Responsible use of artificial intelligence

• Deterministic first: operational figures, charges, deadlines, and statuses are calculated with verifiable rules on the customer's data. A model can explain or draft; it does not invent or correct a number.
• The person decides: a model output is an input. Every action with an effect on data, money, or people requires confirmation from a user with permissions, and is logged with its author.
• Minimal data: only what is essential for the task is sent to an external provider, with license plates and names replaced by placeholders when the task allows it.
• No training with customer data: neither we nor our providers use a customer's data to train models, except under a written, specific, and revocable agreement.
• Declared status: we state what is in production, what is in the lab, and what is only planned. A use that does not exist is not described as available.

When a company enables the Michiva language layer, queries are sent via API through OpenRouter, Inc. (United States), which directs them to the provider of the chosen model. The intended providers are OpenAI and Anthropic, and the subprocessor register indicates which one is in use. We use each provider's commercial API terms, not those of their consumer apps.

• No training: every call requires that the provider not use the data to train or improve its models.
• Zero retention: by default we require providers that do not retain the request or the response. If a company accepts a model without that guarantee, the provider may keep abuse-monitoring logs for up to 30 days, and this is disclosed before enabling it.
• International transfer: sending data to these providers is a transfer outside Chile and the European Union, described in the international transfers policy.

If we later allow a customer to use its own Anthropic or OpenAI API key, that use will also be governed by the API terms the customer has with that provider. The connection will be by API key only, without signing in with a Claude or ChatGPT account. That option does not exist today.

• Human review: no model output by itself becomes a measure concerning a person, a charge, or a tax document.
• Validation of drafted text: text drafted by a model is checked against the data before being displayed; if it does not match, a fixed template is used.
• No conversation memory: Drometra does not store Michiva's question or answer. It logs who queried, for what purpose, which model responded, the result, the tokens, and the cost, and deletes that log after 90 days.
• Limits of use: 200 model queries per company per day and 10 per person per minute, to contain costs and abuse.
• Tracability: every call is associated with the model version used, so that a result can be reviewed later.
• Notice and opt-out: before enabling the language layer for a company, we inform it, and its administrator may request that it not be used or that it be disabled at any time; the fixed questions continue to work without it.

Michiva and any other Drometra assistant explain operational data. They do not recommend investments, loans, prices, or tax strategies, do not interpret the law for a specific case, and do not make financial decisions for the customer. An answer that looks like advice of that kind must be treated as an error and reported.

Nor do they assess a person's health, fatigue or fitness, or give medical guidance, even when they show driving events or Vision alerts. In an emergency on the road in Chile, call 131, 132 or 133.

• A language model can misinterpret a question or write something incorrect with apparent confidence. That is why we limit its role to choosing among fixed questions and drafting from already calculated facts.
• The Vision detector performs worse in low light, backlighting, heavy rain, and with a dirty or misaligned lens, and has only been tested on one phone model.
• A detection does not prove that something happened, and the absence of a detection does not prove that it did not happen.
• The MediaPipe library used by Vision may send Google technical performance and usage metrics when the phone has a connection. It does not send video or images.
• Route optimization works with estimated times based on road data that may be out of date, especially outside urban areas.

• Facial recognition of third parties captured on public roads or of a store's customers.
• Inference of emotions, mood, or personality traits from the face or voice.
• Tracking a person from one day to the next in Drometra Spaces.
• General scoring of people outside the scope of safe driving.
• Fully automated healing, without human intervention or the right to discharge.
• Identification of participation in trade union or representation activities.

This list is a public commitment. If it changes, there will be a record of what changed and why.