three projects · two MCP servers · one measurement lab
Each of these answers a question about something it did not write — a statute book, a stranger's codebase, a benchmark — and each one reports what it cannot answer as part of the answer. Every page below carries real captured output, with the revision and date it came from.
MCP server · Pakistani statute law, as of a date
Answers a statutory question as of a date and cites the provision it answered from. A provision amended in 2022 has one text before that date and another after, so an answer with no date attached is wrong for one of them.
corpus 30,174 provisions from 970 Acts, 1839 to 2026 · over 105 questions: 60 answered correctly, 0 wrong citations, 34 of 34 correctly refused
See a real run → githubMCP server · six readers for a codebase you did not write
Reads an unfamiliar Python repository from its AST: what depends on what, what a change would reach, what the codebase is eligible to modernise. It writes nothing and imports nothing from the tree it is reading.
resolver measured on twelve projects nobody here wrote, 1,063,636 lines at pinned revisions: 61.5% median, 38.0% worst — against 93.7% on its own repository
See a real run → github11 command-line packages · 10 measurement apps · 20 products
A lab for local coder models: ten apps that each answer one question that changes what you would build, eleven zero-dependency developer tools, and twenty multi-agent products on a shared platform.
MBPP accepts 731 of 4,055 single-point mutants of its own reference solutions (18.0%); 362 are provably different programs the benchmark marked correct
See the measurements → github (code private)Each one is built so that the thing it cannot do is visible in its output rather
than discovered by the person using it. Pakistan Law refuses when no provision
contains a word the question turns on, and names the word. Repo Recon reports the
share of calls it could not bind, and publishes the figure measured on other
people's code rather than its own. The measurement lab reports a mutant as
unproven unless an input is found on which it and the original
disagree.
None of the three runs on this host. Two read a corpus or a repository from local disk; the third needs a GPU, Kafka and Redis. These pages are static and carry captured output, labelled with the revision and date — not a live service pretending to be one.