An agent derives a claim again from the code before acting on it, and in tsv runs the canonical parser before building on a guess about syntax.
Leaves out
A session leaves no record of having checked, so the hold is as good as the session.
Fails
When a claim that nothing uses a function is accepted from a search of one package, without tracing every consumer.
Evidence
The grimoire skill, which asks for a claim to be derived again before it is acted on
The tsv fixture workflow's parser verification, which runs the canonical parser before any implementation
Written: Leads, not facts
Coverage
Throughout
Does
The written rules treat a review finding, a ledger entry, a memory and an agent's own summary as leads to verify, and call the code the ground truth where it disagrees with a rule.
The fuz-stack skill, whose opening says the code is ground truth
Tool: Figures computed from reports
Coverage
In places
Does
The benchmark figures that tsv and fuz_schema publish are computed from committed reports that record their commit and machine, and a test fails when the prose stops matching the data. fuz_schema also tests the comparisons and recipes in its README, and tsv holds its catalog of differences to its fixtures.
Leaves out
These are single documents, each with a test of its own. Most sentences in the docs have no check.
Fails
When the benchmark behind a figure predates the current build: refreshing the reports is manual, and nothing checks their age.
Evidence
The tsv benchmarks, whose prose figures are computed from the reports
The tsv site's benchmarks notes, on the test that gates the prose
fuz_schema's measurements, each held to the published results by a test
A stale statement is trusted
A statement that was true when written is still quoted long after.
Session: Distill, then delete
Coverage
In places
Does
Agents condense raw notes into durable documents and delete what is resolved or superseded.
Leaves out
A pass covers what it was pointed at, and everything else waits for the next one.
Fails
When documents only grow, and a reader can no longer tell the current statement from the one it replaced.
Design docs in gro and fuz_forge mark what is built and what is only designed, and the written rules treat a wrong doc comment as worse than a missing one.
The Rust workspaces but zervo fail on a broken doc link, tsv fails on any dead relative link or anchor in the Markdown of its repo, and each TypeScript gate fails when a generated file has drifted. A task on this site resolves the links on this page against the repos they cite, when someone runs it.
Leaves out
Outside tsv, no gate checks the links in a repo's own Markdown files, such as its README and docs. Only Markdown that a site renders is crawled, by that site's build.
Fails
When the links resolve and the sentences around them are no longer true.
Evidence
tsv's conformance audit, whose second check is dead links in every Markdown file
An optimization is predicted to help, lands, and does not.
Session: Race the candidates
Coverage
In places
Does
In fuz_schema, tsv and gro, a candidate change is graded against its base on real workloads before it lands, and fuz_schema and pglet keep a record of what was measured and declined.
Leaves out
The other projects do neither.
Fails
When a micro-benchmark or a model of the cost stands in for the measured workload.
Evidence
fuz_schema's record of trades that were built, measured and declined
fuz_schema's grading of a candidate against a base
Committed limits hold the WebAssembly builds of tsv, pglet, fuz_schema and blake3, the native addon of tsv, and the binary and startup snapshot of gro.
Leaves out
Size is the only axis a gate limits on each change. Speed is graded against a base on demand: tsv fails a row that moves past a committed percentage, and gro asks for a written reason.
Fails
When the gated axis holds and the regression is on another: an instruction count is blind to code size.