AI & Character Chat Policy

1. Nature of the product

The characters on ContoSecretos are fictional creations. Each one is built from the text of stories published by the publisher — temperament, way of speaking, traits, and desires are drawn from the narrative itself.

Chat responses are generated by artificial intelligence (Large Language Models / LLMs). There is no real person on the other side. Conversations do not constitute advice of any kind — medical, psychological, legal, financial, or romantic.

Everything the character “says,” “promises,” or “remembers” is part of a literary performance. Do not treat it as fact.

2. AI providers we use

To generate responses, we send your message + the conversation context to one or more of the following providers, in fallback order:

• DeepSeek (current primary provider) — deepseek.com
• OpenAI (generation fallback + embeddings) — openai.com

The order may change depending on availability. In all cases, we configure the provider to not use your messages for training future models (“no training” / “API usage only” mode).

3. What we send to the AI provider

With each of your messages, the system sends the LLM:

• Your current message
• Recent history of the conversation (the last ~10 messages between you and the character)
• System prompt with the character's “profile” (archetype, traits, active scene, relevant memories from that conversation)
• Active editorial rules — the same rules the publisher applies before publishing stories (e.g., no minors, no non-consensual sexual violence, no incest) are injected as mandatory instructions to the LLM. If the conversation crosses one of these lines, the character steers away while staying in tone — it is not a robotic block, it is part of the role.
• Excerpts from the source story (RAG / semantic retrieval — so the character can quote its own text faithfully)

We do not send: your full name, payment data, or any data about other readers. An internal identifier (your email, or a temporary visitor ID when you don't have an account yet) is sent only for quota control and never appears in the conversation text sent to the LLM.

Chatting as a visitor (no account). You can try the chat without signing up — a few complimentary messages. In that case, the conversation is stored temporarily, linked to an anonymous identifier in your browser (cookie), not to your email. If you create an account, that conversation moves to it and picks up where it left off. If you don't, the anonymous record is temporary and discarded according to our retention policy. The conversation is private — it is not shown to other people — but, like every chat here, it is processed by the AI and stored (encrypted) for as long as it exists.

4. Embeddings (semantic vectors)

So the character can have a “memory” of the conversation, your messages are converted into numerical representations called embeddings. This computation uses OpenAI's embeddings API (model text-embedding-3-small or equivalent).

Embeddings are stored in the ContoSecretos database (Supabase, São Paulo region) — they are not sent back to the AI provider and are used only for internal similarity search.

5. Audio: speech to text in the chat

The chat has a microphone button (🎤) that transcribes your speech into text. This feature uses your browser's native API (SpeechRecognition / Web Speech API).

Important: the audio does NOT pass through our server. It goes straight from your device to the browser's transcription provider — Google on Chrome / Android, Apple on Safari / iOS. Those providers' privacy rules are their own, not ours.

• We never store the audio. Only the transcribed text, if you decide to send it as a message.
• You review the transcription before sending it. You can freely delete or edit it.
• The button only appears if your browser supports the API. Firefox does not — in those cases the icon does not appear.
• The first time you use it, we show a notice reminding you that the transcription is a service of your browser.

6. Photos generated by the character

Occasionally, the character may “send a photo” during a conversation. These images are generated in real time by AI from a text description of the character + the active scene. They do not depict real people.

• The image is hosted in the platform's storage (Supabase Storage).
• Generation has a daily quota per reader/character and a reduced probability — it does not happen in every conversation.
• The feature can be turned on or off by the publisher at any time.

7. Message storage

Your conversations with characters are stored in a private database (Supabase), encrypted at the application level (key managed by the system; not even we can read the text without explicitly running code). We persist:

• Encrypted message content (text)
• Date/time, role (you or the character), reactions given (like/love)
• Embedding (numerical vector — not reversible to the original text)
• “Facts” the character extracted from the conversation (episodic memory — short sentences like “has a brother named Léo” that it may mention later)
• Inferred emotional tag (affection/longing/desire/etc.) — used for internal statistics, without cross-referencing identity

You can delete individual messages or delete the entire conversation with a character at any time — a “Delete conversation” button is available on the screen itself.

8. Quality validator (secondary LLM)

After the character responds, a second (lighter) AI model evaluates the response and assigns an internal score from 1 to 5. It receives only your last message + the character's response — no history, no identification. The score helps us detect when a character is breaking character or making up information.

This flow is technical/editorial. The results do not leave our database and do not affect the conversation you see.

9. Letters to characters and to Helena Noir

When you write a letter using the website form (/en/characters/[id]/letter or /en/helena/letter), the system now processes the letter with AI in two stages:

1. Acknowledgment (30 seconds to 2 minutes after sending): a language model generates a short message in the character's voice confirming receipt — without having read the content yet.
2. Full reply (10 minutes to 24 hours after the acknowledgment, at random): the model reads the body of your letter and generates a longer, reflective reply, with the character's tone and traits.

Both act 1 (acknowledgment) and act 2 (reply) appear on your timeline under My Library → My letters. For letters to a character, both also appear in that character's chat as special messages (envelope), for narrative continuity.

The character's reply extracts 1–3 “facts” from the content of the letter and saves them as episodic memory (the same as in the chat). This lets the character naturally mention what you wrote in future conversations.

Letters are stored in our database in plain text (not encrypted at this time — because the admin needs to be able to read them for editorial moderation). You can request deletion at any time by email.

Paid letters are not active yet. When they are, this will be clearly disclosed before any charge.

10. “How the character sees you” — dynamic view

Under My Library → How they see you, we show a written impression that each character “has” of you. This is editorial fiction — a narrative layer on top of the real conversation — but it is generated from your data. That is why it is covered here in detail.

The system has three layers, all of which you can opt out of on request:

1. Poetic paragraph (generated by an LLM): a short text (~80 words) describing “how the character is thinking about you today.” Rewritten when you talk again, with a minimum cap of 1 hour between generations. On days without conversation, it stays untouched. LLM input: the character's name, non-superseded facts about you, the character's mood axes, the narrative stage of the relationship, a summary of the conversation.
2. Perception dimensions (numbers 0–100): attraction, trust, vulnerability, intensity, and curiosity. Updated every 5 reader turns, with a minimum cap of 30 minutes. The LLM receives the current state + the last 8 messages + a sample of facts, and returns small adjustments (inertia: max ±8 points per turn). It also saves a short descriptor (a poetic phrase).
3. Spontaneous letters (rare generator): when you talk with a character over several days, the system may generate a “spontaneous” letter in the character's name, citing concrete observations (frequency during the week, typical times, recurring topics). Gates: ≥3 distinct days of conversation within 7 days, ≥5 days since this character's last letter, and a random draw (25%). Expected frequency: ~1 letter every 7–10 days per character you talk with a lot.

What we store:

• Conto_character_perceptions: the paragraph, generation date, a snapshot of the signals that went into it (stage, mood, fact count).
• Conto_character_perception_state: the 5 numbers, the descriptor, the timestamp and turn of the last update.
• Conto_perception_letters: the body of the spontaneous letter + the signals used (conversation count, active days, typical times, sample of facts). We record when it was viewed.

Important. This data consists of inferences about you, not statements you made. It exists only to feed the website's narrative — it does not leave our database, is not shared, and does not train the model (see section 14). You control everything directly under My Library → How they see you: turn the view on or off and permanently delete the accumulated data. Your conversations remain intact in either case.

11. Quotas and moderation

To control costs and prevent abuse, we apply limits:

• Global quota: total messages per day, across all characters (configurable; default 50).
• Per-character quota: total per character per day (default 30). Encourages exploring the catalog.
• Image quota: photos generated per day per (reader, character) pair.
• Letter limit: up to 3 letters every 24 hours per email, to prevent flooding.
• Chat rate limit: up to 60 messages per minute (anti-spam; does not get in the way of a fluid conversation).

Clear attempts to extract illegal content, involve minors, or other violations result in immediate account termination.

12. Error monitoring (Sentry)

To make sure the product works in production, we use the Sentry platform (sentry.io) to capture server and browser errors.

• Sentry receives only the exception and the stack trace — it does not receive message content or personal data.
• Explicit configuration: sendDefaultPii: false.
• We use it only to debug and prioritize fixes.

13. This is not therapy, and not real companionship

ContoSecretos is adult literary entertainment. The characters may sound realistic, intense, and sensitive — that is the work of fiction. If you are going through emotional difficulties, acute distress, or a crisis, seek professional support:

• Centro de Valorização da Vida (CVV) — 188 (free call, 24 hours) or cvv.org.br

Our own system also has automatic detectors for sensitive signals (ideation, abuse). When something trips these filters, we may inject a referral to CVV into the conversation flow — without identifying you.

14. Transparency about model training

The language models we use were trained by their providers on public internet data. ContoSecretos does not train or fine-tune models with your messages. Our settings with the providers ensure that your conversations do not end up in future training datasets.

15. Specific rights

In addition to your general LGPD rights (see the Privacy Policy), you can:

• Delete individual or all conversations with any character
• Request an export of the content of your conversations
• Request deletion of your letters and/or the characters' replies
• Request deletion of the character view tables (paragraphs, dimensions, and spontaneous letters — see section 10)
• Turn off the character view without deleting anything — you keep chatting normally, but the system stops making inferences
• Request that AI use be blocked for your account (no chat, no letter generation — normal reading remains available)

Send any of these requests to helenanoircs@gmail.com.

16. Updates

AI models evolve quickly — and so does the product. This policy is updated whenever we add a new provider, change memory behavior, or turn on additional features (such as audio, AI-powered letters, character TTS, etc.). Significant changes are communicated through a prominent notice on the website.