[lab · T3] Ranking & Packing — track thread

[akash-coded]

L.A.B. track thread. One per track. Ask here rather than opening a new thread — the
answers accumulate where the next person will look.

Labs: L05, L06 · Lifecycle stage: Development · the method

What this track is about

Two retrievers' scores are not interpersonally comparable, so fusion consumes ranks. Then the packer has to spend a hard token budget without breaking a citation.

Where people get stuck

L06's hidden check on greedy backfill is the one that catches people. Filling leftover budget with a smaller lower-ranked chunk uses the budget better and makes inclusion depend on the sizes of the chunks above it — so a metric change can no longer be attributed to your ranker. Keep the packer boring so the ranker stays measurable.

How to ask here

Paste the check that failed and its message. The messages are written to be actionable — if
one is not, that is a bug in the lab and saying so is a useful contribution.

python scripts/lab.py run <id>            # public checks
python scripts/lab.py run <id> --hidden   # everything

Submitting

Fork, fill in starter.py, open a pull request. The
Labs workflow posts your results as a comment and
edits it in place on each push. It will not fail the build on red checks — the comment says
what failed.

What this track hands to the next one

Development produces the artefact the following stage needs. If you finish L05, L06 and cannot say what
that artefact is for your own corpus, the track is not done — that is the point of the lifecycle
running underneath the pathway rather than beside it.