This independent teaching example is developed for Yeabin Moon's public learning collection. The prompt, checks and worked review were prepared with AI assistance. The example is not endorsed by a model provider or certification program.
Filing evidence
source.json retains P01 and P02 from the reviewed peer packet: the complete selected intellectual-property paragraphs in Apple's original FY2024 and FY2025 annual 10-Ks. SOURCES.md presents the same text with filing links, dates and section anchors. Source HTML hashes and text hashes are retained in JSON. Original-filing source files were previously inspected locally; internal hashes do not establish remote SEC byte identity. Each local source HTML was one byte shorter than its SEC index size. No new full-filing processing was performed for this lesson.
The paragraphs are whitespace-normalized source selections. They do not cover all AI risks, prove infringement or establish realized financial losses. Fiscal period, filing date and displayed acceptance timestamp are distinct fields; no trading-availability rule is implemented. Issuer passages remain third-party material with original-source attribution, separate from any future license for the accompanying code.
Separate native-assistant run
A fresh native assistant subagent received prompt.txt and source.json, with instructions not to browse or inspect other project files. The requested model selection was gpt-6-astra at high reasoning effort. The first generated JSON artifact is saved unchanged as response.json. This is an observed assistant-produced artifact, not a raw API response envelope or direct SDK call. Native system/tool instructions still applied. The backend snapshot, token usage, API sampling parameters and cost were not exposed and are not inferred.
A separate AI source reviewer developed criteria before reading the output, then identified a claim/type granularity issue in finding 1. A single follow-up using correction-prompt.txt produced corrected-response.json. The original response remains intact. Hashes and known run metadata appear in run.json.
The worked support judgments are AI-assisted editorial judgments, not independently observed human adjudication. Python checks structure, quotation membership and coverage only; it does not classify truth, validate economic materiality or infer causal relationships. A small transparent example is not a model benchmark, and repeated output with the same prompt is not guaranteed.
Controlled examples and future comparisons
The three files under controls/ are intentionally constructed changes to the first response. They illustrate a missing field, a fabricated quote and an unsupported claim with a genuine quote. They must not be counted as observed failures of the selected model. controls.json states their expected mechanical outcomes and support judgments.
Local Qwen run
The qwen/ files retain one actual Qwen3.8 27B MLX 4-bit run through LM Studio 0.4.26+4, executed locally by Yeabin Moon on October 15, 2026. The final message is preserved byte-for-byte in qwen/response.json; its prompt, source, schema and audit inputs match this package. Requested Think off, loaded context 8192 and an output cap of 4096 produced complete JSON with five exact quotations. qwen/run.json records requested settings, input/output hashes, server statistics and client elapsed time. Exact weights revision and backend application of every setting are not independently established.
This was a selected diagnostic after earlier incomplete outputs and a model-load resource refusal. It does not measure first-attempt reliability or compare models fairly. Its 45.59-second client elapsed time is distinct from the server's 8.811-second time to first token and 22.098 output tokens/second; 812 output tokens and zero reasoning tokens were reported.
An assistant source review identified strengthened modality and an ambiguous AI/licensing connection despite a mechanical pass. qwen/editorial-response.json is a separately labeled assistant-proposed revision; qwen/editorial-changes.json identifies each changed field. The original remains untouched. The revision has not been independently human-approved, and passing the checker does not supply that approval. Raw runtime envelopes and reasoning traces are excluded from this public package.
A future controlled comparison would need previously unseen passages, comparable protocols and separate support judgments. Tool selection, retrieval, routing and stateful recovery remain later extensions. No NVIDIA certification, Qwen recognition, measured learner benefit or class duration is established by this example.
Open-weight models and deployment choices
Documentation checked October 15, 2026. These general capabilities motivate the lesson; the saved case does not measure all of their benefits.
- Control and versioning: model files can be downloaded at a specified repository revision, enabling an experiment to retain a particular version. Hugging Face download documentation.
- Customization: parameter-efficient fine-tuning methods can adapt pretrained models to downstream tasks. Applicable licenses, compute and task data still matter. No fine-tuning was performed in this example. Hugging Face PEFT documentation.
- Local data handling: LM Studio documents local inference and document processing that can operate offline after model download. This describes a local deployment; using cloud models or connected tools can change where data goes. LM Studio offline operation.
- Possible cost savings: specialized models and suitable workloads can reduce inference costs; a published sentiment-classification case illustrates that possibility. Its historical prices and performance are not estimates for our Qwen setup. Hardware, utilization, energy and maintenance must be considered in a new comparison. Hugging Face application case.
- Reproducibility: preserving weights, code and settings supports repeatable experiments, but deterministic results are not guaranteed across software releases and hardware. The PyTorch guide illustrates this general issue; our saved Qwen run used MLX. PyTorch reproducibility guide.
Model capabilities and this experiment
Qwen's official documentation describes agent-related capabilities including planning, handling environment feedback, tool-calling integration and adjustable reasoning. These are developer-documented features, not findings from our experiment or proof of the quantized model's agent reliability in LM Studio. Qwen3.8-27B model card and Qwen repository.
Our request supplied no model tool integrations. Python followed a fixed request/save/check sequence, and a separate source review assessed meaning. We have not benchmarked autonomous planning or tool use. Future agent examples would need their own permissions, stopping conditions and evaluations.
Model size and runtime memory
The saved LM Studio pre-request metadata reported size_bytes = 16081678492 for qwen/qwen3.8-27b, with MLX format, 4-bit quantization and loaded context 8192. That is 16.08 decimal GB of reported model size, not measured peak memory. qwen/run.json retains this selected size metadata alongside its other recorded settings. Hardware (Mac Studio M4 Max, 36 GB) was reported by the operator.
For 27 billion parameters stored at exactly four bits each, the simplified storage calculation is 27e9 × 4 ÷ 8 = 13.5e9 bytes, or 13.5 GB (about 12.57 GiB). Real model packages can also contain quantization metadata and components at other precisions. Context storage and runtime buffers require additional memory. No peak-memory measurement or universal minimum-hardware requirement is established here.
The model card's 27B count describes the language model; the simplified estimate is not a byte-level accounting of every component in a multimodal package. The exact installed weights revision remains unknown. Keeping the dated setup and unchanged result preserves the experiment's history when models or software change.