Fractal reconciliation

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

Reconciliation

All four reports use the frozen revision and runtime-client boundary established by boundary. Fractal owns the shipped drivers, configuration and session lifecycle; PredictRLM owns the inner computation. The Source register supplies implementation and doctrine/design, not observed runs or experiments. The reports consistently retain that separation. This reconciliation uses the supplied members only; no source reread was needed.

Record identity and containment

Amendment: EPI-OBJ-session-summary is superseded by MEM-OBJ-session-summary. Both declare the same source-native SessionSummary.turns, containing user text and AgentTurn fields, persisted in session JSON and rendered into signature instructions. Identity evidence is SRC-1, src/fractal/session.py and src/fractal/agent/signature.py, including the SessionSummary declaration in the epistemic record and the field assignments/rendering passages in the memory record. Different lens labels do not establish different referents. This supersession connects the epistemic inventory, retention and rendering ledger entries to the memory account of the same object; it does not replace either analyst's findings or strengthen their statuses. Both IDs remain declared.

RT-OBJ-session-state remains a valid checkpoint container. MEM-OBJ-session-summary and MEM-OBJ-session-history declare distinct contained parts. MEM-OBJ-turn-usage is a numerical part of the summary. EPI-OBJ-agent-turn isolates the answer/outcome fields within that summary and the containing checkpoint. These are containment relations, not replacement identities: the container remains valid and no split supersession is warranted. Their supporting SRC-1 passages in src/fractal/session.py distinguish full checkpoint, accumulating projection, recent detailed tail, and usage/answer fields.

RT-OBJ-edit-workspace is the complete dynamic signature, not another identity for the summary. MEM-BAP-summary-context isolates the retained-summary consumption channel; RT-BAP-workspace-guidance imports root AGENTS.md. Both reach signature instructions but carry different source content and horizons. Object/channel overlap does not warrant supersession.

Connected routes, not duplicate invocations

MEM-RTE-checkpoint-update isolates automatic projection and persistence across RT-RTE-headless-turn, RT-RTE-budget-exhaustion, and RT-RTE-interactive-interrupt. It is not identical to any complete invocation or forcing branch. MEM-RTE-session-reload isolates cross-process lookup, validation and replacement. MEM-RTE-session-context isolates subsequent selection and delivery. No supersession is warranted among those routes.

EPI-RTE-summary-render covers the summary-fields-to-text-to-signature transformation within the broader memory context-delivery route. The latter also selects and delivers detailed history and loads root instructions. Their common rendering step is not whole-route identity. SRC-1, src/fractal/session.py, src/fractal/runtime.py, and src/fractal/agent/signature.py, as retained in these records, support this distinction.

EPI-OBJ-runtime-event is transient reducer input; EPI-RTE-event-projection is its reduction and downstream accounting route. RT-CLM-hook-accounting states the best-effort accounting claim, and MEM-RTE-checkpoint-update consumes the reduced observations alongside result/exception data. These are object, route, claim and wider persistence mechanism, not duplicate referents. MEM-OBJ-turn-usage and EPI-CLM-trustworthy-usage concern trace-derived numeric usage rather than hook-derived file/command observations; they likewise warrant no supersession.

EPI-RTE-provider-check supplements RT-RTE-config-change with a connectivity diagnostic and its command-specific consequences. A diagnostic is not the adapter/configuration mutation route. EPI-OBJ-version-cache and EPI-RTE-version-notice retain and consume a website assertion for advice, not package publication. EPI-RTE-contract-tests checks development contracts, while RT-RTE-package-release publishes a tagged artifact without a test-job dependency. None is a duplicate of another or a task-answer oracle.

WF-RTE-audit-repair is the caller-coordinated two-process procedure, not another declaration of headless submission or session reload. It composes two invocations and a continuation handle. Its dependency on saved context connects to the memory reload/delivery routes without asserting an additional autonomous coordinator.

Independent convergence and integration dispositions

The runtime RT-CLM-result-contract, memory MEM-RTE-checkpoint-update, epistemic authority-route ledger, and workflow WF-RTE-audit-repair agree that output type checks and outer status classification do not establish semantic task acceptance. Direct workspace effects precede those checks; failure, interruption and budget fallback do not roll back edits. Delivery, persistence and later reliance remain distinct from verified repair or accepted truth-apt content. The records retain implementation wiring and uninspected operation separately; the epistemic ledger's implemented architectural status is not an upgrade to observed operation.

The memory integration issues require no correction of the supplied runtime account. Their dispositions are:

  • Summary “compression” means deterministic field selection and activity counts, not semantic shortening of response prose. MEM-OBJ-session-summary, MEM-OBJ-session-history, and EPI-RTE-summary-render converge on this mechanism. The detailed tail is capped; summary rendering is not. This qualifies EPI-CLM-context-rot, not a cross-report contradiction.
  • Dependency trace completeness, recursive context isolation, sandbox retention and model-selected recall remain local unknowns. MEM-RTE-session-context, RT-RTE-direct-agent, and the epistemic bounded conclusion do not infer absent dependency behavior from missing evidence. Caller-supplied history is not universally capped or persisted by the adapter.
  • Persisted modification lists can omit already-applied failure/interruption effects, and before-phase commands can be attempts. MEM-OBJ-session-history and EPI-RTE-event-projection add specificity compatible with the runtime's best-effort accounting claim. Neither supplies a complete audit.
  • Listing has weaker filename/embedded-ID consistency checks than loading. MEM-RTE-session-reload retains that navigation limit. The workflow's resume procedure does not depend on listing; its account separately preserves the possibility of an empty replacement under a continuing ID. Existence or ID capture is not proof of preserved audit context.

The stale workspace-local session path in bundled calling doctrine is identified in boundary and the workflow account. Runtime and memory use workspace-keyed global persistence. This is a source documentation/implementation discrepancy, not disagreement between reports. The audit recipe's missing first-turn exit gate is also compatible with automatic retention of unsuccessful turns, not a claim that such continuation succeeded.

EPI-CLM-readable-trace remains a broader source claim than inspected best-effort hook accounting and bounded retained detail. EPI-CLM-trustworthy-usage remains bounded by supplied trace inputs; deriving numbers outside response prose does not establish complete audit or answer warrant. No report provides the missing observed-run or causal evidence for those benefits.

No unresolved cross-report disagreement or required undeclared part was identified in the supplied members. No correction answers or declined objections were supplied to this attempt. The supersession above resolves the duplicate summary identity only; it settles no disputed support judgment. Later delivery of retained context and editable root guidance remains compatible with independently unestablished learning, theory-building and exercised self-improvement conclusions.