OKF Harness report verification

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

Verification

The five submitted members support a bounded, code-grounded account. I read all supplied reports, their types, reconciliation, the four empty correction-answer files and the structural check. The structural check reports none; there is no structural failure to route. No prior verdict or correction argument was supplied. I independently confirmed the registered checkout's HEAD and clean status against the boundary, then traced source behavior without executing the target or its tests.

Record support and evidence strength

Runtime and acquisition. I checked both component records, both runtime objects, all seven runtime routes, both runtime claims and both behavioral-authority paths. The CLI registrations and command handlers lead to the core operations described, not a target-owned model scheduler. SRC-1: packages/cli/src/index.ts, packages/cli/src/define-command.ts, packages/cli/src/commands/init.ts, packages/cli/src/commands/source.ts, packages/cli/src/commands/workspace.ts, packages/core/src/index.ts. In RT-RTE-source-ingest, file hashing, same-hash reuse, exclusive creation, manifest append and cleanup on append failure are supported by packages/core/src/source/index.ts. URL acquisition is correctly restricted to generated pointer metadata. The metadata-only candidate ranking and five-item cap are supported by packages/core/src/source/ingest.ts. The synthesis and file-count obligations are instructions, not executable semantic gates. The separate implementation, agent-protocol and operation statuses preserve that distinction.

Retained parts and context. I checked all eleven memory object declarations, both memory routes and all three memory claims, together with their annotations. The source/manifest, wiki prose/parsed fields, ledger IDs/notes, checkpoint trees/judgments and graph artifacts have the different consumers described. SRC-1: packages/core/src/source/index.ts, packages/core/src/okf/document.ts, packages/core/src/okf/concepts.ts, packages/core/src/lineage/index.ts, packages/core/src/source/reconciliation.ts, packages/core/src/recovery/index.ts, packages/core/src/graph/index.ts. The scaffold and initialization paths support local Markdown/JSONL state rather than an inferred vector store. SRC-1: packages/core/src/workspace/index.ts. MEM-OBJ-reconciliation-notes correctly distinguishes retained text from an established semantic read-back consumer; history does expose checkpoint judgments. Neither persistence path establishes trace-fed policy improvement.

I traced evidence selection through snapshot/check, risk filtering, graph seals, search, section/range selection and provenance construction, and compared standalone search/read. SRC-1: packages/core/src/evidence/index.ts, packages/core/src/search/index.ts, packages/core/src/read/index.ts. RT-RTE-evidence-answer and its memory annotation correctly retain the three-item selection, capped candidate cards, excerpt-character budget and policy-only continuation limit. Full document citation metadata can accompany a selected excerpt; it is not claim-by-claim warrant. RT-RTE-direct-retrieval is a genuine counterpath to planner-local seals. MEM-CLM-log-exclusion correctly qualifies that route: legacy logs are excluded, while index navigation remains readable. No claimed universal access gate or host-window bound survives in these reports.

Maintenance and epistemic force. I checked the five epistemic objects, three routes, provenance claim and every authority-ledger row against their cited mechanisms. SRC-1: packages/core/src/check/index.ts, packages/core/src/lint/index.ts, packages/core/src/source/reconciliation.ts, packages/core/src/graph/index.ts, packages/cli/src/doctor/index.ts, packages/cli/src/commands/doctor.ts. Rule-bound diagnostic and connectivity deductions are appropriately restricted to their predicate domains. They do not license source truth, semantic acceptance or causal benefit. The doctor finding is limited to orchestration and reported probe outcomes, not an audit of every probe.

The reconcile function admits an exact suspected pair and nonempty note without inspecting wiki edits. Pending revision warnings do not enter the evidence seal mechanism. The maintenance-loop test asserts unchanged evidence eligibility before and after acknowledgment, but supplies no observed test result. This supports MEM-CLM-currency-evidence and the workflow delivery account. SRC-1: packages/cli/test/maintenance-loop.test.ts. I also traced the full reconciliation instructions and workspace stop contract: source reading, semantic revision, check, repair ownership and edge-specific acknowledgment are connected steps, not unrelated capabilities joined into a fictional workflow. SRC-1: packages/agent-pack/templates/okf-harness/references/reconcile.md, packages/agent-pack/templates/okf-harness/SKILL.md.

The answer instructions expressly separate permission to retain from certification of correctness. Empty provenance supports a disclosure convention but cannot distinguish a manual unsourced page from a conversation write-back. MEM-CLM-provenance-signal and EPI-CLM-provenance-origin preserve both the requested policy and its evidential countercase. SRC-1: packages/agent-pack/templates/okf-harness/references/answer.md. Doctrine-only semantic rows do not deny the implemented delivery of instructions. Unobserved model behavior is not converted into evidenced absence; no absence records are declared.

Identity, reconciliation and omission checks

I checked the reconciliation against every declaration and integration issue. Its two supersessions identify the same check and graph operations across specialists, retaining complementary evidence and assessments rather than replacing values. The mixed epistemic source, page and ledger containers already distinguish the parts that memory records separately; no split supersession is needed. Other object/route pairs are products or inputs versus operations, not duplicate identities or operative subroutes requiring a Part of: relation. No Part of: fields are asserted. The standalone check spans several consumers rather than exactly one supplied route. The provenance policy claim and the memory qualification of its inference are distinct claims. All six memory integration issues have dispositions. I found no conflicting findings requiring an unresolved-conflict paragraph or correction.

I selected omission questions from the frozen tree, public exports, CLI registrations and write/subprocess call sites: could an alternate shipped caller bypass withholding, change instructions or configuration, or supply semantic acceptance omitted from the main loop? The direct retrieval counterpaths are already recorded. Recovery stages broad workspace state and restores a selected tree without semantic evaluation, as RT-RTE-workspace-recovery states. Launcher dispatch and pin adoption preserve existing pins and depend on external package resolution. SRC-1: packages/setup/src/index.ts, packages/core/src/config/index.ts. Managed installation and OpenCode synchronization change available instructions under ownership rules, not experience evaluation. Native host text lacks the detailed workspace payload, so instruction parity and actual loading remain unestablished. SRC-1: packages/agent-pack/src/index.ts, packages/native-integration/src/opencode.ts, packages/native-integration/skills/okf-harness/SKILL.md. The refresh helper emits fresh-session advice, not an implemented host-context eviction operation. SRC-1: packages/cli/src/refresh.ts.

One additional shipped withdrawal path is locally omitted from the guidance inventory: bootstrap uninstall dispatches to ownership-checked removal of host entrypoint files. SRC-1: packages/cli/src/commands/bootstrap.ts, packages/agent-pack/src/index.ts.

await rm(absolutePath, { force: true }); result.removedFiles.push(absolutePath); --- packages/agent-pack/src/index.ts

This does not invalidate the reported provisioning mechanism or establish an omitted semantic evaluator. I retain the inventory limit below. Further omission expansion stopped because the selected consequential challenges have supported dispositions, the reports do not emit complete security/integration/evaluator coverage, and no identified material contradiction remains. Individual doctor probes, distribution parity, adversarial filesystem behavior and actual host/model execution remain outside the conclusions established here; their availability alone does not justify an exhaustive audit.

Blockers

none

Limits

  • The guidance-change inventory in RT-RTE-guidance-install, also used by the memory boundary inventory, does not separately characterize caller-requested bootstrap uninstall. SRC-1: packages/cli/src/commands/bootstrap.ts, packages/agent-pack/src/index.ts. This shipped route removes host entrypoint files after ownership/status checks and supports dry-run. The supported installation/synchronization and wiki-memory findings remain valid; withhold any inference that the listed routes exhaust instruction withdrawal or cross-host recovery behavior. This is a partial inventory, not an unsupported absence or complete-coverage value.