1. Research question and model identity

Can a communication device propose useful replies locally while keeping the user’s words under the user’s control? The system must be usable without a cloud connection, but locality alone does not prove response quality or safety.

TinkyMind is the language model. TinkySpeak is the conversation runtime. The model’s proposed choices are not delivered speech; selection and delivery are separate protocol steps. Neither component should be renamed Lois, Peel, or FailFirst.

2. Local inference and bounded requests

1. Partner input
2. Local reply proposals
3. User selects or composes
4. Host confirms delivery

The inspected adapter identifies the checkpoint as child-v6cube-aug, requires a manifest with engine cube-onnx-v1, and checks decoding limits before generation. It launches a Python worker without a shell, sends a bounded request, validates returned choices, and stops active workers on cancellation or shutdown.

The manifest contract fixes maxPairs = 2, maxSeqLen = 256, and maxNewTokens = 20. Those are request/decoding limits, not a model-parameter count or a quality score. The documented CPU dependencies are NumPy and ONNX Runtime.

3. Language, audience, and privacy scope

The current pack declares English input and output, including supported locale variants; it does not provide translation. The adapter refuses unsupported language pairs and refuses an adult profile for this child checkpoint. Changing settings does not retrain the model or broaden its evaluated audience.

The adapter writes no transcript and its workers own no durable conversation data. That is narrower than a claim that every application or selected provider retains nothing. The host owns the conversation session and delivery path; its storage and privacy policy still need review.

Offline execution refers to an installed local pack and its runtime dependencies. It is not a claim that a hosted webpage or every TinkySpeak feature works offline.

4. Evidence and reproducibility

The existing offline documentation describes testing the installed CLI under a macOS network-denying rule, including generation, selection, delivery, history, reset, and refusal of cloud or unsupported-language switches:

node scripts/check-tinkymind.mjs

That is a documented historical runtime check. It was not rerun for this report, and the cited 24 MB artifact size was not remeasured here. Source inspection on 8 October 2026 confirmed manifest validation, language/audience refusals, bounded worker handling, and the separation of proposals from the surrounding protocol.

The separate AAC white paper reports protocol evidence and current failures. Protocol correctness is not a clinical efficacy result or proof that generated choices are appropriate for each user.

5. Limitations and next experiments

• The checkpoint is child-trained and English-only; it is not validated for every age, language, diagnosis, or communication style.
• It is a neural language model. Offline inference does not remove the possibility of an inappropriate or unsupported suggestion.
• Artifact size is not peak runtime memory; CPU inference is not a measured battery-life claim.
• User approval and host-confirmed delivery are necessary boundaries, not a universal guarantee against harm.
• No $50-chip, 90% savings, or clinical-certification result is established here.

Further work should measure appropriate-choice coverage, unwanted suggestions, refusal usefulness, user correction burden, accessible selection, and delivery failures with AAC users and specialists. Hardware evaluation must report the checkpoint digest, device, peak memory, latency distribution, and measured energy under an offline rule.

6. Primary sources and related research

1. TinkyMind models and privacy documentation: documented pack, CPU execution, offline checks, and checkpoint scope.
2. Perslis.AAC: User-Controlled Conversation: protocol, proposal-to-delivery boundary, evidence status, and failures.
3. Inspected implementation: api/_lib/aac_runtime/tinkymind.js and capabilities.js, local checkout reviewed 8 October 2026.
4. Communication-runtime commands: model selection, reset, and explicit delivery.