[ParsaVictor]

Brings the measured results from contributions-log §8.19–§8.21 (#150, #152) into paper/. This PR changes no benchmarks or code.

New tables

• tables/ceiling.tex, the rerank ceiling (§8.20). It shows the share of issues whose every gold function appears in the engine's first k definitions:

  • Lite-274: 0.828 at k=300, 0.672 at k=30;
  • Verified: 0.790 at k=300;
  • dev: 0.776; SWE-Gym: 0.748.

  The text compares this with SweRank's best function Acc@5 of 0.814, but only for scale. It says explicitly that a recall ceiling is not an accuracy. It also names two higher systems, SweRank+ at 0.854 and SWE-Adept at 0.878, so we do not overclaim. The ceiling is now in the abstract, the introduction and the contributions list (new item 3).

• tables/swerank.tex, SweRankEmbed-Small as the short-list encoder (§8.19, §8.21).

  • The k=30 vote passes the pre-registered bar on all four tuning halves. The k=100 vote fails on dev-rest and is reported as a negative result.
  • Run once on each holdout, function Acc@5 goes from 0.460 to 0.496 on Lite (+0.036 [0.011, 0.062], p=0.013) and from 0.428 to 0.472 on Verified (+0.044 [0.024, 0.066], p=0.0001).
  • It is marked research-only because of its cc-by-nc licence. The jina vote stays as the shippable row.

Other updates

• Table 1 gets a SweRankEmbed-Small vote row. The text gives the gap to the same model used as a whole-repository retriever: −0.135.
• The cost table now includes SweRank's per-issue cost (11–15 s on CPU, fp32).
• The text notes that the jina holdout gains are a lower bound, because the stored runs kept only 20 candidates per issue.
• A footnote and a new threat paragraph explain that the lists were regenerated with the release binary (299 of 300 file lists match the stored run).
• The threats section has a new forking-paths paragraph.
• REVIEWER-RISKS.md: P0-1: stable_project_id + .neuromesh/id #1 is rewritten around the ceiling and includes the counter-objection "recall@300 is not Acc@5". P0-5: stop workspace_root drift in reindex_incremental #5 is updated.

python paper/check.py passes: 39/39 citations and 25 labels. The paper has not been compiled, because there is no TeX on this machine.

🤖 Generated with Claude Code