain distinct consumers, channels and force.

Convergence on admission and delivery

RT-RTE-skill-install, memory's annotations and EPI-RTE-function-extraction connect non-failure checking, syntactic extraction, dictionary insertion and second-world installation. Runtime distinguishes prompt rules from enforcement. Epistemic supplies source-deduction counterexamples to universal semantic preservation; memory identifies omitted globals, decorator and multiline-import limitations. These findings do not assert that every extracted function is damaged. The training-reachability condition above remains separate from these internal mechanisms.

All three accounts separate interpreter delivery from model guidance. Code executes before a later output check controls prompt visibility. Memory distinguishes unsuccessful-response suppression from an exception path without guaranteed suppression; this refines runtime's statement that failed installation can suppress prompt text, rather than contradicting it. Clearing prompt text is not rollback of attempted effects.

For normal evaluation, RT-RTE-evaluation-loop, memory's annotation and WF-RTE-failed-code-filter agree that default extraction does not require current-task success. Optional label gating uses the previous task's result. Memory and workflow independently describe delayed failed-code deletion and the final-step gap. Runtime does not claim that every failing chunk is excluded. Substring-based scenario selection in memory and workflow refines, rather than opposes, runtime's identifier-based account.

Workflow's evaluation-loop annotation distinguishes supervisor completion, loop termination, benchmark acceptance, library extraction and processed-task logging. Epistemic's ledger separately traces task-success arithmetic and extraction disposition. Neither a completion fallback nor a progress counter becomes evaluator acceptance in these reports.

RT-RTE-expert-admission has different gates: benchmark success, nonempty first-task extraction and fresh-world definition execution. MEM-OBJ-rollout-functions describes the admitted product. Epistemic identifies evidence-consuming acceptance of benchmark-successful trajectories for training-data use, not generalized function correctness. This does not conflict with the reports' denial of per-function correctness acceptance in ordinary evaluation: targets, criteria and intended uses differ.

Integration-issue dispositions and remaining limits

Memory's Integration issues proposes no supplied-value correction or duplicate. Its delayed-pruning, substring-matching and installation-exception refinements remain connected above. The registered full-code alternative is additional coverage, not a claim that normal evaluation uses API prediction. Epistemic independently covers that alternative without treating its name as evidence of beneficial reflection.

The restart gap in MEM-OBJ-query-list and MEM-OBJ-query-embeddings is local: the writer saves vectors without the corresponding query-list serialization, and n-gram initialization does not load that list. This adds to the runtime's separate output-versus-predefined-input distinction; it does not deny same-process reuse. Unknown predefined provenance, compatibility and alignment coexist with known append and selection behavior. First-name deduplication can hide a later appended function; runtime's within-extraction merge does not claim library-wide replacement or coherent deletion. This content behavior is not supersession of analysis records.

RT-RTE-sft-admission and MEM-OBJ-expert-dataset agree that the live extractor takes the last LM call's messages plus response. The unused helper's slice is not exported-context evidence. Epistemic adds limits on semantic preservation through normalization, not a competing executed path. Background-generator readiness, temperature-directory mismatch, token cutoff/masking and the operator-selected SFT-to-RL bridge remain distinct limits on an automatically successful pipeline.

RT-RTE-policy-update, memory's parameter-retention account and epistemic's bounded conclusion retain separate conclusions about internal update wiring, exercise and benefit. Lexical reuse counts are not invocation traces or causal reuse evidence. Checkpoint recovery is not AppWorld conversation recovery. Missing weights, trajectories, downstream dependence and comparisons limit attribution without erasing implementation findings; the launcher mismatch separately limits reachability. These uncertainties are not disagreements to settle by choosing a stronger status.

No missing part requiring a combined-record conflict was identified among these connected declarations. No source reread was needed; all connections above use the pinned members and their retained evidence.