ListeningKitAI tells and watermark
Every X and Discord send's copy is read for what AI chat interfaces leave in text (the watermark reading), and every draft the voice gate reads is also read for AI phrasing against curated lists and matt's own writing. Rules 111 and 112.
Watermark reading
REQ-AITELLS-01: The reading is clean, stylistic or artifacts
The copy is read by the vendored claude-watermark-checker. Any deterministic finding (zero-width or other hidden characters, odd spaces, pasted Claude HTML) makes it artifacts; only curly quotes or em dashes make it stylistic; nothing found is clean. The statistical watermark is always said to be unknown.
Checked by: code and SubQ
REQ-AITELLS-02: planSend refuses copy with artifacts
planSend refuses a send whose copy reads as artifacts, with the code aiArtifacts, naming what was found. Nothing is sent.
Checked by: code and SubQ
REQ-AITELLS-03: The plan carries the reading and every send logs it
A send planSend allows carries its watermark reading, and both a direct send and each queue send log it in one line.
Checked by: code and SubQ
REQ-AITELLS-13: A Discord send is read the same way before anything goes out
log discord send (MCP discordSend, discordSendLive) reads the whole message with the same watermark reading before the dry run or any send. Copy that reads as artifacts is refused with aiArtifacts, naming what was found, and nothing is sent. Otherwise the reading is logged in one line, printed in the dry run and carried in its --json, live or not. Rule 111.
Checked by: code and SubQ
AI phrasing
REQ-AITELLS-04: vale-ai-tells rules run on the draft
The tbhb/vale-ai-tells rules (vendor/aiTells/valeAiTells/styles, its substitution and existence rules) are run on the draft the way Vale would, and any hit that is not muted fails it.
Checked by: code and SubQ
REQ-AITELLS-05: The index embeds every literal phrase with Jina
log voice aiIndex collects every literal phrase from vale-ai-tells first, then ai-slop-detector, ai-vocabulary, SLOP_Detector, the X complaints and the AI-written samples, and embeds them with Jina jina-embeddings-v5-omni-small at 1024 dimensions. The vectors and the index are stored in the .db, and an index made with another model is rebuilt (rule 137).
Checked by: code and SubQ
REQ-AITELLS-06: Phrases quoted on X are kept only when close to a curated one
A phrase from the X complaints list is kept in the index only when it sits at cosine 0.6 or more from a curated phrase.
Checked by: code and SubQ
REQ-AITELLS-07: matt's own p50 and p99 are measured
The index measures matt's own messages from the voice index against the phrases and keeps his p50 and p99 closeness.
Checked by: code and SubQ
REQ-AITELLS-08: Vale rules matt uses himself are muted
A vale rule that hits in over 2% of matt's own messages is muted and never fails a draft.
Checked by: code and SubQ
REQ-AITELLS-09: Phrases matt says himself never fail a draft
A phrase vale flags that matt says himself in 5 or more of his own messages is his voice and never fails a draft.
Checked by: code and SubQ
REQ-AITELLS-10: Each chunk's closest AI phrase is found and reranked
The draft is cut into lines and word windows, each chunk's closest AI phrase is found by cosine, and the closest are reranked with jina-reranker-v3.5.
Checked by: code and SubQ
REQ-AITELLS-11: A draft closer to AI phrasing than matt's p99 fails
The draft fails when a vale rule hits or its closest chunk sits closer to an AI phrase than matt's own p99.
Checked by: code and SubQ
REQ-AITELLS-12: The voice gate keeps the index current and fails on AI tells
log voice builds the AI tells index itself when it is missing or older than matt's voice index, and a draft whose AI tells reading fails does not pass the gate.
Checked by: code and SubQ