Ecdysis reconciliation

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

Reconciliation

Boundary and record identities

All four reports share the boundary: the shipped improvement plane and its inference interfaces, not the external production task runtime, editor or paper experiments. The Source register supplies implementation, doctrine/design and reported-operation evidence, but no observed run or causal experiment. The connections below use the pinned reports only; no additional source reads were needed. No correction answers or refusal were supplied.

Amendment: EPI-OBJ-review-transcript is superseded by MEM-OBJ-transcript. Both identify the same transcript list of role/content dictionaries in process memory and optional checkpoint JSON, including intermediate replies and normalized Moderator output. Both retain the same append passage from SRC-1, src/ecdysis/fdcr/review.py, and distinguish full retention from twelve-turn model delivery. The memory declaration also covers standalone saving. This is duplicate identity, not replacement of a finding. The epistemic inventory's possible diagnoses, objections and agreements refer to this same object; supersession establishes neither actual criticism nor acceptance.

Amendment: EPI-OBJ-round-record is superseded by MEM-OBJ-round-record. Both identify RoundRecord instances in TrainingResult.rounds, with scalar scores and reference-retention decisions in immutable dataclasses and an in-memory tuple (SRC-1, src/ecdysis/training.py). Their passages concern the construction and retention of those same decisions. Both distinguish this audit object from the harness and deny automatic durable reload or full evidence-to-patch lineage. The epistemic inventory's decision-evidence finding remains attached to this object; it is not another admission route or behavioral input.

Amendment: EPI-OBJ-evolved-skill is superseded by MEM-OBJ-evolved-skill. Both identify EvolvedSkill instruction fields and symbolic identifiers/provenance in dataclasses and artifact JSON (SRC-1, src/ecdysis/artifacts.py). Optional metadata in the epistemic declaration and the memory declaration's focus on instruction content do not identify separate instances. Both distinguish these units from optional raw skill dictionaries in the review specification, and neither establishes learned origin, efficacy or agent consumption. The epistemic inventory's possible truth-apt pattern content remains a finding about the same skill units, not a stronger provenance conclusion.

Other apparent overlaps do not establish duplicate identities:

The workflow member declares no new records. Its annotations concern the existing runtime referents, so no workflow supersession is required.

Convergence and integration dispositions

Review evidence and context. The runtime failure-review record, the memory annotations and the epistemic shaping record converge on different evidence reaching models and editor. Models receive metadata and a twelve-turn history tail; editor receives fuller compacted failures. The epistemic message-change countercase explains why changed editor evidence can leave displayed context and checkpoint hash unchanged. This supplements rather than contradicts the runtime's bounded equality guarantee (RT-RTE-failure-review, MEM-OBJ-trajectories, EPI-RTE-failure-shaping). The workflow's handoff account uses the same visibility distinction (workflow review handoffs).

Checkpoint freshness and rejected advice. The memory integration issue about narrower freshness agrees with the runtime hash/model account. Progress state controls skipped calls; completed state can deliver cached advice without new review. Neither report licenses full evidence identity, authentication or renewed endorsement (MEM-OBJ-checkpoint, RT-RTE-checkpoint-recovery). Memory additionally observes that trainer rejection does not delete the proposal checkpoint. This does not conflict with the trainer retaining the old harness reference: advisory retention and harness admission have different objects and consumers (MEM-OBJ-round-record, RT-RTE-candidate-admission). Workflow recovery is review-local, not a persisted harness-round restart (workflow verification and recovery).

Score admission and later authority. Runtime identifies strict measured-score reference replacement and its live-argument limitations. Memory traces the accepted reference into collection/execution; epistemic separates score production, disposition and retention from acceptance of an explanatory claim. These findings concern different functions of the same path, not incompatible statuses (RT-RTE-candidate-admission, EPI-RTE-score-check, RT-BAP-retained-harness). The workflow independently retains the rejection, no-failure and delivery branches. All retain external callback isolation and evaluator comparability as conditions, not package guarantees.

Unresolved harness behavior. Memory's unknown real harness form and lineage agree with RT-OBJ-frozen-harness. The quickstart editor ignores specification content in runtime, epistemic and workflow accounts. Thus wired delivery through RT-BAP-review-editor coexists with unobserved semantic dependence, improvement and theory-building behavior. No property is inferred from another or promoted to observed operation by this reconciliation.

Skills and outcome feedback. Memory's structural/provenance limits supplement the runtime utility route and epistemic learned-behavior comparison. Explicit loading and first-ID selection do not establish agent interpretation or trainer-score admission (RT-RTE-artifact-utilities, MEM-OBJ-skill-artifact, EPI-CLM-learned-behavior). The scoped MEM-ABS-skill-injection agrees with runtime's lack of automatic injection, without excluding external integrations. Likewise, outcome JSONL is not automatically converted to reward-bearing training input; exception-free reporting is not answer correctness (MEM-OBJ-inference-record, EPI-RTE-outcome-check, RT-RTE-harness-inference).

Supplementary helpers. Epistemic's exposed stability predicates supplement runtime's selected routes. Memory also reports that the trainer does not consult them. The signed-delta countercase therefore limits interpretation of a helper result, not the trainer's strict admission branch (EPI-RTE-stability-checks, memory coverage). The formatter's imperative text similarly supplies no independent agreement gate or shipped coding-agent execution (EPI-RTE-spec-render, memory specification annotation).

No disagreement between the pinned reports was identified. Memory's integration issues are compatible limits or distinct lifecycle functions, not conflicts. The three duplicate identities above require supersession; the remaining declared overlaps are parts, containers or different functions. No supplied correction has invalidated a dependent finding, and no declined cross-report objection was supplied. These dispositions connect the records without judging their support or settling the external evidence gaps.