A patch, and every entry in it is the same act. With the plan finished as far as it can go, the work is taking a claim this project makes about itself and asking what enforces it — then measuring by running rather than by reading. Four findings, one of which reached installable code.
Changed#
And the other rule that constrains everything was enforced by a disclosure
check, not by a run. The deterministic core stays free and offline — no
feature may make a network call a prerequisite for optimising a prompt is the
first of the two rules ROADMAP.md opens with, and the thing a reader is most
likely checking when they open SUPPORT.md.
Two things enforced it, and neither ran the command. outbound-surfaces.test.js
derives every module that can reach the network and requires each to be named
in the prose — a good rule and a different one, because a module can be
disclosed, listed and documented and still sit on the path of trazum optimize.
And security.test.js scans packages/core/src only for fetch, allowing
it in two named modules; a source scan, over one package, of a property that is
about what a command does at runtime — the CLI is not in it at all.
It is now proved the only way it honestly can be — by removing the network and
running the command. fetch is replaced with a thrower before the CLI loads,
and optimize, check and rules have to work with it gone; the report has to
come back byte-identical to a run with the network intact.
And the stub is proved to bite, which is what makes the rest mean anything:
--pricing-live under the same stub has to fail and carry the stub's own
marker, because a stub that silently failed to install would leave every
assertion above passing and nothing proved.
Measured before it was written down: it holds, in the library and through the binary. The guard is the part that was missing.
One of the two rules that constrain everything here was checked on one
English sentence. A locale changes the report, never the optimisation is one
of the two rules ROADMAP.md opens with, and the README says it is enforced by
tests. It was — by a single prompt in two locales.
A rule stated universally and checked on one example is a claim that happens to be true, which is the shape this project named at 1.59.0: a rule you wrote for yourself is a claim like any other.
The sweep now reads the corpus off disk — 35 prompts across the language
corpus, the rule corpus and examples/ — and the locales off LOCALES, and runs
every prompt at both levels in every locale. What "the same" means is enumerated
rather than sampled: the optimised text, tokensBefore, tokensAfter,
tokensSaved, reductionPct, every rule id with its hits and saving, and every
advisory id.
And the opposite direction is asserted too, because a build that returned the English report for every locale would satisfy every equality perfectly: somewhere in the corpus a rule fires, and its title has to come back different.
Measured before it was written down: the invariance holds on all 70 pairs. The guard is the part that was missing, not the property.
Fixed#
The profile was the one contract of the ten whose schemaVersion existed only
if you went through the CLI. docs/format.md promises that every document
carries it and conform rejects a document without it — so profileUsage(), the
function in the package whose whole job is emitting this format, returned a
document trazum conform refuses.
Nobody could have noticed from inside. Every test that checked a profile
against the contract added schemaVersion first, because the CLI does, and a
fixture built the way the CLI builds it can never catch the CLI doing the work. A
connector author reading docs/format.md and using @trazum/core would have
found out from a rejection.
It is stamped by profileUsage now, where every other contract's builder stamps
its own — the plan, the annual record, the roll-up, the cost answer. The CLI's
own stamp is gone, and the compiler caught it as redundant, which is the change
proving itself.
The guard checks the class, not the instance: every document this package can
build with nothing but its own exports is handed straight to its own checker.
Proved by removing the field and watching conform reject it, and the eight
contracts the test cannot reach from the package alone are named, because two
documents checking out is not the format checking out.
Same shape as #290 on this project's record, where outcome-report was a
contract whose only implementation failed it for nine releases.