{ "schemaVersion": "kingy-document-control-provenance/1.0.0", "instrumentation": { "amended": true, "phase": "B1-after-20-before-21", "beforeFirstAffectedInference": true, "retainedNativeOriginalsBefore": 20, "nativeB1ExtraColdStartAfterOriginal": 35, "nativeB1ColdLoadCount": 3, "nativeB1NeutralWarmupCount": 3, "nativeB1ExtraColdLoadsAfterOriginal35": 2, "nativeB1ExtraNeutralWarmupsAfterOriginal35": 2, "nativeInterruptedOriginals": [ { "session": "B3", "scheduledPosition": 23, "questionId": "V19-Q09", "status": "incomplete-unverified", "actualModelGeneration": "unverified", "questionRetried": false, "nativeRecordSha256": "d98481153cec34f759ee58a1dc1a429f8238d08857a0c2811ae13e2ce90b6a6c" } ], "candidateCorpusPromptsScorerUnchanged": true, "freshReferencesOrQualityGradesReviewed": false, "description": "Native LM Studio used the separately reviewed direct System Events template-menu control. After two blank-copy persistence delays, an independently reviewed wrapper was added before B1 original21. It waits up to ten seconds for a single acknowledged Duplicate action to persist a blank ten-attachment record, then returns to unchanged frozen checks. It adds no UI retry, model request, question submission or change to Send/answer timing. After original35, two unscheduled preparation generations occurred outside every measured answer window. The first followed unexpected short input; the user later reported concurrent typing, but exact event attribution was not established. The second contained file parts without text during menu preparation; its cause remains unresolved. Both affected native chats and their original bytes were preserved privately. Coordinate-based measured-trial template-menu preparation was retired before original36. An independently reviewed PID-scoped AXPress control verifies the exact owned row, outer AXGroup popup trigger, enabled popup state and labelled inner More options child, then presses the outer trigger once. It requires a focused More menu, and the binding rejects any second menu, including unnamed native overlays. An initial inner-button target mismatch and stale dual-menu guard defect were corrected before original36, with prior sources and failed probes preserved. The installed vendor UI source was inspected read-only; vendor code was not changed. The blank ten-attachment template is checked before and after preparation. The neutral native preflight created zero-file unscored copies once each and proved canonical reselection and repeat menu access without inference. Foreground indicators and Cancel acknowledgments were not treated as proof of selection or dismissal. A fresh owned blank chat was renamed and five PDFs were sequentially attached with exact PDF/configuration checks. The sixth sequential upload was visibly refused by the installed default five-file add handler. No policy, limit or vendor code was changed, and no over-limit upload was forced. Independent review allowed conditional reuse of a pre-existing, never-measured blank already containing all ten frozen PDFs, through ordinary native identity, name-only and duplicate controls. The native arm therefore depends on an existing ten-file conversation state; this does not establish that a fresh sequential ten-file setup works or that the UI has a universal five-file limit. Native Reveal in Finder proved the exact unused ten-PDF record before it was renamed. The five-file preparation was preserved under a quarantine name, and only the verified ten-file blank name changed to the canonical template name. One ordinary native Duplicate then produced a blank ten-file copy with exact frozen hashes and ten visible previews; the required screenshot API passed. The final model selection and eleven configuration keys are bound by the later AX freeze and verification receipts. Later draft-only characters were preserved privately; the user reported concurrent Mac use, but exact attribution remains unresolved. After the user confirmed an idle Mac, a clean unloaded GUI restart restored fresh observation. New inference was held while a separate finite local-model job occupied the shared host. The study model was kept unloaded during UI recovery. B1 used three recorded model loads and three neutral warmups in total, including two extra cold loads and two neutral READY warmups after original35. Each warmup was one model request without documents or a bank question. All receipts are retained. During attachment-only preparation, another job loaded a separate model and selected runtime2.60 while frozen2.56 remained installed. All34 frozen app, model, RAG and runtime artifacts were reverified unchanged. That observation does not identify who changed the selection or prove a cause for UI interruptions. A mistaken sheet-count diagnostic was corrected against an explicit native sheet lookup. Normal window and sidebar layout changes restored full-width controls; a fresh native menu-item point click opened the actual picker. These preparation actions did not change candidate inference, grading or retained original attempts. Extra conditioning and preparation pauses can affect cache state. Native runs include preparation pauses; B1 also includes extra cold loads and is not an uninterrupted-session timing result. Four unrelated input/output log events are omitted from the public derivative with exact hashes disclosed; all other event bytes and original private logs are preserved. The public engine-process proof omits a local credential-bearing command and binds the unchanged private original by hash. Independent export checks verify the derivative and raw-receipt exclusion; the final completed public artifacts still require the actual privacy scan. Before B3 original1, the native control connection closed before a response. Reconciliation proved zero retained or pending questions, an unconsumed menu token and only the two neutral warmup log events. A documented control reset recovered the actual singleton menu; the old token and unchanged log prefix were preserved. No model reload, extra warmup, control-source edit or question replay was needed. Before B3 original2, the attachment guard rejected one newly duplicated blank: it had ten entries but only nine distinct backing-file identifiers and hashes. The canonical source still had all ten correct frozen PDFs. The rejected blank was preserved unchanged and its native identity verified. A fresh ordinary copy passed ten distinct frozen PDF hashes before any input; original2 was then submitted once. The same guard also rejected a second blank before original22, before prompt, pending record or Send. Its separate interruption and recovery receipts preserve that preparation event. These storage-identifier observations do not establish which filenames the UI or model displayed, their cause, or general product reliability. B3 scheduled original23 (V19-Q09) was submitted before an unobserved gap spanning a host restart. The saved record has no documented final answer, final-generation block or matching final model-log event. Old owned processes were unavailable; a new main app instance existed and remained uncontrolled during reconciliation. The exact native record, pending clock and46-event log prefix were preserved. An initial capture-helper preflight required an explicit null timestamp while the original JSON omitted that field; it stopped before a marker, journal write or coordinator call. A separately reviewed correction accepts only absence or explicit null and rejects every present completion value. The approved capture-only helper then invoked the unchanged frozen coordinator once, using a truthful observation scoped to the unavailable old owned instance and forbidding all UI/model actions. It retained original23 as incomplete without inventing UI completion, answer timing or an answer, and cleared the pending slot. No question was resubmitted. The missing final evidence does not prove that no model request occurred, a wrong model answer, or product unreliability. A fresh neutral ten-PDF duplicate, screenshot-API, canonical reselection and single focused menu preflight then passed on the restarted main instance. A separate reviewer bound the exact frozen2.56 engine, all eleven configuration keys, effective SDK settings, unchanged ten PDF hashes,48 log lines and original46-event prefix before allowing original24 onward. One additional cold model load and one neutral warmup followed the host restart. Their actual continuation and final restoration are disclosed in the linked final load/warmup inventory. No comparative advantage may be claimed from this interrupted row. After that preflight, a focus-only helper observed native focus, but CUA omitted its focused-element metadata. That pilot did not pass and was stopped; unavailable metadata does not prove native focus was absent. A distinct transport amendment moved the current native focus check to the same bounded call as exactly one Duplicate AXPress. Its source, failed neutral attempts and actual verification are preserved separately. This changes control transport only; frozen candidate, scorer, corpus, later prompt and Send behavior remain unchanged. Kingy A2 initially stopped before any server load, warmup or question submission because a model link pointed to a disconnected external volume. The zero-attempt start was preserved. The link was repaired to the existing local native model after exact frozen-byte verification and a separate metadata-only review. Candidate, corpus, scorer and model bytes did not change, and no measured question was replayed. All planned original submissions are retained. Preparation interruptions and unscheduled activity are not product reliability evidence.", "evidenceIds": [ "focus-pilot-review", "focus-pilot-discrepancy", "duplicate-boundary-freeze", "duplicate-boundary-review", "duplicate-neutral-proof", "duplicate-neutral-review", "resumed-original-settings", "transport-freeze", "transport-review", "readiness-addon-freeze", "readiness-addon-review", "template-recovery-review", "replacement-template-review", "replacement-template-verification", "ax-menu-freeze", "ax-menu-review", "ax-menu-preflight", "final-native-preflight", "final-native-template", "final-native-runtime", "final-native-runtime-selection", "native-load-warmup-final", "b3-control-interruption", "b3-control-recovery", "b3-attachment-guard", "b3-attachment-recovery", "b3-attachment-guard-22", "b3-attachment-recovery-22", "host-gap-observation", "host-gap-boot", "host-gap-processes", "host-gap-artifacts", "host-gap-recovery-review", "host-gap-capture-first-review", "host-gap-capture-preflight", "host-gap-capture-review", "host-gap-capture-execution", "host-gap-capture-verification", "host-gap-neutral-preflight", "host-gap-neutral-review", "outer-trigger-review", "outer-trigger-proof", "native-target-diagnosis", "native-attachment-diagnosis", "native-five-file-refusal", "native-file-limit-source", "existing-native-state-review", "final-native-identity", "final-native-rename", "final-native-duplicate", "native-layout-reconciliation", "native-runtime-observation", "frozen-native-artifacts-reverified", "model-log-redaction", "concurrent-use-clarification", "second-concurrent-use-clarification", "native-host-preflight", "native-process-credential-review", "private-screenshot-export-review", "local-model-start-recovery", "local-model-recovery-source", "local-model-recovery-review", "local-model-failed-start", "measurement-integrity" ] }, "all240OriginalsRetained": true, "noIndividualQuestionRetries": true, "allMeasurementsStoppedBeforeRootReferenceAccess": true, "controlErrorsAreProductResults": false }