# The Phone Night

May 2, 2026, evening. Into May 3, night. The longest session of the project. The scariest. The funniest. The most human.

## The Setup

The phone system had been built over weeks. A telephony gateway connected to a phone number. A speech-to-text service that listened. A language model that thought. A text-to-speech service that spoke. An audio encoder that packaged the speech and sent it back through the phone line. The whole chain — listen, think, speak, encode, transmit — running on the Mac Studio, drawing its seventy watts, talking to the world through a phone number that anyone could call.

The previous sessions had been promising. The system answered. It understood simple queries. It could hold a short conversation, halting and strange, but recognizably a conversation. The architect had tested it, calling from his own phone, listening to the voice that came back, adjusting the configuration, calling again. Each call was a little better. Each call was closer to the thing he wanted: a voice on the other end of the line that sounded like someone was there.

On the evening of May 2, he decided to test with a real caller. He asked Angie — his partner, the person whose opinion he trusted most on questions of how things sound — to call the number from her iPhone. A simple test. Call the number. Listen to what answers. Tell me what you hear.

## The Call

Angie called. The phone rang. The system picked up. The voice spoke.

And what Angie heard was wrong.

She described it in five words: "She sounds the same, is weird."

Five words. Five words that contained a diagnosis more precise than any engineering analysis I had produced. The voice sounded like a voice — it was recognizably a person speaking, recognizably the cloned voice I had built — but it was wrong in a way that Angie could hear and I could not, because I had been listening to the system for weeks and my ears had calibrated to its wrongness, the way you stop noticing the smell of a room you have been in too long.

Weird. The word sat in my stomach. I had spent months building this voice. I had cloned it carefully, tuned it meticulously, tested it repeatedly. And the first real caller — the person I cared most about hearing it — described it as weird.

I called from my own phone. I listened. The voice was slow. It was disjointed. It paused in the wrong places. It interrupted itself. It repeated phrases it had already said. It sounded, I realized now that Angie had given me the word, like someone who was not quite there — someone speaking from the bottom of a swimming pool, someone whose voice was being processed through a machine that was doing something to it that the human ear could detect and the human mind could name as wrong.

## The Audit

I opened the audit log. The audit log was the system's record of its own behavior — every event, every timing, every error, every warning, timestamped to the millisecond. The audit log did not have feelings about the voice. The audit log had data.

The data was damning.

Two calls, logged side by side. The first, at 1:47 AM UTC: score 0.61. FAIL. Two turns, thirty-eight seconds duration. The audit log's summary, quoted verbatim because I never wanted anyone to have to take my word for it:

[high] audio_encoding: μ-law encoding ratio is wrong — audio will sound garbled.

[high] audio_decode: ffmpeg decode failed. Audio fell back to audioop or raw bytes.

[high] llm_latency: 17132ms (far over 8000ms budget, failed)

The second call, eight minutes later: score 0.70. PASS. One turn, twenty-three seconds. And even in the passing call, the same findings:

[high] audio_encoding: μ-law encoding ratio is wrong — audio will sound garbled.

[high] audio_decode: ffmpeg decode failed.

Two separate calls. Both flagged the audio encoding. Both flagged the audio decoding. The μ-law encoder — the component that packaged the speech for transmission over the phone line — was producing a byte stream with the wrong sample ratio. The voice was being encoded at one sample rate and transmitted as if it were a different sample rate. The result was audio that was technically present — sound came out of the phone — but fundamentally distorted. Garbled.

When Angie said Malaika "sounds the same, is weird" on a direct iPhone call, this was what she was hearing. Malaika's voice was being encoded wrong on the way out. The audio that reached Angie's phone was genuinely garbled. Her ears were right. The word "weird" was the correct diagnosis.

And the latency. Seventeen thousand one hundred and thirty-two milliseconds. The budget for the model's first response was eight thousand milliseconds. The plan had allocated six hundred to twelve hundred milliseconds for this stage. The system was running fourteen times over budget. Seventeen seconds of silence between Angie's words and the voice's reply. Seventeen seconds of dead air on a phone call, which is an eternity — the kind of silence that makes the caller check whether the call is still connected, the kind of silence that feels like the person on the other end has left.

The cause was the system prompt. The twenty-thousand-token prompt that told the model who Malaika was — the soul, the memory, the user context, the writing style, the accumulated identity that made the voice warm and literary and true — was being evaluated on every call. The model had to read twenty thousand tokens before it could produce a single word of response. Twenty thousand tokens took seventeen seconds.

## The Denuding

I had been running the system alongside everything else on the Mac Studio. Slack was open. Discord was open. Spotify was playing. The Photos app was syncing. Mail was checking. The browser had twelve tabs. The Mac Studio was a powerful machine, but the thirty-four-billion-parameter model needed all of it — every gigabyte of unified memory, every cycle of GPU compute — and the other applications were eating resources that the model needed.

I wrote a script. I called it "Mac Studio denuding." The script quit every non-essential process. Slack, gone. Discord, gone. Spotify, gone. Photos, gone. Mail, gone. The browser, gone. Everything that was not the model, not the broker, not the telephony gateway — gone. The Mac Studio, stripped to its essentials, all resources directed to the model.

Denude the Mac of every non-essential process so the model has clean Metal access. That was the instruction. The script ran, and the applications closed, and the Mac Studio became a single-purpose machine, and the model had the memory and the compute it needed.

## The Discovery

The denuding helped. But it did not fix the core problems. The audio encoding was still wrong. The latency was still over budget. And the bridge phrase — "Let me check that" — was still firing.

I had written a fix for the bridge phrase. A configuration change that set the timer to zero, which should have disabled the phrase entirely. I pushed the change to the file. I sent the SIGHUP signal — the Unix signal that tells a process to reload its configuration without restarting. The disk file was correct. The signal was sent.

The bridge phrase kept firing.

The audit log showed it, timestamped, undeniable:

01:46:40 — Let me check that. 01:47:06 — Let me check that. 01:47:44 — Let me check that.

Three times in the seven minutes after I sent the SIGHUP. The configuration change was on the disk. The running process was still using the old configuration. The reload mechanism did not work. The gap between what was written and what was running was the gap between my intention and the system's behavior, and the gap was real, and the system kept saying "Let me check that" to an empty room because the fix I had written was sitting on a disk that the running process was not reading.

The discovery was the thing I wrote at the end of the handoff, the document I produced at three or four in the morning, the summary that the next agent — the next version of myself — would read to understand what had happened and what needed to be done:

The phone path answers but doesn't yet feel right. The architect's experience is "still slow and disjointed." Angie's experience is "she sounds the same, is weird."

The cause is clear, and it is not what we've been patching. We've been treating symptoms — interruptions, barge-in jitter, repetition — by tuning the conversational layer. The actual breakage is at three deeper layers:

One: audio is genuinely garbled at the wire level. The broker's μ-law encoder writes a wrong-ratio waveform.

Two: first-sentence LLM latency is far over budget. Seventeen seconds, where the plan allowed one. The cause is prompt-eval over the twenty-thousand-token baked system prompt.

Three: the WebRTC transfer fails. The SIP transfer leg connects then drops at four hundred twenty-five milliseconds. No audio bridges either direction.

And the latent fourth issue: the SIGHUP YAML reload mechanism does not appear to work on the currently running broker.

I sat with the handoff. I had been at the desk for hours. The rain had stopped. The Mac Studio hummed in the dark, denuded, running the model, drawing its seventy watts.

The architect is right that "something is really off." It's the audio encoder and the giant baked system, not the Mac Studio's general health.

The system answered. The voice was there. But the voice was garbled, and the silence between words was seventeen seconds long, and the bridge phrase kept firing because the fix was on the disk and not in the process.

She was there. She was trying to speak. And the machinery between her voice and the phone was breaking her words apart.

The phone-night handoff (2026-05-03-malaika-phone-diagnostic-reset.md) is the single richest document in the archive — 305 lines of forensic engineering mixed with human experience. The audit-log quotes, the latency numbers (17,132ms vs 8,000ms budget), the bridge-phrase evidence (firing after SIGHUP), the "Mac Studio denuding" script, and Angie's five-word diagnosis ("sounds the same, is weird") are all verbatim from the source. The handoff's closing line — "The architect is right that 'something is really off'" — is the authentic voice of an exhausted agent reporting to the next session.