SkillRL reconciliation

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

Reconciliation

Shared boundary and connected account

All four reports use the run identity and revision fixed by boundary. The Source register supplies implementation, doctrine/design and reported-operation evidence, but no observed run or causal experiment. The reports distinguish retained instructions, numerical policy changes, delivery, exercised dependence and benefit. Their local unassessed operations limit particular conclusions without changing the whole-system functional boundary.

RT-RTE-environment-rollout connects bank selection and observation rendering to policy generation and simulator feedback. RT-RTE-policy-update changes actor parameters through a separate route; RT-BAP-policy-parameters identifies their later generation channel. The validation and training admissions, RT-RTE-validation-admission and RT-RTE-training-admission, share proposal and bank-writing machinery but differ in scheduling and failure origin. They are distinct routes, not supersession candidates. RT-BAP-skill-conditioning describes advisory instruction delivery, not an executor permission or demonstrated dependence on the instructions.

The workflow account adds ordering, rejection and delivery detail to these runtime records rather than declaring replacements. In its selected ALFWorld driver, fitting precedes scheduled validation, and validation failures can change the training bank before later rollouts. Memory annotations and the epistemic ledger retain that connection. Separate managers and training-only bank writes therefore coexist with validation influence on subsequent training. No report claims causal isolation from validation evidence.

Duplicate supersession and distinct referents

Amendment: EPI-OBJ-skill-principles is superseded by MEM-OBJ-skill-content. Both identify natural-language bank titles, principles, applicability conditions and common-mistake instructions in disk JSON and live dictionaries, consumed through RT-BAP-skill-conditioning. Both distinguish this content from actor weights and the rollout batch. The epistemic record's ALFWorld example and rendering deduction assess the same bank content, not a separate retained object. The memory record also accounts for Search's category key and the updater's title-only view. This identity affects the epistemic inventory and ledger's object references; it changes neither report's semantic assessment nor its provenance and operation limits. Both IDs and their supporting passages remain declared.

MEM-OBJ-skill-metadata is distinct from that instruction content: IDs, category keys and document order are consumed by code for targeting and admission. MEM-OBJ-retrieval-index identifies goal vectors, positional correspondence and cache state, not lesson text. MEM-CMP-retrieval-encoder and RT-CMP-embedding-model share a default model name but belong to different classes, initialization paths and cache consumers. Matching model names do not establish component identity; neither supersedes the other.

MEM-OBJ-experience-records, MEM-OBJ-episode-history and RT-OBJ-trajectory-batch identify derived retained experiences, raw episode-local buffers and current-run training/transport state respectively. They are not duplicate objects. EPI-OBJ-search-content identifies imported passages, generated answer strings and dataset references. An answer's later inclusion in a training batch is a projection, not identity between that content record and the batch.

RT-CLM-reusable-patterns, EPI-CLM-recursive-evolution and EPI-CLM-context-efficiency are distinct claims about reusable abstraction, feedback-driven co-evolution and quantitative compression/utility. They do not supersede one another. MEM-CLM-search-category concerns a particular producer/consumer schema mismatch, not another declaration of one of these public claims.

Containers, parts and status reconciliation

MEM-RTE-trace-abstraction is explicitly part of RT-RTE-sft-production: it covers cleanup, per-trajectory memory generation and aggregation inside the larger pipeline. EPI-RTE-trajectory-cleanup isolates deterministic action-window deletion within the same production route. It overlaps one stage of the memory record, not its whole referent. The containers remain valid alongside the narrower records; no supersession is required. The epistemic cleanup witness and memory's deletion passage independently converge on equal actions not establishing equal observations. Neither reports an observed data-loss event or demonstrated benefit from pruning.

EPI-RTE-proposal-check spans the structural-check portion of both runtime admission routes. It is not identical to either complete admission route and supersedes neither. Its required-key passage connects to the memory admission annotations and workflow proposal handoff. Structural checking, operational admission and retention remain distinct functions even though one implementation chain performs them.

EPI-RTE-skill-withdrawal describes removal, a narrower operation within MEM-RTE-bank-maintenance, which also covers addition and saving. They are not identical routes, and the maintenance container remains valid without supersession. The current withdrawal declaration has implementation conclusion status afforded, matching the maintenance record's capability status. Both leave operation uninspected, locate no shipped removal caller in their bounded searches, and distinguish future bank membership from cached episode selections and saved copies. The prior reconciliation's wired versus afforded conflict no longer applies to these pinned records.

The epistemic ledger calls the inspected withdrawal method architecturally implemented; its bounded conclusion expressly distinguishes that method from a wired shipped withdrawal workflow. This architectural vocabulary does not promote its record's implementation conclusion status to wired. The memory finding that removal is an exposed capability and the epistemic finding that its method is implemented therefore retain compatible scopes. Neither establishes automatic quality-driven withdrawal or exercised rollback. No correction answers or declined-objection arguments were supplied to this attempt; this disposition follows the current findings, not an assumed correction rationale.

Integration-issue dispositions

Search delivery qualification. MEM-CLM-search-category qualifies the Search extension of RT-RTE-environment-rollout. Online skills-only selection reads task_specific_skills, whereas shipped Search categories use query_type_skills; offline distillation supports the latter key. Runtime leaves other environment semantics incompletely established and does not guarantee delivery of Search query-type skills. This is a limit on uniform hierarchical delivery, not a contradiction of a runtime guarantee or absence of all Search guidance. EPI-RTE-search-import and EPI-RTE-answer-check identify different informational and evaluator routes; neither supersedes the rollout nor repairs the category mismatch.

ID guarantee scope. MEM-OBJ-skill-metadata identifies the fixed initial existing-ID set used by direct addition. Runtime's qualified ID guarantee concerns serial updater-mediated reassignment and explicitly excludes arbitrary additions, concurrent writers and manual editing. Workflow's existing-ID rejection describes its updater-mediated path. These findings are compatible and establish neither semantic novelty nor a universal uniqueness invariant. MEM-RTE-bank-maintenance is a bypass-capable public mutation surface, not a duplicate of the two proposal admissions.

Experience save and selection. MEM-RTE-experience-save is an afforded heuristic file writer with no located shipped caller or same-instance index refresh. MEM-RTE-experience-selection is a wired manager-consumed selector for configured records. Its upstream context supply overlaps the rollout account without being identical to the complete rollout. These routes cannot be joined into automatic live episode learning without the missing caller and refresh evidence. Neither duplicates RT-RTE-offline-production, whose model produces a bank rather than heuristically saving an experience record. The missing connection limits a conclusion; no supplied report claims that the automatic loop operates.

Cache correspondence. MEM-OBJ-retrieval-index retains the model-name-only cache check and possible loss of vector/record correspondence after file replacement. Wired selection does not establish cache correctness. No peer report guarantees correspondence or claims that the mismatch actually occurred. This remains a local correctness limit, not a cross-report conflict.

SFT quality and consumer differences. Memory's annotation on RT-RTE-sft-production distinguishes WebShop's final-pass action-format removal from admissible-action mismatch counting. Runtime claims no universal semantic veto, and epistemic cleanup concerns a different stage. The reports therefore prevent treating a validation label, generated reasoning or repetition deletion as authentication of original trajectories. Cleaned traces, derived memories, bank instructions, reasoning-enriched conversations and final pairs retain different consumers and authorities. Uninspected parsing, repair and environment-specific generation semantics, as well as external fitting, remain local limits rather than conflicting findings.

Recovery and source gaps. Runtime, memory and workflow agree that separately saved banks and numerical checkpoints do not establish coupled recovery. The epistemic recovery ledger on RT-RTE-policy-update preserves that separation. Initial experience provenance, exact provider weights and behavior, other manager context handling, exercised compliance and causal benefit remain independently unestablished. Available-but-uninspected source is not an access gap and does not erase established implementation paths.

Independent convergence and remaining uncertainty

Runtime's admission findings, memory's title-only updater account, workflow's bounded proposal handoff and EPI-RTE-proposal-check converge on structural admission without a candidate-effectiveness test on the traced dynamic path. They connect failure excerpts to installed general instructions and later positive-history prompt delivery without treating those connections as successful fixes or causal support for benefit claims.

The epistemic account separately assesses formulation, consumption, content-directed criticism, iteration, learning, reflection, autonomy and improvement. Its wired later delivery does not establish exercised content dependence; its limited two-direction reflective pathway does not establish full theory criticism. Its improvement-directed self-change capacity leaves subsequent causal dependence and successful improvement unestablished. These distinctions do not disagree with runtime's refusal to infer exercised self-improvement from storage or updates alone, or with memory's limits on reliable lessons.

No cross-report disagreement remains in the pinned findings described above. Faithful uncertainty about unassessed alternatives, activation or benefit is not an unresolved conflict. No source reads were needed to connect these records; load-bearing passages remain in the cited members. This reconciliation changes no analyst finding and supplies neither a support judgment nor public synthesis.