v0.1.0 — the first release whose numbers you can check

[akash-coded]

v0.1.0 · the harness is tagged.

Ten notebooks, a from-scratch toolkit, an eval gate that blocks a merge on a metric regression,
12 auto-graded labs, and a documentation site at
akash-coded.github.io/nanorag.

What is worth your attention

Three measurements that contradict the material, kept rather than tuned away. RRF losing to
weighted fusion, comparison starvation not reproducing, and no retrieval-score threshold separating
answerable from unanswerable. Each carries how to re-test it and the number that would
falsify it — see the README.

A fourth was added after a discussion thread: the learned reranker made retrieval worse at
every k, because a reranker over the same signals as the first stage is an expensive identity
function.

Eight ADRs, each with a falsifier. "What would change this" is a required section. A decision
without one is a preference.

What this release does not claim

Every number here was measured on a synthetic corpus this repository generated itself. That
choice is deliberate — gold labels are true by construction, so there is no annotation-error floor
under any result — and it means every finding has an untested precondition.

The extension project exists to find out which
of them survive contact with real data. E0–E6, 17 issues, and a board field with three values:
reproduced, contradicted, inconclusive.

Contradicted is the most valuable outcome. It would say precisely what a synthetic corpus
cannot teach, which is a thing almost nobody writes down.

Where to start

You are	Start
new here	docs/00-orientation/start-here.md
wanting to practise	the L.A.B. simulator
preparing for an interview	the question bank
here for one page in a meeting	the cheat sheets

If you find a number that does not reproduce

Open an [errata] thread in Q&A. That is the highest-value contribution anyone can make here, and
it has already happened once — #85 caught this repository quoting the
Landis–Koch κ bands as though they were a standard.