OpenResearch reconciliation

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

Reconciliation

Shared boundary and evidence limits

The four reports concern the same enclosing runtime and frozen revision in boundary. Their implementation and doctrine/design findings use SRC-1 in the Source register. They establish neither a newly observed research workflow nor causal benefit. The workflow's representative local experiment is a source-supported scenario, not an execution trace. Missing traces, opaque native internals and unassessed included operations remain limits, not disagreements or system-wide absences. No source reads were needed for this reconciliation; the supplied records retain the supporting passages.

Duplicate supersessions

Amendment: EPI-OBJ-run-log is superseded by MEM-OBJ-run-log. Both identify the filesystem log selected by log_path(run_id) in SRC-1 src/commands/logs.rs, and retain the same path-construction passage. The epistemic focus on measurement-bearing portions does not identify a different stored object. Epistemic inventory and checking findings continue to concern that log; the memory record's broader lineage and consumer account does not grant its contents scientific warrant.

Amendment: EPI-OBJ-research-outputs is superseded by MEM-OBJ-project-artifacts. Both identify the mixed files under the project artifacts root in SRC-1 src/local/files.rs and agent-skills/orx-reports/SKILL.md. Both explicitly include prose, data and opaque or non-truth-apt files; the epistemic record is a container, not a narrower accepted-claims store. Its inventory, retention and chart findings retain their consumer-specific limits. Neither declaration establishes file-level provenance or scientific acceptance.

Amendment: EPI-RTE-log-preview is superseded by MEM-RTE-log-retrieval. Both describe the same resolved-run CLI operation in SRC-1 src/commands/logs.rs: local suffix read, at most 2048 bytes, a 500-character preview and full-file locator. The matching constant and warning passages establish identity, not merely similar purpose. The epistemic non-ampliative-reshaping assessment concerns this delivery operation; the memory record's later explicit-read affordance does not imply an automatic full-log read.

Amendment: WF-RTE-log-retrieval is superseded by MEM-RTE-log-retrieval. Both identify the same run-ID resolution, persisted-log access, compact stdout and subsequent targeted-read handoff. SRC-1 src/commands/logs.rs supplies matching preview and warning passages. The workflow also retains the requested continuation from agent-skills/orx-logs/SKILL.md; that instruction is not evidence that the foreground agent performed the read. The workflow record's containment within the broad experiment-run account remains compatible with selecting the memory declaration as the common referent. Both declarations distinguish wired delivery from unobserved consumption and interpretation.

These supersessions connect identities only. Both IDs remain declared in each case; no finding or status is replaced by a stronger one.

Containers, parts and distinct routes

MEM-RTE-turn-context is a declared preparatory part of RT-RTE-chat-turn, not a duplicate. The latter includes admission, dispatch, native execution and persisted return. The former isolates ordinary composition and its selectors, including the prepared_input bypass. Its supported shell read-back resolves the runtime's conditional statement: consecutive composer-shell exchanges on active ancestry enter ordinary prepared input as untrusted data. This positive refinement does not conflict with RT-RTE-user-shell, which does not claim that no later consumption exists. It does not extend to arbitrary old shell results or every entry point.

MEM-OBJ-chat-messages, MEM-OBJ-native-session and MEM-OBJ-bootstrap-context are distinct objects used by the broad turn route. The stored transcript is not the native payload; the bootstrap seed is not either object's complete serialization. MEM-RTE-chat-compaction is a maintenance route with native and fallback alternatives, while MEM-RTE-branch-context selects continuation ancestry. Neither is identical to the enclosing turn. Recovery and fallback's all-branch flattened snapshot versus side-chat active-path selection refine, rather than contradict, the runtime's lossy-recovery account. Native opacity does not erase the known fallback summary-to-later-turn chain.

MEM-OBJ-session-goal is distinct from RT-OBJ-playbook: persisted caller guidance enters ordinary turn composition, whereas the rendered playbook contains shipped instructions and changing project orientation. The memory annotation on the playbook identifies dynamic counts, not complete experiment evidence. MEM-BAP-bootstrap-injection and RT-BAP-playbook-injection therefore remain distinct authority paths: prepared prior-context text versus spawn-time appended system guidance. No supersession is warranted.

MEM-OBJ-experiment-tree is a valid container for experiment metadata and branches. EPI-OBJ-experiment-description isolates the free-form description field; it is not identical to the tree or RT-OBJ-source-snapshot. EPI-RTE-description-store retains caller text; EPI-RTE-tree-organize changes parent/archive metadata; RT-RTE-experiment-run selects and executes committed source. These are different operations. Current comparison parent need not remain historical code ancestry after a move. This qualification does not contradict the runtime's branch-tip snapshot binding at submission.

Amendment: EPI-RTE-artifact-retain is superseded by MEM-RTE-artifact-access. Both describe filesystem-presence admission, listing, contained access and deletion in SRC-1 src/local/files.rs and immediate visibility under agent-skills/orx-reports/SKILL.md. The matching immediate-artifact passage and common endpoints establish route identity. The memory account's linked-reader and withdrawal detail complements the epistemic retention assessment; neither route requires scientific acceptance or establishes automatic later-model injection. The epistemic retention ledger continues to assess this same mechanism.

EPI-RTE-chart-render is a distinct producer/checker, not the artifact-retention route. It gates this helper's new write on numerical/schema validation; arbitrary writers bypass it. An artifact container can validly coexist with this producer-specific record. The numerical and semantic portions retain different consumer rules without treating every artifact as checked.

Research interpretation and delivery

WF-RTE-result-handoff is a broader reporting handoff than EPI-RTE-results-check: it includes the requested evidence check, selected explanation, references, delivery and possible next action. The epistemic ledger separately distinguishes checking from conditional disposition. EPI-RTE-winner-promotion describes repair/refill/promote/stop policy beyond that reporting step. Shared results-skill passages establish overlap, not identity of these entire referents. The valid workflow container need not be superseded by its narrower analytical functions.

The reports independently converge on the separation of execution, scheduling, evidence and scientific interpretation. RT-RTE-experiment-run binds code and maintains process state; RT-RTE-run-wakeup schedules a later turn with status-only text; the reconciled log route supplies a locator and bounded preview; the results check requests evidence-consuming judgment. Wired launch, wake-up and log delivery do not conflict with doctrine-only semantic checking. These are separate functions, not competing statuses for the same guarantee. The chart helper's enforcing structural check likewise does not contradict advisory scientific guidance.

RT-CLM-research-assistance, EPI-CLM-results-evidence and EPI-CLM-experiment-loop are distinct claims: a broad research assignment, a reporting criterion and an iterative experiment policy. EPI-BAP-results-guidance concerns focused skill guidance, not the playbook injection mechanism itself. Instruction delivery is not evidence that checks, winner promotion or theory criticism occurred.

Remaining integration dispositions

The memory report's opaque-checkpoint, seed-lineage, native-discovered-skill and remote-log questions are retained local limits. MEM-CLM-independent-chat is a documented independent-chat contract, not a demonstrated native subagent frame or return route. The workflow's compute delegation is to a controller, not that independent chat or a reasoning subagent. No conflicting child-isolation findings require settlement.

RT-RTE-skill-upload and the slash/native consumers described by MEM-RTE-turn-context connect structural instruction admission to distinct later delivery channels. Native loading/compliance and initial discovered-skill provenance remain unresolved. RT-RTE-release-update concerns maintainer-supplied production machinery and channel policy, not the research winner-selection route. The reports do not infer learning, reflection, autonomous theory building or self-improvement from persistence, instruction distribution or external research-code optimization alone.

No incompatible cross-report findings were identified in the supplied members. The open questions above prevent stronger conclusions but do not supply a conflict to settle. No correction answers or declined objections were supplied. No additional record identity or required missing part was established by this comparison.