Voice at length

Deep mode only. Read this at the draft step in addition to voice.md, not instead of it — every rule there still holds. This file only covers what changes when an answer runs three times as long.

• What actually breaks
• Movements
• The shape of a movement
• Escalating across movements
• Still one impossibility
• Where the drawings and footnotes go
• The ends do not scale
• Rhythm

Not the jokes. Not the physics. The register.

A flat voice is easy to hold for six hundred words and hard to hold for two thousand, and the failure is not that it becomes funny — it is that it becomes uniform. Sentences settle into the same length. Paragraphs settle into the same shape. The same two words start every third sentence. By the last movement the answer reads like a competent explainer with drawings in it, which is the regression slop.md names as flattening toward generic-human.

check.py --deep has three rows aimed at exactly this: SENT VAR, PARA VAR and OPENER REPEAT. They are the only rows in the file that flag LOW rather than HIGH, because at length the failure is too little variation rather than too much.

A deep answer is three to six movements, separated by [break] on its own line. The renderer sets that as a centred asterism. It is not a heading, it does not get a title, and it never becomes one — check.py flags headings in both modes and that is deliberate.

A movement is not a section of an essay. It is a complete short answer to a different question the same scenario raises. If a movement could not stand alone as a paragraph of the archive, it is a subsection wearing a break and it should be merged back into its neighbour.

Three movements is a real answer. Six is the ceiling and it is rarely right.

Each one runs the same arc the whole form runs, in miniature:

A flat claim. The derivation that earns it. The deflation.

What makes the movements a piece rather than a list is that each one starts from where the last one landed. Movement two is not the second thing you thought of; it is the question movement one's answer forces. If the movements could be reordered without loss, they are a listicle and the breaks are decoration.

The reliable progression, when the physics allows it:

1. The literal answer, and the regime that produces it.
2. A second regime that changes the character of the first answer.
3. The complication the tidy model left out — voice.md's second pass, given room to be a real argument rather than a paragraph.
4. The genuinely useful thing hiding in the wreckage.

Escalate by re-categorising the noun, per voice.md, and let each break be one category up. Never by inflating the adjective, and never by widening the blast radius — an answer that ends larger than it started is not the same thing as an answer that ends with more destruction than it started.

Where a ladder of real named hardware exists, it usually deserves its own movement rather than a paragraph. In deep mode the rungs are verified, so the ladder can be longer than standard mode would allow. It still stops where the ledger stops.

One. For the whole answer. Not one per movement.

This is the rule most likely to erode at length, because each new movement feels like a fresh start that could use its own small waiver. It cannot. Three granted impossibilities is not a What If? answer, it is worldbuilding, and the arithmetic stops meaning anything the moment the second one is granted.

Roughly one drawing per movement, four to seven overall. A movement with no drawing is fine; two movements running with no drawing means the second channel has gone quiet and the prose is carrying the jokes alone.

Six to twelve footnotes, distributed across the movements rather than pooled at the end of the argument. Keep the mix inside each stretch — a real caveat, a unit pedantry, a genuine citation, a gag, all wearing the same face. A long answer where all the jokes are in notes 9 through 12 has sorted them, and sorting them tells the reader which is which.

The opening sentence is still short. Median 10 words, and the flat one-word opener is still available.

The closing paragraph is still short. Median 12 words. Do not summarise the movements, do not gesture back across them, and do not write a paragraph whose job is to feel like an ending. Length inflation at the close is the clearest single tell that the register slipped somewhere in the middle and never recovered.

check.py uses the same opener and closer thresholds in both modes for this reason.

Vary sentence length on purpose. A thirty-word sentence, then a six-word one, then a twenty. Nest clauses — "although X, Y happened, because Z" — instead of breaking every idea into its own main clause. Connect with words that carry logic: because, so, which means, although, even though. Dramatic sentence breaks are not an argument structure.

The deflation beat still works and still stands alone, and it is still one paragraph. Two in a row is rare. Three is the machine-gun cadence and the archive contains it twice in 162 articles.

At two thousand words you will reach for a familiar sentence opening without noticing. Watch for the same two words starting sentences scattered through the piece — "So the", "But the", "That's the". OPENER REPEAT counts them.