The remaster is ready. Does it sound right to you?

The opening, and the fault the audit exists to catch

This is the part that is genuinely new since you last looked.

Which opening, and how that was settled rather than guessed

A replacement episode has to open the way the published one opens. That is a fact to read off the published file, not a choice to make again. So the head of the published episode was compared against every opening in the library, and the answer was not close: the Sting matched at 1.00 out of a possible 1.00, and the other two scored below zero, meaning they do not resemble it at all. The bar for accepting a match without asking you is 0.80. Anything under that and the pipeline stops and hands the choice to you, because a wrong opening is exactly the kind of mistake that ships quietly.

The score is a correlation: 1.00 means the two shapes rise and fall together perfectly, 0.00 means no relationship, and a negative number means they move against each other. It compares the loudness shape over time rather than the raw waveform, so a difference in level or in mp3 encoding between the two files cannot throw it off.

The opening is now audited before it is allowed anywhere near an episode

This run it passed. Its last fraction of a second measures -54.8 dBFS — quiet, but real sound, a room rather than nothing.

That check is not routine paranoia. The sting in the old library ends in 522 milliseconds of digital zero: not a quiet room, literally no signal at all. The voiced opener had that trimmed off years ago; the sting never was, and it rode into every episode built from it. You can hear the consequence in the file that is live today, and you can hear that it is gone in the new one, two paragraphs down. The pipeline now reads from a repaired copy of the opener library, so this episode needed no patching — the fault was fixed at the source rather than papered over once per episode. The audit stays in place because an asset used eighty times carries its defect eighty times, and the cheapest place to catch that is before it is used, not after.

What this run did repair: the story's own first 27 milliseconds

The raw recording begins with a sliver of digital silence before the first sound — 27 milliseconds of true zero, sitting inside a 55 millisecond stretch that reads as silence to the detector. Completely harmless for as long as it is the first thing in the file, because nobody can hear silence before the audio starts. The moment an opening goes in front of it, that same silence lands in the middle of the episode, between the sting and the first word, and becomes audible. It was filled with room tone taken from the quietest half-second of this same episode, measured at -58.8 dBFS. Room tone is the specific hiss of one room, one microphone and one evening, so a stock hiss underneath a specific recording would be heard immediately as a join.

The result, and the bit you can actually hear

In the published file the opening's silent tail and the story's silent head sit back to back, and together they make one continuous 680 millisecond hole starting at 6.50 seconds — a square-edged one, where the sound stops dead instead of fading. In the new master that region measures zero holes and zero milliseconds of digital zero. So does every other second of the file.

The pale blue block is the opening. The red block on the top row is the 680 millisecond hole in the file that is live today. The green sliver on the bottom row is the filled head of the story in the new master. There is nothing else marked on the bottom row because there is nothing else to mark.

Listen to the handover from the opening into the first words, in both files. Each button plays about six seconds.

On a phone speaker in a noisy room this may be hard to catch. On headphones it is the moment where the room disappears completely for two thirds of a second and then snaps back as the first word arrives.

The opening was kept out of the volume setting

The opening arrives already mastered to the level it was designed at. If it were included when the episode's overall volume is set, every correction would quietly drag the opening away from that level. So the levelling stage works on the story only, and the final audit then measures the whole episode, opening included, because that is the file people get.

Listen to the whole thing

Three versions of the same episode. Play any of them straight through.

The 28 loud moments, one tap each

These are the spots where a sudden peak, mostly one of the children shouting or laughing, was held down so it would not jolt a child who is already drifting off. Nothing was ever pushed up, only held down, and never by more than 6 dB. Together they cover 49.5 seconds, which is 9.0 percent of the story.

The pale blue block at the left edge is the opening, the first 7.03 seconds. No moment falls inside it.

Each button plays only that moment, with 2 seconds of run-up so you hear what leads into it, and stops half a second after it ends. The third button plays the new version and then the old one back to back, which is the fastest way to tell whether you can hear the difference at all.

The times shown are positions in the new master, so they include the 7.03 seconds of opening at the front. The pipeline's own record lists these moments against the raw story instead, which is why every time here is 7.03 seconds later than the number in the record further down the page. When you press "hear it before", the old file is seeked further on again, because its copy of the story sits later inside it — usually by 0.08 of a second, but not everywhere, so the seek for each of these 28 moments was measured separately rather than assumed. The five moments where the two files are further apart than usual say so above their buttons.

"Up to" is literal. The volume slides down over 0.6 seconds, sits at the stated depth across the middle of the moment, then slides back up over 0.6 seconds. Only the middle is held by the full amount, which is exactly why the change itself does not register as an event.

What actually changed

In plain words, comparing the three files.

The original recording arrived clean. The audit of the raw phone recording found exactly one flaw in nine minutes: a 55 millisecond gap of silence right at the top, before the recording properly starts — the one the head fill above dealt with. Nothing else. No dropouts, no clicks, no repeated fragments. That is why the repair stage had nothing to do and did nothing. A second, entirely separate detector built by other people, Essentia 2.1b6.dev1389 (standard mode), was run over the same recording as a check on our own detector, and it reached the same conclusion: independent detector agrees the recording is clean. It found zero clicks and zero dropouts of its own. It does not see our gap at the top, which is expected, because that one sits before the recording starts rather than inside it. The pipeline treats that second opinion as advice, never as a gate.

The published version has real damage in it. Running the same detector over the Auphonic file that is live today turns up 396 holes of digital silence where the room tone vanishes completely, 9.0 seconds of pure zero in total. 74 of those have square edges, meaning the sound stops dead rather than fading, which is the kind you can actually hear as a small click or a hiccup. It also has 6 duplicated fragments, short pieces of audio repeated — and every one of them repeats at the same 0.43 second distance, which is the signature of a misaligned paste rather than of somebody genuinely saying something twice. Plus 3 further micro gaps. Pointing the independent Essentia click detector at it turns up 5 click hits, but they cluster at only two distinct places, 0:32.4 and 7:04.0 — the rest are the same instant reported more than once. The same detector finds 0 in the new master and 0 in the raw original.

The figures for the raw original come from the pipeline's own record. The ones for the new master and for the published version were measured while this page was being built, by running the pipeline's own detector and the Essentia detector over both files, because the pipeline only ever compares delivery checks against the published file and does not audit it for damage. One caveat the pipeline itself records: a raw hole count overstates audible damage, because most holes are gated near-silence sitting inside a pause. The number that matters is the square-edged subset, 74.

The life in the recording survived. Loudness range, a measure of how much the volume moves between the quiet parts and the loud parts, measured in LU or loudness units, is the number that tells you whether a story still breathes or has been ironed flat. The raw recording measures 12.2. The new master delivers 10.8. The published Auphonic version was squashed to 6.8, close to half the movement. That difference is the reason to replace it.

Volume is set for Apple Podcasts. The new master delivers −16.5 LUFS as heard. LUFS is the standard unit for how loud something actually feels to a listener rather than how big the waveform is, and Apple asks for −16.0 give or take 1, so this sits inside the window. "As heard" matters here: the file is mono, and a mono file played through two speakers sounds about 3 dB louder than the meter reading for one channel, so the pipeline measures it the way it will be heard rather than the way it is stored. The file-only reading is −19.6.

Nothing is near the ceiling. True peak is −2.7 dBTP. dBTP means decibels true peak, that is, the highest level the sound reaches once it is turned back into an actual waveform in the listener's player, including the little overshoots that mp3 encoding creates but a plain waveform view never shows. The ceiling for this pipeline is −1.0, and −2.7 leaves comfortable room. Zero samples hit full scale.

The old file's timeline does not run straight. Lining the two files up against each other two seconds at a time, the published one normally carries the story 0.08 of a second behind the new one, steadily, for about 84 percent of the episode. In three stretches it does not. In each one it slips further behind, holds there for around half a minute, and then snaps back:

The biggest of them covers almost exactly the span where all six duplicated fragments were found, between 1:06.4 and 1:34.0 of that file, and its size is close to their shared repeat distance of 0.43 seconds. Read together that is not a mystery: audio was pasted into that stretch of the published file, pushing everything after it late, and taken back out at the end of it. It is the same class of fault as the holes, seen from a different angle. The new master does not do this — checked at 54 points spread across the episode it tracks the raw recording at a flat 7.03 seconds throughout, never out by more than 0.167 of a millisecond.

Measured while this page was built, by sliding an eight second window along both files and finding, at each step, the shift that makes their loudness shapes line up best. Inside those stretches the match at the usual 80 ms collapses to almost nothing while the match at the value shown stays above 0.9, so these are real displacements in the file, not the measurement wandering.