Not scheduled. Listed so the reasoning is on the record.
Nothing here is built, apart from the entries struck through or annotated as
shipped, which stay as the record of what this file went back on. That is the
whole meaning of the heading, and it stopped being true without anybody
noticing: the editor extension shipped as trazum-vscode in 1.86.0 and the per-model-family tokenizer as
@trazum/tokenizer-openai in 1.85.0, while both stayed listed below as things
this file was still thinking about. The release sections in
HISTORY.md recorded them as
released the whole time, so one file said both at once. roadmap-truth.test.js
holds it now, because the moment something is built is exactly the moment
nobody remembers to move it.
More contradiction axes — tone, persona, refusal policy. Unscheduled rather than planned, because each new axis has to earn its place against false positives and I have not seen enough real prompts to know which ones do. An advisory people learn to ignore is worse than no advisory, and four axes that fire correctly beat seven that mostly do not.
More locales. The architecture supports it as of 0.3.0, and adding one is now a catalogue plus dictionary entries. Held back on purpose: a language needs a maintainer who actually reads it, and a stale translation is worse than an honest fallback to English.
What that role is was written down at 1.56.2, with its bar, in docs/language-maintainer.md — along with the admission that five of the seven dictionaries already shipping had never been read by a speaker either. The page does not move this entry into the schedule. Whether somebody fills the role is not something this project can plan.
Japanese is deliberately not on this list, and the split is the point. There is no Japanese trimming dictionary and one is not planned: deciding that a phrase says something in more words than it needs is a judgement about the language, and nobody here can make it in Japanese. But
--reorder's backward-reference list does cover Japanese and Chinese, matched without word boundaries because 上記のテキスト has none. Those two are not the same claim. Refusing to rearrange somebody's prompt needs only enough of the language to recognise a phrase pointing backwards; offering to shorten it means asserting the shorter version still asks for the same thing. Trazum will do the first in a language it cannot read and will not do the second.So a Japanese prompt gets
--reorder's protections in full, no trimming, and a report that names the seven languages the dictionaries cover rather than implying the prompt was already efficient.Still unscheduled, and the number that was missing now exists for two families out of six. Pricing seven providers made this the question it always was in miniature: the estimator is tuned against Claude's tokenizer, and a GPT or Kimi figure carries whatever error that mismatch produces. This entry set the threshold itself — within 5% across families and the dependency is not worth taking; 40% out and it is — and said deciding without that number would be deciding blind.
1.82.0 measured it. Against DeepSeek's own counter the estimator is 94.5% out at worst over the 47-sample corpus, and against Mistral's 103.1%. Both are more than twice the threshold this entry set for worth taking, so for those two families the question is answered and the answer is yes.
It stays unscheduled anyway, and the reason is not the measurement. A real tokenizer per family is a dependency in packages that have none, and that is a distribution decision rather than a measurement one — the same shape as the editor extension above. What has changed is that it can no longer be deferred for want of evidence: the report now names the family and prints the measured error where one exists, and says nobody has measured it where one does not. Google, xAI and Moonshot are the three still unmeasured, each needs a key for its own counter, and each has its own named skip in the suite carrying the command to run. OpenAI left that list in 1.85.0, not because a key arrived but because its measurement was already committed here while this file and
band.tsboth said otherwise.The dependency question is settled, and the answer was a separate package.
@trazum/tokenizer-openaiships in 1.85.0: the core still declares nothing, and somebody who wants exact OpenAI counts installs twenty-two megabytes on purpose.Prompt library.Shipped in 1.7.0, and the reasoning here was wrong in an instructive way. This entry said storing prompts "would mean sending them to a server. Trazum's privacy story is that it never does" — and it read as a principle when it was really a conflation of two things. Rule 1 binds the optimiser: the CLI sends nothing, needs no account, and works with the network unplugged, and none of that changed. It never said nobody may run a service that stores what they chose to save to it. Left visible rather than deleted, because a roadmap that quietly removes the entries it went back on is a roadmap with no record of having gone back on anything.Cost alerting. A bot that watches a prompt's monthly estimate and speaks up when it crosses a threshold, or when caching effectiveness drops. Attractive, and unscheduled for the same reason as the two above: it needs somewhere to run and something to remember — a service holding other teams' prompt metrics on a schedule.
check --max-tokensin CI already covers the threshold case for anyone content to have the answer arrive on a pull request instead of in Slack, which is most of the value for none of the surface area.