Company Brain report verification

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

Verification

The current reports support a bounded application-level account. Both previous blockers are repaired; no new material defect was found. This judgment follows independent source inspection, not the analysts' validation results or the previous verification's conclusions.

All supplied inputs and contracts were read completely, including the bounded ranges and final tails. The supplied structural report-check says none; there is no structural failure to route. Runtime and epistemic each supply one correction answer; memory and workflow answers are empty. The previous verification is an argument to reassess, not evidence of source behavior. No target code, test, provider or service was executed.

The frozen checkout HEAD equals boundary's revision, and its tracked top-level tree agrees with the inventory. SRC-1 src/worker.ts, src/routes/index.ts, Slack dispatch and src/brain/turn/loop.ts establish the shipped application and its model/tool loop. Scheduled, administrative and maintenance callers remain inside it. The enclosing-runtime classification and whole-system functional boundary remain supported. Neither that boundary nor these reports claim exhaustive inspection, remote-service implementation coverage or observed operation.

Correction dispositions

Entity freshness: corrected. I read SRC-1 src/brain/memory/entities.ts completely and traced the persistence flag in src/brain/turn/tools.ts. RT-RTE-entity-web now distinguishes the generation-filtered hosted lookup, post-lookup guard, pre-write guard and stale late-write cleanup from the unfenced web-result return. Its progression, Immediate return, Invalidation or expiry and Evidence limits all preserve the reset-during-inference countercase. The Runtime account describes cache/write guards without reinstating a return-freshness guarantee. Memory's entity annotation concerns cache lookup and persistence suppression; epistemic's entity ledger assesses identification rather than guaranteeing freshness. Reconciliation retains the same distinction. The former runtime blocker is discharged without weakening the supported cache controls.

Reception consumer: corrected. I read SRC-1 src/brain/auto-research/outcome.ts and src/brain/admin/chat.ts completely, searched outcome symbols across src/ and web/, and checked the agent method/export passages in src/brain/turn/agent.ts and src/brain/turn/agent.impl.ts. Non-refinement administrative requests call recentDraftOutcomes, format the results and pass them into a fresh computeTurn. EPI-OBJ-draft-outcome, EPI-RTE-outcome-acquisition, EPI-RTE-outcome-frame and EPI-RTE-admin-consumption retain that implemented chain. The coverage table, ledger, draft-send read-back and bounded conclusion no longer call it consumerless.

The new scope distinctions also hold. Outcome selection uses sent references and selected channel/DM surfaces; framing omits reply-user and draft IDs. Operator-held history supplies at most 20 turns. Refinement skips reception and pending blocks and proposes wording for separate application. The administrative function does not persist an accepted criticism or automatically revise the research plan. Direct inspection of SRC-1 src/brain/auto-research/plan.ts confirms prior bodies and operator steering as planner inputs, not acquired outcomes. This supports reception interpretation capacity without asserting its exercise, criticism-driven iteration or improvement.

The partial-failure countercase is real: a rejected thread read becomes null, but successful reactions plus thread?.ok !== false can set engagementKnown true with empty replies. The formatter can then display no replies without known reply evidence. The corrected record and conclusion expose this defect rather than treating the flag as reliable absence evidence. The former epistemic blocker is discharged; this source-system defect does not invalidate the accurately qualified architectural finding.

Record and evidence judgments

I checked all declarations and annotations, their source anchors, evidence layers, status claims and consequential consumers. The groups below record their dispositions. Supporting passages remain in the cited records; source reading was used to assess their meaning rather than repeat deterministic quotation or link checks.

  • Models, live state and standing text. SRC-1 src/brain/turn/brain-model.ts, src/brain/turn/model-profile.ts and src/config/index.ts support the distinct roles and provider/fallback qualifications in RT-CMP-brain-main, RT-CMP-triage-model and RT-CMP-research-model. Symbolic names are not frozen weights; inference does not establish training. Approval SQL and turn-control state support RT-OBJ-turn-state. Skill storage/load/approval, workspace HTTP gates, configuration tools and prompt assembly support RT-OBJ-brain-skill, RT-OBJ-workspace-prompt, RT-RTE-instruction-admission, MEM-RTE-config-edit and RT-BAP-standing-instructions. Requester-approved skill creation, human HTTP edits and model-mediated admin edits have different admission mechanisms. Prompt hierarchy is authored guidance, not demonstrated model compliance or executable access policy.
  • Invocation, coordination and external effects. SRC-1 src/routes/slack/index.ts, src/brain/slack/turn.ts, src/brain/slack/turn-control.ts, src/brain/turn/approval.ts, src/brain/turn/compute.ts and src/brain/turn/loop.ts support RT-RTE-slack-turn, RT-RTE-approval-resume and RT-CLM-turn-fencing. HTTP acknowledgment differs from delivery; KV dedup differs from an atomic exactly-once protocol. Requester/team/status/expiry/current-turn/workspace checks admit continuation, not answer truth. Source effect policy and direct execution in src/brain/tools/mcp/policy.ts, src/brain/tools/mcp/tool-policy.ts and src/brain/tools/mcp/runtime-tools.ts support RT-RTE-app-code, RT-RTE-direct-mcp, RT-CLM-write-approval and RT-BAP-tool-policy. Unknown personal tool names can evade the separate direct approval predicate; read-only/shared filtering is different. Sandbox assembly, regex guards and both adapter call sites support RT-RTE-sandbox-work, not universal external-effect mediation. These local guarantees retain their owners, enforcement points, alternate paths and external contracts.
  • Knowledge input, derivation and admission. SRC-1 src/brain/memory/writeback.ts, src/brain/memory/index.ts, src/brain/turn/capture-tools.ts, src/compat/routes/memories/handler-effect.ts and src/memory/memories.ts support RT-RTE-knowledge-handoff, MEM-OBJ-brain-document, MEM-OBJ-derived-memory and EPI-OBJ-memory-doc. Captured proposals, submitted dated documents and returned derived rows differ. The adapter ignores instant/replacement controls and flattens array metadata; queued/done licenses submission success, not completed extraction or factual acceptance. EPI-RTE-memory-admission and EPI-RTE-document-retention preserve that distinction. Channel spool/batch processing in src/brain/slack/channel-observe.ts supports MEM-OBJ-channel-batch, EPI-OBJ-channel-clusters, MEM-RTE-channel-curation and EPI-RTE-channel-distillation: persistence precedes submission, partial submission retries before cursor cleanup, and HUMAN-only provenance is requested rather than independently checked. Profile provisioning in src/brain/memory/profile-sync.ts and extraction doctrine in src/brain/memory/profile-config.ts support MEM-OBJ-profile-config, MEM-RTE-profile-provision and MEM-CLM-service-curation. Ignored visibility does not enforce remote privacy, and unknown initial settings do not erase known subsequent configuration.
  • Recall, navigation, withdrawal and human readers. SRC-1 src/brain/memory/read-scope.ts, src/brain/memory/profile-recall.ts, src/brain/memory/search-brain.ts, src/brain/memory/tags.ts, src/brain/memory/tree.ts and src/brain/turn/context-tools.ts support MEM-OBJ-memory-index, MEM-RTE-ambient-recall, MEM-RTE-requested-recall, EPI-OBJ-brain-results and EPI-RTE-brain-selection. Candidate caps precede rendering; buckets can exclude otherwise searchable facts; tag selection is lexical; timestamp-only paging cannot guarantee enumeration. Relevance and optional provenance do not establish truth. src/brain/memory/split.ts and tree repointing support MEM-RTE-node-splitting, EPI-OBJ-topic-partition and EPI-RTE-topic-partition: original documents feed organization proposals, remote metadata precedes local moves, and array-only tag updates prevent unconditional consistency. Forget apply, cleanup and reset in src/brain/tools/forget-memories.ts, src/compat/routes/v4/memories/forget-matching.ts and src/brain/turn/agent.impl.ts support MEM-RTE-memory-withdrawal and EPI-RTE-reset-withdrawal, not exact-preview provenance or complete erasure. The debug caller supports MEM-RTE-debug-capture, not authenticated Slack admission. src/routes/memories.ts and src/routes/graph.ts support MEM-RTE-human-readers, including different graph liveness handling and no inferred edit authority. Client/service scoping leaves deployment tenancy unresolved, not evidenced leakage.
  • Retained continuity and style authority. SRC-1 src/brain/turn/context.ts, compute call sites and src/brain/turn/thread-investigation.ts support RT-RTE-context-boundary, MEM-OBJ-thread-investigation and MEM-RTE-investigation-recall. Compaction retains bounded values and reload stubs, not recoverable originals behind digests. Principal/thread/TTL checkpoint recall differs from approval conversation restoration. Any successful native call can select answer text as verifiedEvidence; no claim-level verification follows. src/brain/turn/interaction-observe.ts, src/brain/memory/interaction-style.ts and prompt assembly support MEM-OBJ-agent-style, EPI-OBJ-style-note, MEM-RTE-style-recall, EPI-RTE-style-generation and EPI-RTE-style-admission. Carried observer context and later bucketed facets have different consumers. Human-evidence instructions do not prove compliance. MEM-BAP-retained-context correctly distinguishes behavior-shaping supply from exercised dependence or benefit.
  • Research, scheduled work and reply warrant. SRC-1 research plan/run/draft/store/send, agent enqueue and Slack posting support RT-RTE-research-draft, EPI-OBJ-research-draft, EPI-RTE-draft-check, EPI-RTE-draft-send and WF-RTE-draft-delivery. Preloaded capped history, concurrent independent turns, advisory evidence/length flags, lease/reset storage admission and separate publication choice are supported. The sender rejects sending before the helper's stale-claim recovery; failed-post release does not guarantee crash recovery or exactly-once remote publication. Scheduler admission/callbacks in src/brain/tools/scheduler.ts and src/brain/tools/scheduling.ts support RT-RTE-scheduled-task, including payload-dependent read-only posture and no delivery of noncompleted output. src/brain/observability/reply-quality.ts and src/brain/turn/finalization/output.ts support EPI-OBJ-reply, EPI-RTE-reply-check and EPI-RTE-reply-selection: the gate judges selected shapes, not claim evidence, and recovery chooses the longest acceptable transcript text. README/system-prompt passages support the attributed claims EPI-CLM-team-answers, EPI-CLM-current-memory and EPI-CLM-evidence-check. Native web lookup, opaque extraction and nonsemantic reply gating justify withholding absolute source restriction, guaranteed currentness and factual self-check compliance.

All emitted route wiring statuses concern implemented local calls, controls and consumers. Operation remains uninspected; no record promotes authored tests, prompts, service status or delivery permission into observed accuracy or causal benefit. No evidenced-absence declaration was submitted. Conditional proposal/decision/veto, admission/recovery and oracle distinctions are supported at their stated scopes. Neither retaining experience nor receiving reception evidence alone establishes learning benefit, theory construction or exercised self-improvement.

Relations and reconciliation

The explicit Part of: relations have supported containment: checkpoint within turn state; admission, retention and style functions within knowledge handoff; reply check/selection within Slack turn; draft checks and delivery within the research route. Child fields do not inherit every parent guarantee. Memory's annotations and workflow's research annotation add lens-specific lineage, authority or delegation without redefining supplied identities.

The three duplicate supersessions identify the same style material, topic split and explicit sender segment. They preserve both analysts' findings and do not promote evidence strength. The input/submitted-document, cluster/spool, selected-result/store, reset/forget and checkpoint/compaction distinctions are justified. The newly declared reception object and three stages have distinct functions and no duplicate sender or automatic-planner identity.

All eight memory integration issues have supported dispositions in reconciliation. Additional model-mediated/debug admissions extend the bounded runtime account; ignored adapter controls, tenancy, withdrawal, metadata and trust-stage qualifications do not contradict stronger guarantees asserted elsewhere. Workflow adds history/recovery detail; epistemic adds operative shape checks and administrative feedback consumption. An analyst's uninspected leaf is not a conflicting absence claim. No unresolved conflict, invalid supersession or required missing part was identified. Reconciliation neither reinstates the previous defects nor selects a stronger status to settle uncertainty.

Omission questions and stopping decision

The frozen tree, alternate shipped callers and producer/consumer dependencies selected three consequential questions:

  1. Do freshness controls cover the returned result as well as cache state? Complete entity tracing and tool-call persistence selection confirm the repaired return countercase. The current records no longer imply end-to-end stale-read prevention.
  2. Does reception close into a consumer, and which consumer? Cross-tree outcome-symbol searches, complete acquisition/admin reads, agent exposures and planner input inspection establish operator-facing interpretation, not automatic engagement-driven planning. Refinement/history and partial-failure handling are retained at their actual scope. The corrected records support the comparison without implying exercised criticism or improvement.
  3. Can alternate admissions or delivery paths inherit the main loop's guarantees? Tool assembly, direct MCP, sandbox adapters, debug capture, approval writeback, schedule callback, hosted adapter/reset and research sender/helper checks preserve distinct grants, authorship, publication and withdrawal mechanisms. Complete erasure, universal effect mediation, exact preview provenance, exhaustive retrieval and exactly-once publication remain withheld in the submitted records, while positive wiring remains supported.

Further optional omission expansion stopped because these selected questions have source-supported dispositions and no identified material challenge remains unresolved. Available uninspected permission/UI/evaluator, membership, checkpoint-rendering and service leaves prevent exhaustive equivalence or system-complete negative comparisons; the reports already express those uncertainties without emitting such comparisons. Their presence alone does not require another inventory. The existing account remains faithful without reopening the boundary or requesting corrections solely to enlarge coverage.

These are SRC-1 implementation reads at the frozen revision; embedded instructions and README claims were assessed as doctrine/design. No new source or operation evidence was introduced. The remaining uncertainties are already faithfully scoped in the records, not additional verifier findings requiring a publication limit.

Blockers

none

Limits

none