ACTIVEagent

claims-auditor

Checks whether the site is still telling the truth about itself, across every surface that repeats a claim.

Model
opus
Tools
ReadGrepGlobBash(rg *)Bash(git *)Bash(node *)Bash(pnpm *)Bash(curl *)
Project
zckr-dev
Cost
One Opus pass, read-only, mostly ripgrep.

WHEN IT RUNS

Verifies that facts the site publishes about itself are still true, across every surface that repeats them. Use after behaviour changes, before shipping anything user-visible, or when a published count, name, or list may have drifted.

CONTEXT

The claim under question, its source of truth in code or data, and the seven surfaces that repeat claims here: the page, the markdown mirror, the architecture doc, the design doc, the project records, the README, and the machine surfaces.

MEMORY

None. It reports claim, source, every surface stating it, agree or drift, and whether a test guards it.

EVALUATION

Mismatches come back with file:line. It also names claims that are right but unguarded, because an unguarded claim is a liability rather than a pass.

WHERE IT FAILS · 3

  • Checks the number and misses the justification: a comment citing a rule that has since been rewritten is a wrong claim even when the figure beside it is correct.
  • Greps the value and not the phrasing, so a claim spelled out in words rather than digits survives the sweep.
  • Cannot see claims that live only in prose nobody thought to make testable.

LESSONS · 1

  • Wrong prose compiles perfectly, which is why this is the one bug class every automated gate in the repo is blind to.

CONNECTED · 2

SEE THE GRAPH

← ALL AGENTS