OrangePro report verification

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

Verification

The current reports support a bounded source-grounded account. Both earlier blockers are resolved: credentials are no longer treated as a necessary condition for all drafting, and no-AI is no longer presented as a veto on explicit deterministic generation. I judged the current members afresh, read the correction answers and previous verification, and independently traced the consequential branches. The supplied structural check reports none; there is no structural failure to route. The checkout's HEAD equals the reviewed boundary and Git reported no working-tree changes. No target code, tests, examples, provider or service was executed.

Correction and control checks

For RT-RTE-start-run, I traced SRC-1 src/local/cli.ts start dispatch, src/local/localConfig.ts provider resolution, src/local/operations.ts generation-provider resolution and start controls, and src/local/autoProve.ts through both lanes. CLI accepts and forwards no-AI together with an explicit deterministic provider. The start resolver supplies that provider without credentials, and the existingOnly expression excludes that deterministic case. The automatic driver checks no-auto, static-only and attempt limits, tries existing tests first, then permits deterministic generation without a model configuration. The separate later gap-filling branch still requires AI enabled. These are reachable supported inputs, not merely an unchecked internal branch; generation remains conditional on eligible targets and budgets. Keyless configured Ollama is also supported. Runtime's revised forcing comparison, route fields and admission account, workflow's request/handoff account and annotation, and reconciliation's connected finding now preserve these distinctions. I dismiss both previous blockers as repaired, rather than carrying them as publication limits.

I also challenged static-only independently. RT-RTE-generation-compare correctly identifies a different path: CLI default comparison, MCP compare and background job compare reach opCompare, which omits static_only from both generation calls. SRC-1 src/local/generate/generator.ts passes only the explicit option into goCompile; that helper does not read the process flag. The raw branch can reach its temporary-file write and go test -run ^$ . after an executable response, suitable Go target/module and installed tool are available. Environment inheritance does not repair the missing argument. The reports correctly state conditional source reachability, not an observed bypass, and withhold a global static-only guarantee.

Records, support and evidence strength

I assessed every declaration and its cited source anchors, the annotations and dependent prose, status choices, operative identity, containment and claimed authority. The following checks use SRC-1 implementation at the registered source boundary. Record quotations supply the load-bearing passages; source comments and prompts remain doctrine rather than observed behavior.

  • Providers and prompt authority: RT-CMP-openai, RT-CMP-anthropic, RT-CMP-ollama and RT-BAP-generation-prompt. Reading src/local/generate/providers.ts supports configured model names, externally held unpinned parameters, text requests and adapter-specific truncation handling. No inspected request updates parameters. The supplied prompt passages request output constraints and weak-context caution; neither their delivery nor parser acceptance demonstrates semantic compliance. Opaque service state is not classified as absent learning.
  • Graph, retained tests and context: RT-OBJ-graph, RT-OBJ-generated-tests, RT-RTE-test-generation and MEM-RTE-context-selection. I checked src/local/workspace.ts, generation/preservation/update branches in src/local/operations.ts, and context assembly, flow selection, pinning and raw Go generation in src/local/generate/generator.ts. Provider resolution precedes retained-draft serving; fresh-only append avoids duplicating pinned tests. Full-analysis preservation and incremental stale marking differ. Weak candidate reasons enter prompt context despite exclusion of candidate flows from deterministic flow-chain selection. Source excerpts and selected graph strings are not host transcript memory. Fingerprint eligibility and reference resolution establish neither dependency-wide currency nor assertion truth. The reports retain those qualifications.
  • Caches and saved-result consumption: MEM-OBJ-parse-cache, MEM-OBJ-resolver-cache, MEM-OBJ-run-key and MEM-RTE-start-reuse. src/local/analyze/parseCache.ts, src/local/resolve/resolverCache.ts, src/local/runReuse.ts and the shortcut/save code in src/local/operations.ts support the stated symbolic substrates, version/shape gates, key selection and later consumption. Ordinary workspace outputs are excluded from the code digest, with ledger and proof-manifest input exceptions; large files use size/mtime. Reuse returns saved findings after rerendering, not fresh proof. Malformed saved data can fall through to full work. No exhaustive invalidation or authentication guarantee is emitted.
  • Candidate admission: RT-RTE-candidate-admission. Direct apply reads in src/local/aiGraph/links.ts and src/local/flows/llmFlowDiscovery.ts support current-graph validation, deduplication, workspace hash and shown/current anchor intersection. opStart invokes generation then apply without per-proposal human approval. Candidate-lane storage and weak-context delivery do not grant public proof credit or semantic warrant.
  • Proof, ledger and later authority: RT-OBJ-dynamic-proof, MEM-OBJ-ledger, RT-RTE-targeted-proof and RT-BAP-proof-credit. I traced both ledger writers in src/local/operations.ts, the certificate shape in src/local/ledger.ts and selection in src/local/rtm.ts. Static record processing resets credit to unproven even when a scoped confirmer found new hard edges. Dynamic credit uses the common baseline/mutant gate. RTM separately requires closed certificate predicates and a current target fingerprint, choosing the newest matching success rather than the latest attempt. This supports retained decision authority, not tamper-proof execution history or full runner soundness.
  • Oracle, fault evidence and classifier: EPI-OBJ-oracle, EPI-OBJ-fault-evidence and EPI-RTE-fault-probe. src/local/contracts/oracle.ts accepts proposed or approved metadata and does not authenticate its author. src/local/faultProbe.ts compares repeated witness outputs and selected test identities, with additional conditions for full-suite scope. It does not compare output truth with expected-value prose. The epistemic account preserves acquisition, checking and disposition as separate functions and does not infer semantic acceptance from shape validation.
  • Draft feedback and automatic recovery: RT-RTE-draft-verification and WF-RTE-proof-recovery. src/local/draftTest.ts initializes model feedback locally while loading durable attempt counts, then feeds candidate/verdict information to bounded repair. Success ends repair; the first candidate submitted through verification supplies the repair baseline, not a prior successful verified-gap requirement. I read runtime's “actually verified candidate” procedurally in that context and dismiss it as a contrary mechanism claim. src/local/candidateTest.ts checks repeated originals/faults, assertion position, text controls, restoration and suite conditions while leaving assertion-to-oracle semantics to review. src/local/autoProve.ts relocates generated imports and rechecks the destination before writing/proving; a non-runnable flag alone is not an absolute veto. Its caught proof exception leaves the written file. Workflow preserves this transformation and partial-result recovery without claiming successful repair or rollback.
  • Comparison evaluation: EPI-OBJ-comparison-scores, EPI-RTE-comparison-scoring and EPI-CLM-judge-accuracy. src/local/generate/compareScore.ts, src/local/generate/compareJudge.ts and opCompare support the lexical witness, score formulas, coercion and fallback. One assertion with the stated empty metadata yields the reported numerical values by deduction, not execution. Metadata-conditioned model opinion and exact syntactic counts have different warrants. Neither establishes semantic correctness or causal graph benefit, and the result does not install a new prompt/provider policy.
  • Export, machinery and delivery: EPI-OBJ-verification-receipt, EPI-RTE-export-replay and RT-RTE-actions-delivery. src/local/exportTests.ts checks patch-byte equivalence and recomputes candidate classification before fresh replay and A–H receipt disposition. src/local/actionsRun.ts replays a configured candidate or returns needs-test; it does not silently invoke drafting. src/local/actionsInstall.ts gates managed writes/conflicts, and src/local/publishTests.ts separately requires confirmation, receipt/base/branch checks and non-force draft-PR delivery. No merge follows. src/local/boundary.ts requests the documented container restrictions; the install branch has a distinct network/writable-mount contract. Source wiring is not evidence of live isolation or adoption.
  • Claims and external adapters: RT-CLM-local-generation, EPI-CLM-proof-scope and RT-RTE-platform-call. The retained doctrine passages are compared with narrower executable mechanisms rather than endorsed as global guarantees. Package entry points, src/orangeproTools.ts, src/apiClient.ts and packages/mcp-remote/src/index.ts support separate local and backend-adapter responsibilities. Header-selected endpoint/key/tenant values and REST requests do not establish incoming authentication, backend evaluator quality, local graph synchronization or tenant isolation. Hosted state remains uninspected rather than absent.

All six explicit containment relations are supported: generated tests belong to the graph; context selection to generation; start reuse and proof recovery to start; scoring to comparison; export replay to the delivery chain. A directly callable export stage can still be that chain's part. Ledger history, certificate, oracle premises, fault observations, scores and receipt are distinct referents. There are no declared evidenced absences or supersessions to approve. No missing operative split invalidates these container identities.

I checked reconciliation against all declarations and memory integration issues. Its dispositions preserve the freshness differences, weak-edge versus candidate-flow distinction, historical-success selection and separate delivery/adoption gates. Runtime's intended approved-brief role is not a claim that approval is authenticated; epistemic's more precise shape-check finding therefore does not require an unresolved conflict. Different export inspection depths likewise do not contradict one another. No cross-report disagreement requiring an anchored conflict remains.

Omission questions and stopping decision

The frozen top-level tree, CLI/MCP/job dispatch and producer/consumer dependencies guided the omission checks, not only the listed anchors. Could another shipped caller defeat a control exclusion? Explicit deterministic start and foreground/detached comparison were traced above, including supported-input reachability and execution prerequisites. The repaired account now accommodates both. Could another writer grant proof from caller-reported success or could a later failure revoke historical success? Searching certificate and ledger writers and tracing both record_run and RTM selection supports the stated protocol and consumer-relative retraction policy. Neither conclusion is generalized to hostile local ledger edits.

Could feedback state conceal a learning route omitted from the memory inventory? Additional SRC-1 implementation inspection of src/local/feedback.ts found invitation preferences and cooldown read-back, plus hosted-form URL construction. Its local consumer changes report invitation presentation, not test-generation policy. The narrow state-to-consumer witness is:

if (prefs.invitations === "off") return false; --- src/local/feedback.ts

I dismiss omission of that presentation state as immaterial to the non-exhaustive memory mechanism account; it does not overturn any emitted learning or epistemic finding. This is implementation evidence, not observed form use or backend behavior.

Further optional search stopped because the selected consequential questions have supported dispositions and no identified material challenge remains unresolved. Complete language-mutator, coverage-parser, recipe and renderer audits could add route detail, but the reports already withhold full mutation correctness, universal security/freshness guarantees and system-complete negatives about knowledge production or improvement. Host/provider/backend internals and unsupplied runs prevent operation and benefit conclusions; those uncertainties are already expressed beside the affected findings. They are not additional defects or evidence of absent behavior. The whole shipped-artifact boundary remains viable without attributing external host responsibilities to OrangePro.

Blockers

none

Limits

none