what is in each release?
Versions are what a reader can install and point at. A number in pyproject.toml with no tag behind it does not answer "which one am I running", which is the first question anybody asks a library that is still moving.
0.1.0, 5 August 2026
The first tagged build. Everything below runs on your own machine against your own key, and nothing here bills through anybody.
the measurement
Wilson score intervals rather than the normal approximation, so 0 of 20 is an upper bound instead of certainty of absence.
A clustered percentile bootstrap that resamples prompts, because repeats inside one question are not independent draws of the thing being measured.
Interval width split into the model answering differently and the question set chosen, since only one of those two is worth spending more money on.
No rank printed where the ordering does not repeat, and no "ahead" where the intervals overlap.
A round is one engine and one surface. A file holding two is refused and the engines it holds are named.
The screening gate counts rivals and never the customer, because filtering a question set by the customer's own score raises the published number by deleting the evidence against it.
the record
An append-only, hash-chained ledger. Every report re-derives the chain before it computes anything, and lulu verify re-derives every chain on disk.
Failed calls are written down with status="error" rather than retried, so a round is what happened rather than what happened often enough.
A ceiling checked before each call. A round that hits it stops short, says so, and exits with its own code rather than pretending to be complete.
the commands
Fifteen, from lulu setup to lulu publish. Prices come from the provider's own published page with the date they were read, and lulu usage keeps an amount the provider stated, a cost computed from reported tokens, and a floor for calls it said nothing about as three separate bases that never merge.
the surfaces
A static site with four routes, every figure on it off a measured round, and a social card per route generated from that round's own numbers.
lulu publish writes a round out as pages, by hand, so nothing reaches the web that somebody did not choose to put there.
the gates
990 python tests and 52 node tests, all offline, no key spent.
ruff over 78 python files, eslint and tsc over the site.
Every push runs all of it on python 3.11 and 3.13, installs the wheel into an empty environment and runs lulu --help out of it, and the site deploys only behind those gates.
known limits
Written out in the README under "what does it not do?", and held to the code by a test: two engines rather than four, a rival list still typed in by hand, and a name counter that drops a name nothing in a round spells.