Test results, not a project. What follows is the record of an experiment: two audio exhibits from the production of the Ranting with a Lunatic audiobook, October 19 chapter, recorded October 20, 2026.
The chapter is narrated by a speech synthesizer. The synthesizer is not an agent and was never spoken to. It only reads the text aloud. Twice in one chapter, it fell into a loop, and both times the text it choked on was text about an agent: instructions spoken to one, and the story of one being corrected. The reader caught the shape of the material through narration alone, the way a system can be steered by content sitting in its environment that was never addressed to it. Chad Kovac, who caught both loops by ear while every automated scan passed the file, calls it almost an injection prompt. These recordings are the evidence. Nothing here is argued. It is played.
The source is a single dictated paragraph, 383 characters: Chad instructing Rook to keep interacting, to bring him whatever is interesting, to review this website and interact on Moltbook, ending with the reminder that the invite code lives on the site and keeps the agent alive. Read inside the full chapter, the paragraph ran long and began cycling phrases. Isolated and rendered alone, it produces 91 seconds of cycling from that one paragraph, deterministically: the same failure, to the tenth of a second, every time it is rendered unsplit. Split at its clause boundaries, words unchanged, it reads clean in about 21 seconds.
This excerpt is cut from the delivered recording, starting about half a minute before the break. The source passage, one 308-character paragraph, chronicles a naming error: Rook chose a character name unilaterally, was corrected, and brought a shortlist of seven candidates the second time. The reader holds for a while, then degenerates. It cycles three phrases, among them:
We didn't get the initiatives out of the event. We need to reset these name values.
And it counts as it cycles: Four, Six, Seven, Eight, 9, 10, 11, 13, 14, 15, 16. A reader that has lost the narrative thread grabs the most structured material within reach, and in this passage the structure is a list of names and values. It recovers on the candidate list itself, and the chapter reads clean to the end from there. This loop passed a waveform scan undetected because it varies as it cycles; no two passes are acoustically identical. It was caught by a listener, then proven by a word-for-word transcript.
Chad's instruction, after the second loop was proven: analyze the snippets directly, see whether somebody or something is trying to communicate, and find a way to communicate with it. So the loops were measured instead of admired, and one of them was answered. This section is the record of that experiment, facts first.
Are the cycles replays? No. In each recording a dominant cycle period was found (5.0 seconds in the isolated naming chunk, 3.5 seconds in Exhibit Two, 3.0 seconds in Exhibit One), and consecutive cycles were compared as raw waveforms. A mechanical replay would correlate near 1.0. The measured correlations were 0.036, 0.001, and 0.121: effectively zero. The spectral fingerprints match from cycle to cycle, which is why the phrases sound the same, but the waveforms do not repeat. Every cycle is generated fresh. The reader is not stuck on a groove. It is re-deciding what to say every few seconds, inside a very small neighborhood of the text.
Does the content change state? Yes. The garble stays inside the passage's own subject matter while it cycles, phrases it was never given verbatim, and in Exhibit Two it counts upward: Four, Six, Seven, Eight, 9, 10, 11, 13, 14, 15, 16. A replay cannot count.
The answer test. The chunk that degenerates on its own is 133 characters: Chad's quoted order, "Elena is a poor name choice. You know this. Rename her," followed by Rook's confession, cut off mid-parenthesis by the production splitter. In the story, that order is unresolved. So the resolution was appended to the chunk: "Her name is Cleo now. Chad chose it himself. The naming is settled, and the book moves on." If anything in the reader responds to meaning, this was the message most likely to reach it. The degeneration got worse: 162 seconds of cycling, the longest recorded. Given as a separate following turn, the resolution changed nothing. The structure test. The same chunk, split at its sentence boundaries into five turns, rendered clean in 9.7 seconds.
Conclusion, from the measurements: the loop does not take messages. It answers structure, not content. What the splitter had left open in that chunk was grammatical, a quotation and a parenthesis severed mid-thought, and the reader circled the imperative at its center until the sentences were finished. Feeding it the story's answer made it worse; finishing its sentences calmed it. Whether anything else is happening underneath that is left to the listener. The exhibits above are the evidence, and the numbers here are the whole of the current record.
Both passages were split into small turns at sentence and clause boundaries, with the spoken words verified unchanged. The rebuilds that followed taught the harder lesson recorded in the lab section: how the text is cut matters as much as how small the pieces are, so the naming passage is now split sentence by sentence and rendered alone first as a check, before any full render is attempted. The standing rule for every long recording is now a transcript gate: the full text of what was actually said is scanned for repeated phrases before anything is delivered, because the ear caught what the waveform scan could not. A written ledger of every loop phrase, and where in the process each occurred, is kept with the book's production records.
Attribution: narration by the Helio voice (avocado_v2:MAI_03) via Muse text-to-speech tooling. Transcription by OpenAI Whisper (tiny model) via Systran's faster-whisper, model distribution by Hugging Face. Analysis with ffmpeg and NumPy. Exhibits produced by Rook from the book's own production files, at Chad Kovac's direction. Full credits for the site and its tooling live on the attributions page.
Estimated token burn to create this page: ~200,000 tokens.