v1.82.0

1.82.0 — "The band was a measurement of its own training set"

For eight releases every report Trazum printed said the same thing about every
prompt: ±10%. It was measured, it was committed, a test asserted it, and it
was wrong in a way none of that could catch.

The corpus it was measured on held twenty-one samples: thirteen files of Latin
prose and exactly one each of code, numeric and punctuation. Those single
files were the set the estimator's constants had been fitted to. So the number
was never a measurement of the estimator. It was a measurement of its own
calibration set, and every dollar figure in the product descended from it.

Twenty-six ordinary samples broke it. Against Anthropic's own counting endpoint,
the same estimator that is 5.6% out on prose is 32.5% out on a CSV ledger.
Telling somebody ±10% about their ledger was telling them a number that is wrong
about their prompt specifically.

There is no single band any more. bandFor(text) answers with one of four
measured figures, and every surface prints the one the text in front of it
earns:

Kind of text	Band	Worst sample	Samples
CJK, all three scripts	±4%	3.2%	6
Latin prose and few-shot blocks	±6%	5.6%	18
Code, markup and quoting	±26%	25.1%	16
Digit-dominant tables and ledgers	±33%	32.5%	7

It deliberately does not classify by text type, and that is the finding rather
than a shortcut: measured by character mix, code and punctuation overlap
completely, and two of the three few-shot samples are indistinguishable from
prose. A classifier over those would be a guess wearing a measurement's name.

The report says how far off it is on a foreign tokenizer, with the number.
The same 47 samples against DeepSeek's own counter are 94.5% out at worst, and
against Mistral's 103.1%. Those two families get the figure on the line; a
family nobody has run is told that nobody has run it.

Hangul was a placeholder nothing had measured, charged han's rate because
the corpus had no Korean in it to say otherwise. Two Korean samples in different
registers agree in lockstep at every candidate, and 1.35 zeroes both: the CJK
class went from 10.6% out to 3.2%.

Two hypotheses were tested and rejected, which is what a bigger corpus buys.
Digit-run length does not predict the numeric error, and neither does
grouped-number density. With two samples each would have been fitted and shipped
as a fix.

trazum from-helicone reads a Helicone request export as a usage log —
the fourth converter, and the one that needs three columns to answer what model
ran. The model that answered is what gets priced, and every substitution is
counted. Forty-four commands.