ReframeWeb report verification

Type: agentic-system-analyses/types/agentic-system-verification.md

Verification

The reports support publication of a bounded implementation account. No material contradiction or unsupported central classification was established. One local omission limits comparison of workspace input-selection routes, as recorded below. This judgment does not establish successful execution, reliable model compliance or improved capacity.

Evidence and record dispositions

I reviewed every supplied declaration, its evidence status and source anchors, the annotations, every Part of: relation, and the reconciliation's identity and conflict dispositions. I independently traced consequential behavior in SRC-1 rather than adopting the analysts' reported checks. The checkout HEAD matched the frozen revision and its working tree was clean. The boundary includes the shipped application and native services; the repository tree agrees with its top-level coverage. The supplied structural report says none; all four correction-answer files are empty. There are no structural failures or correction arguments requiring disposition.

Model and effect boundaries. RT-CMP-model-transport, RT-CMP-codex-model, and RT-CMP-whisper-transcriber distinguish model identifiers from frozen weights and leave opaque parameter behavior unresolved. Their adapter/configuration evidence does not support stronger fixity or training conclusions. RT-OBJ-tool-attempt correctly locates assignment checks and retry uncertainty in the Host attempt, not in a universal sandbox. Inspection of ToolAttempt.call and _execute confirms call-ID reuse, argument fingerprints, receipt recording and the bounded uncertainty ledger. RT-BAP-store-callback and RT-BAP-panel-import have different consumers and enforcing checks: activation/revision and operation-specific grants versus invocation capabilities and native request/result bounds. These support the reports' rejection of a universal task-tool monitor. Evidence: SRC-1, agent-host/src/reframe_agent_host/task_tools/attempt.py, agent-host/src/reframe_agent_host/store_host_boundary.py, semantic-store/crates/store-runtime/src/panel/host.rs.

Invocation, execution and delivery. RT-RTE-request-cycle, RT-RTE-model-task, RT-RTE-codex-task, and RT-CLM-agent-output survive the forcing cases examined. The orchestrator filters context before composition, dispatches effects/output recording before completion, requires PASS together with null failure fields, and returns a bounded result without guaranteeing request satisfaction. Direct execution appends linked instructions. Codex execution uses a fresh ephemeral process while retaining the prepared workspace on retry. The executable summary path extracts one executor feedback item; the routing comment describing another conversation is not implemented by that path. The workflow account and annotations correctly describe these same routes, including the unused failed_execution argument, accumulated summary/refusal retry context, no-action threshold and absence of automatic rollback. Evidence: SRC-1, agent-host/baml_src/ns_voice_turn/task_flow.baml, agent-host/baml_src/ns_voice_turn/agent_routing.baml, agent-host/baml_src/ns_task/refusal/flow.baml, agent-host/baml_src/ns_task/ns_internal/failure_resolution.baml, agent-host/src/reframe_agent_host/agent_execution/tasks.py, agent-host/src/reframe_agent_host/agent_execution/summary.py, agent-host/src/reframe_agent_host/task_tools/outputs.py.

Retained material and actual consumers. MEM-OBJ-context-memory, MEM-OBJ-session-memory, MEM-OBJ-instruction-memory, MEM-OBJ-filesystem-memory, MEM-OBJ-workspace-store, and MEM-OBJ-task-history keep text, locators, backing bytes and receipts distinct. Preference initialization is not inferred from the candidate writer. MEM-RTE-context-reload loads collection text without a numerical cap in those loaders; MEM-RTE-graph-selection instead uses hints/depths, model-kept IDs and subsequent transformations. Read marking occurs before relevance selection. Store-project search ignores global tags and has its own 32-result cap. MEM-BAP-linked-instructions establishes a separate prose-delivery path, not enforced model obedience. MEM-RTE-history-summary has a concrete later completion consumer without establishing a durable summary index. These findings are supported, subject to the additional workspace-input route below. Evidence: SRC-1, agent-host/src/reframe_agent_host/agent_flow/memory_contexts.py, agent-host/src/reframe_agent_host/agent_flow/session_context.py, agent-host/src/reframe_memory/graph_retrieval.py, agent-host/src/reframe_memory/store_projects.py, agent-host/src/reframe_agent_host/workspace/task_preparation.py, agent-host/src/reframe_agent_host/task_tools/history.py, agent-host/src/reframe_agent_host/agent_flow/action_history.py.

Admission and maintenance. RT-RTE-guidance-review stages nonempty proposals, reviews nonempty groups, intersects kept IDs and writes only selected title/description content. Its memory annotation correctly distinguishes a trace-fed guidance architecture from proven learning or complete provenance. MEM-RTE-project-locator distinguishes attempt-local verification receipts from the registration/path gate on saving. MEM-RTE-node-curation and MEM-RTE-memory-migration supply genuine alternative changes, with guarded instruction replacement, incident-edge deletion, snapshot/conflict checks and locator-visibility migration. They do not imply universal human approval or semantic validation. RT-RTE-saved-store separates checkpoint publication, metadata CAS, membership and live activation/recovery. RT-RTE-application-publish correctly restricts its matching-receipt gate to initial application publication and retains partial multi-write failure. Evidence: SRC-1, agent-host/baml_src/ns_memory/candidate_tools.baml, agent-host/baml_src/ns_voice_turn/memory_review_flow.baml, agent-host/src/reframe_agent_host/agent_flow/candidate_memory_writeback.py, agent-host/src/reframe_agent_host/task_tools/project_memories.py, agent-host/src/reframe_agent_host/task_tools/save_artifacts.py, agent-host/src/reframe_agent_host/task_tools/store_instructions.py, agent-host/src/reframe_agent_host/memory_browser/records.py, agent-host/src/reframe_memory/graph_merge.py, agent-host/src/reframe_memory/locator_migration.py, agent-host/src/reframe_agent_host/workspace/publication_service.py, agent-host/src/reframe_memory/store_memories.py, agent-host/src/reframe_agent_host/startup_activation.py, applications/session-management/publish.py.

Epistemic parts and licenses. EPI-OBJ-task-input, EPI-OBJ-output-summary, and EPI-OBJ-candidate-entry do not assert observed semantic preservation or conjecture. EPI-RTE-prompt-assembly confines its positive preservation finding to deterministic packaging. EPI-RTE-completion-check, EPI-RTE-completion-gate, and EPI-CLM-completion-evidence correctly distinguish an asymmetric model criterion from the enforcing success conjunction. EPI-RTE-guidance-check, EPI-RTE-guidance-disposition, and EPI-CLM-supported-guidance preserve the important input gap: review requests interaction-supported lessons but receives purpose, candidates and stored text, not a separate interaction trace. Implementation is supported; actual claim-level acceptance remains unestablished. Evidence: SRC-1, agent-host/baml_src/ns_task/prompting/flow.baml, agent-host/baml_src/ns_task/ns_internal/prompt_text.baml, agent-host/baml_src/ns_task/completion/flow.baml, agent-host/baml_src/ns_memory/candidate_review_prompt.baml, agent-host/baml_src/ns_memory/ns_internal/candidate_review_flow.baml.

EPI-OBJ-test-result and EPI-RTE-candidate-startup accurately bound their predicate to private startup/selection and cleanup. Generic publication does not require consuming that result. EPI-OBJ-rendered-evidence and EPI-RTE-panel-evidence distinguish preserved PNG bytes, lossy preview and newest-eight image attachment from factual or interaction acceptance. EPI-RTE-benchmark-checks and MEM-CLM-selection-evaluation correctly distinguish exact reference/shape metrics from production admission and causal benefit. Sorting preserves multiplicity; prompt-shape success does not inspect factual meaning. Evidence: SRC-1, agent-host/src/reframe_agent_host/candidate_testing.py, agent-host/src/reframe_agent_host/task_tools/store_publication.py, agent-host/src/reframe_agent_host/task_tools/screenshot_artifact.py, agent-host/baml_src/ns_task/execution/tool_media.baml, agent-host/baml_src/ns_benchmarks/evaluation.baml, agent-host/src/reframe_agent_host/benchmarks/task_prompt_execution.py.

Relations and reconciliation. All declared containment relations preserve valid enclosing routes: context reload/selection, normal-history summarization, request prompt/summary/completion parts, guidance content/check/disposition/criterion, and isolated candidate startup. Cross-kind containment names carried values or criteria rather than competing route identities. The reconciliation appropriately makes no supersession: overlapping checkpoint production, benchmark checking and summary production do not make their different objects, claims and consumers identical. Its no-conflict disposition is supported. The implementation/comment discrepancy is not an unresolved disagreement between reports. No missing conflict citation or concealed contradictory finding was established.

Omission questions and stopping decision

The frozen tree, host registration and producer/consumer dependencies guided three consequential challenges beyond the reports' listed passages. First, could workspace inputs bypass task links and relevance selection? Yes: live saved-panel projects have an automatically wired projection, retained below as an inventory limit. Second, could other shipped publishers invalidate the admission account? Inspection of SRC-1, applications/session-management/publish_update.py and semantic-store/authoring/publication.py found additional explicit maintenance writers. The update publisher has its own receipt predicate and preserves existing task/instruction state; the authoring publisher writes directory and instruction memories. These fit the reports' already plural, nonexclusive admission account. They do not invalidate the initial-publication gate or establish per-content human control. Their presence does not justify an exhaustive publisher comparison.

Third, could a separate evaluator force generic production publication, authenticate model summaries or establish measured learning? The traced candidate-test, publication, completion and benchmark consumers do not provide those stronger links. The reports already withhold them rather than emit system-wide absences. Uninspected acceptance suites and native backend/storage leaves therefore remain possible additional checks, not evidence against the supported local wiring findings. Further omission expansion stopped because the selected challenges have dispositions and no identified material contradiction remains. A full native security, durability, evaluator or publisher audit is not necessary for the bounded conclusions these reports actually retain.

Blockers

none

Limits

• MEM-RTE-graph-selection, MEM-BAP-linked-instructions, and RT-RTE-codex-task do not inventory the separate automatic live-panel input projection. SRC-1, agent-host/src/reframe_agent_host/host_services.py wires live_panel_projects; agent-host/src/reframe_agent_host/workspace/panel_manifest.py selects nonclosed saved panels with resolvable project entries; agent-host/src/reframe_agent_host/workspace/task_preparation.py adds their filesystem memories and writes a project manifest plus generated instructions. This is a local omission, not a false graph-selection finding: explicit task links and relevance-selected Store inputs still work as reported, and neither report claims an exhaustive workspace-input inventory. Withhold comparisons that treat those two selectors as the only sources of prepared files or generated instruction context. The additional source establishes wiring, not exercised projection or model compliance.

self.agent_tasks.panel_projects = lambda: live_panel_projects(self) --- agent-host/src/reframe_agent_host/host_services.py

file_ids.extend(filesystem_memory_ids(panel_projects)) --- agent-host/src/reframe_agent_host/workspace/task_preparation.py