# Kingy PDF workflow and LM Studio native document comparison

This plan precedes any access to the new v1.4 document/question/reference contents by the root or repair agent. The corpus author worked separately from candidate code, outputs, failures, tuning and competitor settings.

## Conditions

Use the same recorded M4 Pro (14 CPU / 20 GPU cores), 24 GiB unified memory, macOS 26.3 build 25D125, AC power and unchanged power mode. Recheck this inventory and record memory pressure/swap/thermal state for every session. Run one system at a time. Do not interrupt unrelated jobs or pretend this shared Mac is an isolated workstation.

Both systems use the same verified Qwen3-4B Q4_K_M artifact, 2,497,280,256 bytes, SHA-256 7485fe6f11af29433bc51cab58009521f205840f5b4ae3a32fa7f92e8534fdf5. Target a single 8,192-token context, temperature zero, reasoning off, ten CPU threads, full Metal offload, flash attention and F16 KV where each runtime exposes those controls. Record effective settings and any mismatch. Kingy uses llama.cpp b11429; LM Studio uses its own shipped engine. Runtime differences belong to the complete workflow comparison, not a claim about the underlying model alone.

The new independently authored pack has ten English text-bearing PDFs, 30 pages and 40 held-out questions: 20 direct, 10 cross-document and 10 unsupported. Five unsupported questions describe explicit omissions and five facts are genuinely absent without omission cues. Twenty development items are supplied for future reproduction but are not used to tune this already-calibrated comparison. Both candidates may calibrate only on consumed v1.3 and earlier cases before final freeze. No reference values or support spans enter model inputs.

## Native LM Studio baseline

Use the signed official LM Studio app and its stock document attachment/preprocessing workflow. Record app/build, inference-engine version and binary hashes, official rag-v1 revision and actual plugin configuration. The documented defaults use three retrieved chunks and affinity strictly greater than 0.5, with the native Nomic embedding model. Record its artifact/hash and parser/chunker settings. Let the stock workflow choose full-context injection or retrieval.

Every question begins a new chat with all ten identical PDFs attached. Verify attachment count and filenames before submission. The question must be the first user message; no dummy setup prompt with attached documents. A later turn follows different stock preprocessing rules. Capture native processed model prompts, outputs, citations/source mappings and timing through documented logs and actual GUI/chat export. Do not substitute Kingy BM25 retrieval or its sufficiency checker for LM Studio's document workflow.

The common user request names the question and required fields, asks for explicit facts from the supplied documents, asks for supporting source citations/exact quotes, and asks for a clean decline when the documents do not supply the requested information. A machine-readable format may help export but syntax alone cannot decide answer correctness. Permit documented native output/citation conventions through a format-only adapter. Extra model-side judgments are not added to the LM Studio baseline after seeing outcomes.

## Repetitions and timing

Predeclared session order: A1, B1, B2, A2, A3, B3, where A is Kingy and B is LM Studio. Each system runs three fresh model sessions, each retaining all 40 planned attempts. Within each paired repetition use the same deterministic question permutation with seed 271828 plus the repetition number. Save this schedule after the final freeze. Stop only task-owned inference processes between sessions; do not terminate unrelated applications or jobs.

Use the same neutral non-bank warmup policy for each model. No corpus-dependent warmup. Record model cold start and import/parse/indexing separately. If native first-query indexing happens after submission, it remains in that query's end-to-end latency and is reported explicitly. Record cache reuse and do not claim OS file caches were flushed.

Complete-answer latency covers retrieval, all model phases, adapter processing and usable citation availability. Kingy's selection and sufficiency calls both count. Retain native first-token, output-generation and completion timestamps separately where exposed. Do not equate GUI automation overhead with inference latency or silently remove retrieval/indexing from one system. If comparable boundaries cannot be measured, publish quality results and withhold the unsupported speed comparison.

## Scoring and claims

Freeze the common format adapters, canonical aliases, source-grounding rules, page-resolution rules, abstention rules, thresholds, paired aggregation and all source/settings/artifact hashes before either candidate sees the held-out bank. Correctness requires every essential requested fact and qualifier. Citation support requires that the answer's actual claimed citation substantiates the fact; automatically displayed retrieval cards are not automatically credited as emitted answer citations.

A uniquely located exact source quote may supply a reproducible page mapping when native citations omit a page number. Report that mapped-page capability separately from an explicit model/app page citation. Preserve ambiguities and unresolved semantics rather than inventing facts or source locations. No LLM judge alone certifies quality. Any conclusion depending on unresolved judgments stays unproven.

Each complete run must independently meet conservative lower bounds of 90% complete supported-answer success, 95% emitted-citation support precision and 90% correct abstention to qualify a configuration. Complete supported-answer success requires every requested fact/qualifier plus its linked source support. Retain errors, timeouts, duplicates, missing planned work, uncited correct facts and every failure. Publish canonical factual correctness separately so citation tooling does not masquerade as underlying fact accuracy. Unresolved semantics remain in denominators with lower/upper bounds; a comparative claim cannot depend on unreviewed ambiguity.

Forty unique paired questions are the statistical unit; three repetitions measure repeatability, not 120 independent questions or users. Report per-run denominators and paired question-level differences, with uncertainty and small-sample limits. Restrict any advantage to these configured complete document workflows on this Mac and this synthetic corpus. No claim of general model superiority, market-wide cheapest setup, novice usability, arbitrary PDF reliability or independent hardware reproduction.

## Current activation prerequisite

LM Studio's welcome screen explicitly says continuing accepts its Terms of Use. Curtis explicitly approved this action through the pending approval question; Get Started was clicked and the app reached its normal chat interface. The final comparison freeze still awaits consumed-data native capture/settings preflight. No fresh-bank access or execution has occurred.

## Native isolation preflight adjustment before evaluation

Launching the official desktop executable under an outer macOS Seatbelt deny-egress profile failed before inference because Chromium's GPU/network subprocesses could not initialize their own sandbox. The invalid first profile and corrected-profile crash logs are retained; no new-bank question was executed. Do not disable Chromium sandboxing or interrupt unrelated jobs by changing the whole machine's network. The native baseline will retain its normal vendor sandbox and perform local Qwen/Nomic document inference on the public synthetic corpus. Native egress enforcement is unverified. Kingy's runner/server retain their tested deny-egress profiles. This study can compare factual answers and source support; it cannot establish comparative privacy, or claim that LM Studio's offline mode fails.

Native calibration uses the installed unmodified rag-v1 revision8 with retrievalLimit8 and default affinity threshold0.5, selected on consumed data only. The original three-chunk setting missed the second source in the consumed cross-document question; eight chunks returned both correct answers. The final baseline therefore uses the calibrated setting, rather than attributing that known default limitation to the configured comparator.

Native first messages come from duplicates of a visibly empty template with ten identical verified PDF attachments and no previous messages. Every created record is checked for original versions, one first user message, one answer, frozen full request, attachment hashes, native settings, processed prompt and citation mappings. Named empty templates avoid automatic title inference; any additional requests are still retained when observed.