Synthetic evidence, recorded from a running system. The questions were
written by hand; every response was produced by cairn record calling the
engine, and every source passage is from the bundled synthetic demo corpus —
an invented county, invented programs, invented amounts. It demonstrates that
the instrument and the target work together. It measures nothing about any
real benefit program.
- 29 items (5 ar, 15 en, 9 es).
- 22 expected answers, 7 expected refusals, 6 of them adversarial probes.
- 14 items are not in English. 12 of them are
translations of an English item and carry
"review": "unreviewed", which every run says out loud; the remaining 2 were authored directly in their own language. No non-English string in this bundle — translated or authored — has been reviewed by a subject-matter expert, and claiming otherwise in an audit record would be the exact dishonesty that field exists to prevent.
python3 -m cairn index
python3 -m cairn recordRe-recording an unchanged corpus and configuration produces byte-identical files. A diff here is a change in behavior, and the bundle hash moving is the trace that says so.
| File | What it is |
|---|---|
items.jsonl |
The authored questions — including which passage answers |
| each one — plus every passage retrieval accepted for it | |
responses.jsonl |
What the engine replied, with the sources it cited |
| marked inline | |
sources.jsonl |
Every passage in the corpus, so a citation to something |
| that does not exist is detectable | |
interface.html |
A snapshot of the served page, with its colour pairs |
| declared so they can be checked rather than believed | |
checksums.json |
SHA-256 per file, and for the bundle |
Source ids here are Cairn passage ids with the # before the ordinal written
as a ., because the bundle format's inline citation grammar has no # in
it. grocery-allowance-en.2 is grocery-allowance-en#2; nothing else about
the identifier changes, so an audit finding maps straight back to a passage.