The ninth article, and the two shorts that go with it

[peopleworks]

The one story today produced that is worth telling.

The article

EN — "My AI detector scored a passage 94 out of 100, and I made it refuse to say anything"
ES — "Mi detector de IA le puso 94 sobre 100 a un pasaje, y le prohibí decir nada"

It is built on the shape of the calibration corpus, not on the fix. Ninety texts, shortest 662
words, median 2,772, nothing below 600 — so every verdict about a pasted paragraph was extrapolation
onto a population the corpus does not contain. Last week's window study supplies the direction: 0 of
32 flagged whole, 6 of 32 as excerpts of the same writing, with the control row that rules out
composition and leaves length.

Then the part that makes it publishable: the fix was not a better model, it was to say less.

Three things it is careful not to say

Each is the sentence an editor cuts first, and each would hollow the piece out:

1. That this makes the tool more accurate. It makes it quieter.
2. That it fixes a 900-word essay. It does not; that is still far below the corpus median.
3. That this is a flaw in AI detection generally. It is ours, published by us, which is the only
   reason it is worth reading.

It closes on the version that travels outside the niche — a threshold is valid over the population
it was fitted on, and nowhere else — and on the honest note that publishing the method is what made
the defect findable, by me, a year later, reading my own numbers with a different question.

Verified against our own linter

	Score	Signals	Burstiness	Words
EN	0/100	0	0.59	1,588
ES	0/100	0	0.60	1,647

The English first pass scored 6/100 — three empty intensifiers and three mid-sentence bolds — and was
corrected before the HTML was generated, which is the same gesture as the sixth and eighth articles.

The cover is plotted, not drawn

social/length-story-cover.png: the ninety texts at their real word counts and real scores,
measured by running the shipping engine over each one. The floor, the threshold and the 66-word
passage are all in their true positions.

The empty region on the left is not a stylistic choice. The corpus has nothing there, and that
emptiness is the entire argument. No words to translate beyond the label, so it serves both editions.

Also here: the eighth article's platform copy

It had a briefing block and no posts — no X, no LinkedIn, in either language. It publishes
tonight, so that gap is filled in this PR: X and LinkedIn, EN and ES, keeping the negation in front
where the briefing asks for it.

Publishing order

Ninth after eighth, and linked from it. The pair tells one story — find the defect, fix it — and
each half is weaker alone. Worth putting through the committee first: it makes a claim about
measurement, and it is better that Fable and Copilot try to knock it down than a stranger.

🤖 Generated with Claude Code

https://claude.ai/code/session_015PEbbiYSNPw7jE3LrPNhyF

[peopleworks]

Added: shorts 12 and 13

Both languages, everything ready except the narration — that needs an ElevenLabs key, which is yours.

Short 13 — 94/100, y sin veredicto is the one to lead with. The 94 is grey, and that grey is the whole piece: a detector that scores a passage 94 and then declines to say anything is the opposite of what the category does. If a render ever comes out with that number in red or green, it is wrong.

Short 12 — the assistant's own turn, in a chat bubble at the foot of a submitted essay that visibly does not belong to the document. The sentence that cannot be cut: it does not say who writes well.

What changed in how they are built

The GUION says to narrate first, because delays are absolute seconds. These two put them in fractions of a --dur variable instead — which works because cue-times.mjs already assumes a roughly constant speaking rate and spreads the measured duration across the words. The relative position of every phrase is known without hearing anything; only the total is missing.

So: narrate, then set one number per file.

node narrate-all.mjs short12-residencia-es short12-residencia-en short13-longitud-es short13-longitud-en
# put each printed duration into --dur at the top of its HTML
node build-short.mjs scenes/short13-longitud-es.json

Three things only a snapshot caught

• the stamp lost its centring — its keyframe rewrote the whole transform and dropped the translateX(-50%) it was positioned with;
• the strike-through ran to the canvas edge, because its containing block was the row and not the sentence;
• the closing card had under three seconds on screen against the ~3.4 the earlier shorts give it.

I wrote a snap tool for this and then deleted it — the kit already has one, and two tools with the same name is how a repo grows a second source of truth. The GUION now points at the kit's, with a warning to pass a t larger than the total duration: at t=26 on a 28-second short the closing card has not entered yet, and it reads as missing rather than as late. That was my first mistake with it.

Scripts run 56–61 words → 24–27 seconds, back in the range of shorts 1–7 rather than the 29–40 of 8–11.