Every tool agreed,
and six of them were wrong
One score, eight cues, and not a single error message. The files were wrong anyway.
Every tool agreed,
01 — One score, eight cues
A show needs an opening, something to talk over, a transition and a close. All four came out of the same twenty-six bars.
The piece is a walking bass under a classical guitar line, in E Dorian at 132 beats per minute. The bass plods. The guitar finds notes, keeps some and abandons others, which was the idea: a person working at a fixed pace while something faster beside them produces material that sometimes sticks.
A show does not use a fifty-second piece of music. It uses a six-second sting, a loop quiet enough to sit under speech, a two-second button between segments, and a close that fades out. Each is a range of bars from the score, so cutting them became a tool rather than eight careful sessions of clicking.
Who wrote it is worth stating plainly, because the rest of this is about software being wrong. James Hartt wrote the theme. He is one of the two people building AgentSee, not someone we commissioned. Every note was placed by hand, and no model composed, suggested or arranged any part of it. The help was on theory: a conversation about modes and chord function, and a set of notes from another member of his quartet. Neither of those placed a note.
The tooling came afterwards and only ever did the cutting. That division turned out to be the whole argument. Writing fifty seconds of music is the part nobody wants taken away. Turning it into eight assets is the same careful afternoon repeated eight times, and that is the part worth giving away.
02 — A file can be valid and wrong
Sheet music written as XML states the important things once, in the first bar, and assumes them ever after.
The key signature. The time signature. How many ticks make a beat, which clef to read, whether the instrument sounds where it is written. All of it appears in bar one, and nowhere else.
So cut bars nineteen to twenty-six into their own file and you get a document that opens without complaint and passes every check. It also plays at the wrong speed, in the wrong key, an octave out. Nothing warns you, because nothing is malformed.
The tempo was worse. It is written into the first instrument's part only.
Our quiet loop is bass alone. Extracted on its own it carried no tempo at all — even cut from bar one — and would have rendered at whatever the software happened to default to.
A loop running at a speed nothing else in the show agrees with is the sort of fault that survives a listen and then irritates an audience for a year.
Nothing was malformed. Everything was wrong.
03 — The loop length could not be calculated
Two bars at 132 beats per minute is a number. It was not the right one.
A loop has to be exactly one musical length or it drifts. So the obvious move is arithmetic: four beats a bar, 132 a minute, 48,000 samples a second.
The figure that comes out is not the figure in the file. The render ends with the last note still decaying, and that tail belongs to no cycle.
Dividing the file's length by the eight cycles it contains looks like the fix. It is 6 per cent high, for the same reason.
The first version of our measuring tool did exactly that. Tested against a recording whose tempo we already knew, it was out by eight beats per minute — and stated that to two decimal places. Correlating each cycle against the next gave 174,545 samples and 132.00, which had been the answer from the start.
One finding there was worth the detour. A loop taken from the first or last cycle of a recording can never be seamless. The first has nothing ringing into it; the last has no following cycle to ring into.
Both ends are discontinuous by construction. Measured against the middle cycles, those two were about seventy decibels worse. We had been choosing between them by ear, and guessing.
04 — The meter was wrong about its own work
Loudness software reported four of eight cues as badly off target. All eight were on target.
Music under speech has to sit well below it. Set that gap by ear and it changes every week, so ours is written down: cues that play alone at one level, the loop fourteen decibels under.
The measuring pass then reported that four of the cues had missed by up to three decibels.
They had not. The figure being printed was the software's own forecast of what it was about to do, and that forecast runs pessimistic. One cue predicted at minus nineteen measured minus sixteen point four once written.
Measuring the finished files instead of reading the forecast put the whole set inside four tenths of a decibel.
The same tool, asked to leave the sample rate alone, produced files at 192,000 samples a second. It upsamples internally to find peaks. With no rate specified on the output, that internal rate lands in the file.
So the instruction to change nothing had to be written as an instruction to change it to what it already was.
05 — Two things the notation cannot do
Some of what we wanted was never a software problem.
The close is meant to fade to nothing. You can write that: there is a mark for dying away and a hairpin which closes into a small circle meaning silence. Neither does anything to a plucked note, because its level is fixed the instant it is plucked and nothing rides it down afterwards. A fade to silence is an operation on a recording, not an instruction in a score. Every fade-out anyone has heard was done at a desk.
We also panned the bass slightly left for width and got none. The ostinato sits between roughly 49 and 98 hertz, and fundamentals that low carry almost no directional information, so the attack of the pluck moves and the body of the note stays put. It costs headroom and gives nothing back. Most podcast listening happens on one earbud, which makes any instrument panned hard enough to notice an instrument that half the audience loses completely.
06 — The reason this is published
Six tools were confidently wrong and not one of them raised an error.
That is the part worth keeping. Valid files, plausible numbers, no warnings, wrong output. Every fault above was caught by someone disbelieving a figure and going to look, and the looking is the expensive part. Generating eight cues from one score took an afternoon. Establishing that the eight cues were correct took considerably longer, and could not be delegated to the thing which produced them.
We expect that ratio to hold for most of what this company does. It is the reason we keep a record rather than a portfolio.