AI & automation transparency

• Context halo ranking. The eight adaptive candidates beside the permanent grid are chosen by explicit rules over the configuration an adult entered plus device-local signals — time of day, the selected routine, place class, the chosen communication partner, recently used concepts and personal associations. It is ranking, not generation. It has no write path to the permanent grid, and it cannot move a permanent cell.
• Bounded phrase expansion. Selecting words such as MUM and CAR can offer three or four fuller phrasings drawn from a defined rule set. The options are bounded and inspectable; nothing is invented by a language model.
• Partner speech to response candidates. Where a listening mode is enabled and the browser provides speech recognition, recognised partner speech is mapped by rules to candidate responses. Recognition itself is performed by the browser; we do not receive or store the audio. Where the browser lacks the capability, a clearly labelled manual mode is shown instead.
• Word finder. Search over vocabulary that highlights the route to a word. It never relocates vocabulary to make it easier to reach.
• Derived insight measures. Arithmetic over configuration and interaction counts. No model, no classification of the person.

The rule that governs everything above: software may propose, rank or expand; only the communicator may author. No generated or suggested candidate can be spoken, added to the utterance ribbon, sent, or recorded as the communicator's words without an explicit selection by the communicator.

Partner-facing model features can highlight the corresponding cells to teach a route, but they cannot insert into the communicator's message or trigger their speech. This gate is covered by automated tests in our build, so a regression fails the build rather than shipping quietly.

• no cloud or third-party AI service applied to a communicator's utterances;
• no training of any model on a communicator's messages;
• no emotion recognition, sentiment scoring or affect detection;
• no inference of behaviour, competence, compliance or hidden intent, and no assertion of what someone "really meant";
• no automated decision-making with legal or similarly significant effects;
• no profiling for advertising, and no advertising technology at all;
• no auto-speech, auto-send or auto-selection of a suggestion.

Communication-gap reporting is deliberately literal: it says "no permanent vocabulary item exists for trampoline", not "the child wanted the trampoline". The first is an observable fact about the system; the second would be a guess about a person.

Wherever suggestions could otherwise present a fixed menu, EyeSay always includes a MORE or SOMETHING ELSE escape, and the permanent agency words — STOP, NO, HELP, BREAK, TOILET — are always available. A person must never be limited to the options a system thought of for them.

Insights shown to guardians and professionals use derived measures — vocabulary breadth, independent selections, unresolved searches, mix of communication functions, access errors. They do not present raw message transcripts by default, and they are not a behavioural profile. See the child privacy notice.

We may add stronger on-device models, or optional server-side assistance for adult-facing work such as drafting a vocabulary set. If we ever do, then before it operates we will update this page and the privacy notice, name any AI provider in the provider register, and keep the authorship gate intact. Applying cloud AI to a communicator's own utterances would be a material change requiring notice and an explicit adult choice — never a silent default.

Where EU AI Act transparency obligations apply to a future feature, we will meet them, including making clear when a person is interacting with an automated system.

Questions about how a specific feature works are welcome — Prodia Systems Limited, 27 Pembroke Street Upper, Dublin 2, D02 X361, Ireland, contact@prodiasystems.com.