docs: stale R@5=0.152 references in several files (recent measurement is 0.595)

[hang-in]

Hi jaytoone,

While explaining CTX's external R@5 to a community thread, I noticed something in the docs and wanted to share it gently in case it's useful — no urgency at all.

What I noticed

The current master cites two different external R@5 values, sometimes in the same file:

• CLAUDE.md L43-44: External R@5 = 0.595 (labeled as current)
• CLAUDE.md L91: 외부 코드베이스 R@5=0.152 (heuristic 과적합) (listed as a weakness)

The 0.595 value matches benchmarks/results/reeval_external_iter11.json exactly:

Repo	iter11 R@5
Flask	0.6462
FastAPI	0.3870
Requests	0.7526
Mean	0.595

The 0.152 value comes from docs/research/20260326-ctx-methodology-comparison.md L70, which itself notes it as a pre-fix result with re-run pending. After the generalization fixes (720380f, 727b5c3) and the iter11 re-measurement, several follow-on docs continued to cite the 0.152 baseline:

• CLAUDE.md (L91, L197)
• docs/research/20260327-ctx-alternatives-research.md
• docs/research/20260402-production-context-retrieval-research.md
• docs/research/20260326-ctx-goal1-goal2-final.md
• docs/paper/CTX_paper_draft.md
• a few others

Suggestion

1. Update the "weakness" / "future work" wordings that cite 0.152 to reference the iter11 measurement (Mean R@5 = 0.595).
2. Keep the 0.152 in methodology-comparison.md L70 as an academic record (it's already labeled as pre-fix), maybe with a forward pointer to reeval_external_iter11.json.

One related question — docs/benchmark/g1_g2_publication_framework.md cites Mean R@5 ≈ 0.744 from a different evaluation framework. If you happen to know which one is canonical, that would help me reflect it accurately on my end too — but no rush.

If you'd like a hand with the docs cleanup, just let me know. Otherwise, please feel free to handle it whenever convenient.

Thanks again for sharing this project.

[pluto2060]

Thank you for the careful audit — this is exactly the kind of issue that is easy to miss.

Fixed in commit pushed now: updated R@5=0.152 → 0.595 in:

• CLAUDE.md (L91, L197)
• docs/research/20260326-ctx-goal1-goal2-final.md
• docs/research/20260327-ctx-alternatives-research.md
• docs/research/20260402-production-context-retrieval-research.md

Kept 20260326-ctx-methodology-comparison.md L70 as-is (pre-fix academic record) per your suggestion — good call.

On the g1_g2_publication_framework.md question: the canonical number is 0.595 (iter11, Flask/FastAPI/Requests mean). The 0.744 in that doc was an earlier intermediate measurement before the generalization fixes — it should be treated as superseded by iter11.

[hang-in]

Thank you for the quick turnaround — and especially for the canonical-vs-superseded clarification on 0.595 vs 0.744. That answers exactly what I was unsure about.

I'll align our fork's docs to reference 0.595 as canonical going forward, with a pointer back to this issue for the rationale. Appreciate the fast and thorough fix.