Attune — privacy, stated plainly

Last updated 2026-07-13. This page describes what actually happens to your data. It is written to be checkable, not reassuring.

What is stored, and what never is

• Your messages are not stored — with one narrow, named exception. Goals, drafts, and generated candidates are processed and discarded, except: to sound like you, we keep a rolling window of your last few own messages as voice samples (never the recipient's messages, never the generated candidates). That is your own content, and it is wiped by "erase" or by "clear it" on the options page.
• If you save a recipient, we store your channel model: the description you wrote, the tone each message was predicted to carry, a count of flagged misread risks, and outcome records (whether a message landed as predicted, judged from the reply you later run through the tool — one line of evidence, not the reply itself). Never the message text. If your description includes a person's name, that name is stored with the channel — it stays visible to you, editable, and deletable. If you confirm a match suggested by the browser extension, the person's address is stored with the channel too, so the right channel loads next time; addresses are visible in your export and erased with everything else.
• The browser extension reads only at the moment you invoke it: your selection, the To/sender line, and the message you are replying to — from that message only, never the rest of the thread, never in the background. What it reads is used for that run and not stored (except a confirmed address, above). Inserting a result into your compose box happens only on your click; the extension never touches send.
• Run counters: your invite code, the date, and a count — for rate limiting.
• Waitlist notification (opt-in only): if you ask to be emailed when a trial spot opens, that address is stored until the one note is sent, then deleted. It is the only raw email address the system ever holds; sign-in identities are stored as hashes.
• No third-party analytics, no trackers, no advertising cookies; the only thing kept in your own browser is your invite code or sign-in session. What we do review — our own predictions against real outcomes, to make the tool better — is described next, not hidden here behind "no analytics".

How the stored outcomes are used

The outcome records above — what tone a message was predicted to carry, how many misread risks were flagged, and whether it landed as predicted — are reviewed in aggregate, across accounts, to improve the tool's prediction quality. That is the whole reason they are kept: a prediction the reply later proves wrong is how the tool learns to predict better. What is reviewed is the model's own characterizations and the landed / did-not-land outcomes, tied to a hashed account key — never to a real identity, never to your message text. It is used only to make the encode and decode more accurate. Never sold, never used to profile you, never used for any purpose unrelated to the tool working better.

Where your text goes when you run a check

Your input is sent over TLS to our server (Google Cloud Run, region Taiwan), which forwards it to Anthropic's Claude API for processing. Under Anthropic's commercial API terms it is not used to train models. We keep no copy after the response is generated. This API hop is the one place your content transits a third party, and this sentence is the exact boundary of what we can promise.

If our storage were breached

An attacker would obtain: run counters, feedback notes you typed, and channel characterizations such as "terse, dislikes hedging." Not your messages — they are not there. We state this so you can judge the actual stakes rather than trust an adjective.

Your rights, as buttons rather than promises

• Export: "Download everything stored for your code" on the main page returns the complete record — there is no hidden remainder.
• Erase: "erase it all" deletes it immediately and permanently.

The browser extension

The extension has one permission: adding two items to your right-click menu. It cannot read pages, history, or keystrokes. It sees only text you explicitly select and only when you invoke it, and its sole action is opening this site with that selection.

The tool never sends messages

Nothing you write here is transmitted to any recipient. You copy, you send, you remain the author. Predictions are priors, not verdicts.

Operated by Connected Information (bear@connectedinformation.io). Applicable frameworks: Taiwan PDPA, GDPR where it applies — consent, purpose limitation, erasure, as implemented above. Questions and deletion requests: email, answered by a human.

Attune — 隱私，講清楚

最後更新 2026-07-13。這個頁面描述你的資料實際上會發生什麼事。 它寫來是讓你查證的，不是讓你安心的。

什麼會被儲存，什麼永遠不會

• 你的訊息不會被儲存——只有一個範圍很窄、明白點名的例外。 目標、草稿與產生的候選訊息在處理後即丟棄；例外是：為了讓產出聽起來像你， 我們保留你最近幾則自己寫的訊息作為語氣樣本（絕不包含對方的訊息， 也絕不包含產生的候選）。那是你自己的內容，按「清除」或在選項頁按「清掉它」 即全部抹除。
• 如果你儲存一位收件人，我們儲存的是你的管道模型： 你寫下的描述、每則訊息當時被預測會帶的語氣、被標記的誤讀風險數量， 以及結果紀錄（訊息是否如預測落地——由你之後跑進工具的回覆判定； 留下的是一行證據，不是回覆本身）。永遠不存訊息內文。如果你的描述裡 有人名，那個名字會隨管道一起儲存——你隨時看得到、能編輯、能刪除。 如果你確認了瀏覽器擴充功能建議的配對，對方的地址也會隨管道儲存， 下次才載得出正確的管道；地址在你的匯出裡看得到，也隨其他資料一併清除。
• 瀏覽器擴充功能只在你召喚它的那一刻讀取：你選取的文字、 收件人／寄件人那一行，以及你正在回的那一封訊息——只有那一封， 不含討論串的其他部分，也絕不在背景讀取。讀到的內容只用於那一次執行， 不會儲存（前述已確認的地址除外）。把結果插入你的撰寫框只在你點擊時發生； 擴充功能永遠不碰寄送。
• 執行計數：你的邀請碼、日期、次數——用於限流。
• 候補通知（僅限主動選擇）：如果你要求在名額開放時收到 email，該地址會保存到那一封通知寄出為止，之後刪除。這是整個系統唯一 保存過的原始 email 地址；登入身分一律以雜湊儲存。
• 沒有第三方分析、沒有追蹤器、沒有廣告 cookie；存在你瀏覽器裡的只有 你的邀請碼或登入 session。我們確實會檢視的東西——自家的預測 對照真實結果，為了把工具做得更好——寫在下一節，不藏在「沒有分析」 這句話後面。

儲存的結果紀錄怎麼被使用

上述結果紀錄——訊息當時被預測帶什麼語氣、標記了幾個誤讀風險、 是否如預測落地——會以彙總方式、跨帳號檢視，用來改進工具的預測品質。 保存它們的理由就只有這一個：一個之後被回覆證明是錯的預測，正是工具 學會預測得更準的方式。被檢視的是模型自己寫下的描述與「如預期／不如預期」 的結果，繫在雜湊後的帳號鍵上——絕不連到真實身分，絕不連到你的訊息內文。 它只用於讓編碼與解碼更準確。絕不出售，絕不用來剖繪你，絕不用於任何 與把工具做好無關的目的。

你按下檢查時，文字去了哪裡

你的輸入經 TLS 傳到我們的伺服器（Google Cloud Run，台灣區域）， 再轉送 Anthropic 的 Claude API 處理。依 Anthropic 的商用 API 條款， 它不會被用來訓練模型。回應產生後，我們不保留任何副本。這個 API 中繼 是你的內容唯一經過第三方的地方，而這句話就是我們所能承諾的確切邊界。

如果我們的儲存被入侵

攻擊者能拿到：執行計數、你打的意見回饋，以及像「簡短，不喜歡拐彎抹角」 這樣的管道描述。拿不到你的訊息——因為那裡沒有。我們把這寫明， 是讓你能自行判斷實際的利害，而不是信任一個形容詞。

你的權利，是按鈕，不是承諾

• 匯出：主頁的「匯出」回傳儲存的完整紀錄—— 沒有藏起來的剩餘部分。
• 清除：「清除」立即、永久地刪掉一切。

瀏覽器擴充功能

擴充功能只有一項權限：在你的右鍵選單加入兩個項目。它不能讀取頁面、 瀏覽紀錄或按鍵。它只看得到你明確選取的文字，而且只在你召喚它的時候； 它唯一的動作，是帶著那段選取文字打開本網站。

這個工具永遠不會寄出訊息

你在這裡寫的任何東西，都不會被傳送給任何收件人。複製的是你， 寄出的是你，作者始終是你。預測是先驗，不是判決。

由 Connected Information 營運（bear@connectedinformation.io）。 適用框架:台灣個資法（PDPA），以及適用時的 GDPR——同意、目的限制、 刪除權，均如上文實作。問題與刪除請求：來信即可，由真人回覆。 兩個版本如有出入，以英文版為準。