ListeningKitVoice gate
Every message log sends as matt passes the voice gate first (log voice, MCP voice). The gate asks SubQ whether matt typed it himself, reading his own writing across all his conversations. Rules 90, 98, 111 and 112.
Reading
REQ-VOICE-01: Every draft is read against matt's writing across all his conversations
Every draft is sent to SubQ with matt's writing guide, the DMs he wrote himself, and his real replies and closest wording from all his conversations, whether or not there is a chat with the person.
Checked by: code and SubQ
REQ-VOICE-02: A chat with the person adds that chat's evidence
When there is a chat with the person, the draft is also read against matt's replies, closest wording and latest messages in that chat, and SubQ is asked whether it reads like him writing in that chat.
Checked by: SubQ
REQ-VOICE-03: Phrases matt never says fail the draft
A draft that contains a phrase from the list of phrases matt never says fails, whatever SubQ answers.
Checked by: code and SubQ
Opener
REQ-VOICE-04: A first message is only the opener
A first message to someone matt has never talked with fails when it carries an ask: a pitch, a question about a tool, product or service of his, a call, or a link to his own stuff. Rule 98. A chat where matt has already sent messages (the voice index's latest messages in that chat) is never judged as a first message: SubQ is told so (alreadyTalkedInThisChat) and the opener check passes.
Checked by: SubQ
Watermark
REQ-VOICE-05: A draft with AI artifacts fails
A draft whose watermark reading finds artifacts (zero-width or other hidden characters, odd spaces, pasted Claude HTML) fails. Rule 111.
Checked by: code and SubQ
Pass
REQ-VOICE-06: A draft passes only when every answer passes
A draft passes only when the voice answer reaches its threshold, the context, friend-like and opener answers reach theirs, the chat answer reaches its threshold when there is a chat, the builds-on answer reaches its threshold when the draft is spoken (REQ-VOICE-13), no never-said phrase is found and the watermark finds no artifacts.
Checked by: code and SubQ
His own phrasing and the scrambler
REQ-VOICE-07: A draft written in word pairs matt has never written fails
Every reply in the voice index is cut into pairs of words. The gate counts how many of the draft's pairs matt has written at least twice, leaving out pairs that use words from what the other person said. A draft whose share is below matt's own 10th percentile fails, measured over his replies with each reply's own pairs left out. The pairs he has never written are named in the result and handed to SubQ, so "is wild" fails because matt never wrote it.
Checked by: code and SubQ
REQ-VOICE-08: The vendored watermark scrambler reads every draft
The draft goes through vendor/aiWatermarkScrambler (novatechflow/ai-watermark-scrambler, layer A) for invisible and lookalike characters. A draft with any fails, and its count and kinds are handed to SubQ. The statistical watermark is not checked here.
Checked by: code and SubQ
REQ-VOICE-09: The gate runs inside every send and refuses a failing draft
Every send as matt (Discord, Slack, an X dm, the X queue's dms, replies and quotes, Instagram text) runs the whole gate right before it goes out, after the account has caught up and before the chat is claimed, whether or not log voice ran first. A failing draft is refused with voiceGateFailed and nothing is sent; in the X queue the item is held with what failed. A pass is one line; a fail names only the checks that failed and why. Rule 137.
Checked by: code and SubQ
REQ-VOICE-10: The chat's context is ranked for the moment and given in full
The gate takes the chat's messages from the .db (both ways), embeds them and what is being answered, keeps the closest 40 by cosine, reranks those with jina-reranker-v3.5, and gives SubQ the top 12 in full, in time order, with its question on whether the draft answers them in this conversation. Rule 137.
Checked by: code and SubQ
REQ-VOICE-11: Only embeddings are stored, each once, in the .db
A vector is stored in the .db's embeddings table by the text's hash, the model and the dimensions, and an embedder looks there first and embeds only what has no vector yet. The AI tells index's phrase list and matt's baseline are in aiIndexes. No reading, score or verdict is saved. Rule 137.
Checked by: code and SubQ
REQ-VOICE-12: The gate inside a send reads matt's real replies in that chat
The gate inside a send finds the chat in matt's saved conversations by what the send knows (chatIdFor): its conversation id on that platform, or, on X, the conversation whose pair holds the person's user id, or their handle on that platform only. It gives SubQ his real replies there, the same evidence log voice --with gives. When the chat isn't there, the gate still reads his writing everywhere. On 2026-11-04 drafts that voice passed at 0.99 were refused inside the send at 0.36, because the send gave SubQ none of them. The opener rule (rule 98) only reads a first message: a chat with matt's own messages in its history, or his replies from that chat or that person, is already talking (alreadyTalkedIn), and openerOnly is not asked. What the send answers (momentAnswered) is every message they sent after matt last spoke, oldest first, one per line, so a run of messages is answered as a whole; the chat is ranked for the gate against that together with the draft itself, so the earlier moment a draft builds on is in front of SubQ (on 2026-11-05 a pitch to kewlkid passed voice at 0.89 and was refused in the send at 0.34, ranked against his last line alone). When matt's own message is the newest, the send is news of his own, so nothing is passed as what they said and the gate asks whether it picks up on something specific about them. Rules 98 and 137.
Checked by: code and SubQ
REQ-VOICE-13: A voice note is graded as spoken and must build on the conversation
readVoice in src/twitter/copy/voice/client.ts takes spoken. With it the voice question asks whether matt would say it himself in a voice note, its length and punctuation never counting against it; toThem asks the same of how he talks in this chat; and a buildsOn question asks whether it takes what they are talking about further, across the whole conversation, rather than repeating what matt already told them, which must reach CONTEXT_PASS to pass. Without spoken there is no buildsOn (it is null). gateReading in src/cli/twitter/helpers/voice.ts passes spoken on and skips offProfile for it, gateSend takes spoken, failedParts names a low buildsOn, and log voice --spoken gates a voice note the same way. With spoken, gateReading also hands readVoice howMattTalks(): the AssemblyAI transcripts of matt's own voice notes (mattsOwnVoiceNotes in src/extensions/voiceClone/helpers/recordings.ts, only the ones recorded before firstClonedAt, so none the clone made), given to SubQ as howMattTalksInHisOwnVoiceNotes, which the spoken voice question reads first. Rule 149.
Checked by: code and SubQ
REQ-VOICE-14: matt can send one dm over the gate's refusal
log twitter dm takes --overGate (OVER_GATE in src/cli/twitter/helpers/send.ts). With it, the send skips gateDeps.gate and says so on its error stream; without it the gate runs and a refusal is voiceGateFailed. The copy's AI artifact check (planSend, aiArtifacts) runs either way, before the gate. log discord send takes --overGate the same way (its flag in DISCORD_SEND_INPUT, src/cli/discord/index.ts): with it the send skips gateDeps.gate and says so, without it a refusal is voiceGateFailed, and its AI artifact check still runs first. matt, 2026-11-05: a message to kewlkid in his own words, and even his GitHub username on its own, were refused by the gate. Rule 152.
Checked by: code and SubQ
REQ-VOICE-19: A task is gated as a brief, and its files go with it
log discord send --brief passes brief and the attached files to gateSend, which, after the sameness check, hands the draft to gateBrief instead of the voice reading. gateBrief refuses a draft the scrambler finds suspicious, then asks SubQ (readBrief, src/voice/brief/client.ts) the five BRIEF_CHECKS (clear, why, doneWhen, kind, steerable) and fitsTheRoom when there is a room, with the chat's ranked moments, matt's writing guide and each attached file (attachedOf: its name, and the first ATTACHED_TEXT_CHARS characters of a text file). It passes only when every answer is at least BRIEF_PASS (0.7); a refusal names each failed check (briefFailedParts). --file attaches files to the last message only (filesOf): each must be under OUTBOX or ~/.logCli and at most 10 MB, else fileRefused before anything is sent. matt, 2026-11-07. Rule 162.
Checked by: code and SubQ
REQ-VOICE-18: A draft that just passed is not sent to SubQ again
When gateSend or the log voice command passes a draft, it records the pass by the draft's text (recordPass, src/voice/passes). gateSend then answers a send of the same text as passed without asking SubQ (recentPass) when that pass is under PASS_REUSE_MS (10 minutes) old and nothing has come in on the chat since (incomingSince). The sameness check (REQ-VOICE-16) still runs first. On a Discord DM or group whose .db is current (readSourceFor), the room the gate reads comes from chatMessages, not a read from Discord. matt, 2026-11-07. Rule 161.
Checked by: code and SubQ
REQ-VOICE-17: A refused draft says where to look before it is rewritten
When gateSend in src/cli/twitter/helpers/voice.ts refuses a draft with voiceGateFailed, or sameAsSent refuses it, the refusal ends with rewriteSteps: call voice with the person (and what they said) and no text, to read their thread and matt's real replies to things like it; call search for what matt has said on it before; then write from those, unlike anything sent today. matt, 2026-11-07: "you don't even try to look ... you don't try to search". Rule 158.
Checked by: code and SubQ
REQ-VOICE-16: No two messages read alike
sameAsSent in src/cli/twitter/helpers/voice.ts reads what the account sent in the last SAMENESS_WINDOW_HOURS (24) from the .db (ownTextsSince in src/inbox/store/client.ts, one text per conversation) and refuses the draft when samenessOf in src/voice/sameness/client.ts finds half or more of its word pairs (SAMENESS_LIMIT, 0.5), with links left out, already in one of them; a draft with fewer than SAMENESS_MIN_PAIRS (3) word pairs is never refused. gateSend runs it through gateDeps.same before the voice gate's reading, and the twitter dm and Discord sends run it before their overGate skip, refusing sameAsSent. itemProblems in src/twitter/queue/plan/client.ts adds a sameAsSent problem for every pending item that batchSameness finds repeating an earlier item of the same queue. matt, 2026-11-07: the share-asks that day differed only in their first line. Rule 157.
Checked by: code and SubQ
REQ-VOICE-15: Every send and voice note is gated against the room it goes to
gateSend in src/cli/twitter/helpers/voice.ts passes a room to gateReading: the one given (room), the one its readRoom reads, or one built with roomOf from the chat in the .db. roomOf in src/voice/room/index.ts gives the chat's kind, who else spoke in the last 24 hours, the last of them and how many days ago, how many of matt's messages end the chat, whether nobody else has spoken lately, and addressTo, which for a channel or group with nobody else lately says to talk to it as a whole. readVoice in src/twitter/copy/voice/client.ts gives it to SubQ as theRoom and asks fitsTheRoom, and a draft scoring under CONTEXT_PASS fails. The Discord send and voice note pass a readRoom that reads the chat's last 20 messages through the engine with roomFromRead. When nobody else has spoken lately in a channel or group (nobodyToAnswer), there is nothing to answer, so inContext doesn't count towards the pass and isn't named in a refusal; fitsTheRoom, buildsOn and the voice still do. In a direct message where matt sent the latest messages and they haven't answered yet, fitsTheRoom counts a follow-up that carries on from what matt just sent them (a question about their own post he just shared) as addressed to them as the chat really is, and fails it only when it acts as if they had replied. Rules 155, 156.
Checked by: code and SubQ