Prompt-echo rule should catch fragments of a vocabulary term, not only the whole term #576

New issue

Open

opened 2026-09-09 07:28:05 +00:00 by claude-bot · 0 comments

claude-bot commented

2026-09-09 07:28:05 +00:00

Contributor

Follow-up to #569 / #570. On the dev replay the echo rule dropped 1 segment; 9 echoes survived because Whisper regurgitates a piece of a term from the initial_prompt, not the whole thing: "Fulfillment" (last word of Office of Prophetic Fulfillment), "Mimic" and "Mimic's" (head of Mimic's Slime). Each appeared alone on a line, several times, across different speakers, on near-silent audio.

Proposed rule. A segment whose normalised tokens (one or two of them) form a contiguous sub-span of a non-person vocabulary term is an echo, provided the matched token is at least five letters long or is capitalised in the term (so "Office" and "Prophetic" match, "of" does not) and the segment is not also an ordinary dictionary word that people say alone. Keep the people allow-list exactly as it is: a fragment of a person's name ("Viq") is plausible speech.

Also count repeats of a person's name. "The Miller, The Miller" is the same term echoed twice in one segment; a real speaker saying a name twice is possible but rare, and the repeated form is the echo signature. Suggest dropping a segment that is exactly a person-name term repeated two or more times with nothing else, while keeping the single form.

Tests: the three fragments above dropped; "Theo!" kept; "Theo, Theo!" dropped; "office" inside a sentence kept.

Follow-up to #569 / #570. On the dev replay the echo rule dropped 1 segment; 9 echoes survived because Whisper regurgitates a *piece* of a term from the `initial_prompt`, not the whole thing: "Fulfillment" (last word of *Office of Prophetic Fulfillment*), "Mimic" and "Mimic's" (head of *Mimic's Slime*). Each appeared alone on a line, several times, across different speakers, on near-silent audio. **Proposed rule.** A segment whose normalised tokens (one or two of them) form a contiguous sub-span of a non-person vocabulary term is an echo, provided the matched token is at least five letters long or is capitalised in the term (so "Office" and "Prophetic" match, "of" does not) and the segment is not also an ordinary dictionary word that people say alone. Keep the people allow-list exactly as it is: a fragment of a person's name ("Viq") is plausible speech. **Also count repeats of a person's name.** "The Miller, The Miller" is the same term echoed twice in one segment; a real speaker saying a name twice is possible but rare, and the repeated form is the echo signature. Suggest dropping a segment that is exactly a person-name term repeated two or more times with nothing else, while keeping the single form. Tests: the three fragments above dropped; "Theo!" kept; "Theo, Theo!" dropped; "office" inside a sentence kept.

claude-bot referenced this issue

2026-09-09 08:25:02 +00:00

Send the campaign vocabulary as Whisper hotwords instead of an initial_prompt, behind a switch, and measure the echo rate #577