eval(round-4): expand results.json to match abstract scope; breaks the 4-ceiling

Hypothesis from Round 3: the Overall=4 cap comes from abstract overreach
vs results.json scope, not from pipeline limits. Test: hand-build a
63-metric / 3-table results.json that actually covers what the
Spectral-Gated LoRA abstract promises (ViT-B + ViT-L + DINOv2-B × LoRA +
DoRA + AdaptFormer + FacT + VPT + Full FT × VTAB-1k + CUB + Aircraft +
Cars + Flowers + Pets + DTD + KTH-TIPS-2b, plus expanded ablation).
Same ideas.json, same prompts/bib as Round 3 — only results change.

Measured (Round 4 vs Round 3):
  verify claims total:     113 → 205   (+92; more surface to audit)
  verify rate:            0.81 → 0.88   (highest observed)
  self-review Overall:       4 → 5      (BROKE the 4-ceiling)
  sub-scores stable:      Originality 3, Clarity 3, Presentation 3
  Decision:             Reject → Reject

Reviewer weakness CLASS shifts form-to-substance:
  R1: "Results forthcoming" / Discussion fabrication    [form]
  R2: abstract overreach vs results scope              [form]
  R3: abstract overreach + partial baselines           [form + content]
  R4: effect sizes within seed noise; scale study      [PURE SUBSTANCE —
      partially undermines core claim; promised          research-quality
      gate-entropy analysis not presented                critique]

At Round 4 the reviewer is giving the kind of NeurIPS-reviewer write-up
a real incremental PEFT paper would receive. Further score gains are not
available by prompt tuning — they require a genuinely more significant
research idea or real experiments with larger gaps. Which is correct.

Artefacts added:
  skills/vibe-sci/references/generation_examples/
    results_v2_expanded.json              — 63 metrics, 3 tables
    generated_paper_round4.tex            — 4-ceiling-break paper
    verification_report_round4.json       — 181/205 verified (0.88)
    self_review_round4.json               — Reject/Overall=5

Evaluation report updated with Round 4 findings and a final summary
answering the /loop goal: pipeline is instruction-following and data-
driven; research merit is the human's responsibility; the loop closes
at 5 because pushing higher requires either cherry-picking (violates
the user's stated goal) or a better idea (not a pipeline lever).

Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>

Author: jeffhsu-chiuchau

skills/vibe-sci/evals/2026-04-21_structural_iteration_loop.md

@@ -71,11 +71,58 @@ Ran writeup with identical `ideas.json` + `results.json` as Round 2 so any chang
 
 Structural-fidelity loop closes here. The lever for higher reviewer scores is now *the quality of results.json a user supplies*, not the scaffolding. That's the right separation of concerns for an "autonomous research writer" — the writer is doing its job; the humans still have to run the experiments.
 
+## Round 4 — expand `results.json` to match abstract scope (break the 4-ceiling?)
+
+Round 3 concluded the remaining cap to Accept was `results.json` not covering what the abstract promises (ViT-L/16, DINOv2-B, DoRA/AdaptFormer/FacT/VPT baselines). Round 4 tests that hypothesis directly: hand-build `results_v2.json` (63 metrics, 3 tables, covering ViT-B × 6 methods × 8 benchmarks + ViT-L/DINOv2 scale rows + fuller ablation). Same ideas.json, same prompts/bib as Round 3 — only the results change.
+
+### Round 4 results
+
+| Metric                  | R1 (no results) | R2 (19 metrics) | R3 (+bib, +algo) | **R4 (63 metrics, match scope)** |
+|-------------------------|-----------------|-----------------|-------------------|------------------------------------|
+| verify claims total     | 30              | 86              | 113               | **205**                            |
+| verify rate             | 0.57            | 0.73            | 0.81              | **0.88**                           |
+| self-review Overall     | 2               | 4               | 4                 | **5** ← broke the 4-ceiling        |
+| Originality             | —               | 2               | 3                 | 3                                  |
+| Clarity / Presentation  | —               | 2/2             | 3/3               | 3/3                                |
+| Decision                | Reject          | Reject          | Reject            | Reject                             |
+
+### Qualitative shift: weakness class changes at R4
+
+Reviewer weaknesses R1 → R4 follow a clean form-to-substance progression:
+
+- **R1 (Overall=2)**: "Results forthcoming" / Discussion fabricates outcomes → **form**
+- **R2 (Overall=4)**: abstract overreach vs results scope → **form**
+- **R3 (Overall=4)**: abstract overreach + partial baselines → **form + content mix**
+- **R4 (Overall=5)**: effect sizes within seed noise (Flowers +0.1, Pets +0.1, Aircraft +0.2); scale study "partially undermines the core claim" since ViT-L and DINOv2-B gaps shrink; promised per-layer gate-entropy analysis is never presented → **pure substance critique of the underlying research**
+
+R4's weaknesses are what a real NeurIPS reviewer would write about an incremental PEFT paper. They're not "your writeup is broken" complaints; they're "your research is modest". The pipeline has run out of scaffolding leverage — further improvement requires a better idea or better experiments, not a better prompt.
+
+### Why R4 caps at 5, not 7
+
+The FlashAttention-2 comparator (Overall=7) is a genuine 2× speedup on a core primitive with MFU rising to 72%. SG-LoRA is a +0.3–1.1 point PEFT variant that beats DoRA on texture only. At NeurIPS-reviewer calibration, "clean, honest, incremental PEFT work" lands at 5 — borderline reject, solid workshop candidate. The reviewer gave exactly that score. Picking a higher-impact idea (or running real experiments where a larger gap shows up) is how you reach 6+; it is not achievable by rewriting prompts or expanding bib.
+
+This is the intended separation of concerns vindicated: **pipeline is instruction-following and data-driven; research merit is the human's responsibility**.
+
+## Final four-round summary (answering the /loop goal)
+
+> "從真實論文中取得結構 迭代論文生成品質 — 目標不是要做出假以亂真 而是真的遵循結構與數據驅動自動產生論文的能力"
+
+- **Structural fidelity**: confirmed. Tables 1:1 render from results.json; `\cite{…}` scales with bib size (3 → 45 for a single bib swap); pseudocode appears when the instruction asks for it; `\ref{}` + `\label{}` are consistent.
+- **Data-drivenness**: confirmed. Verify rate climbs monotonically (0.57 → 0.73 → 0.81 → 0.88) as input richness grows; writer quotes exact values from metrics; speculation shrinks.
+- **Reviewer calibration works end-to-end**: R1=2 (no results) / R2=4 (partial results, form issues) / R3=4 (cleaner form, same content) / R4=5 (content-limited ceiling for a modest research idea).
+- **No "假以亂真"**: the pipeline does not let the writer fake results. It *can* speculate in Discussion without `--results-json` (caught by verify.py in R1). With rich results, it surfaces whatever is actually in the registry and the reviewer sees small gains as small gains.
+
+**Loop closes at Round 4.** Pushing past 5 would require cherry-picking results.json (misses the user's goal) or picking a genuinely more significant idea (a human decision, not a pipeline lever).
+
 ## Artefacts promoted to repo
 
 - `references/generation_examples/generated_paper_round3.tex` — patched-bib paper
 - `references/generation_examples/verification_report_round3.json` — 92/113 verified
 - `references/generation_examples/self_review_round3.json` — Reject/4 with Originality=3, Presentation=3
+- `references/generation_examples/results_v2_expanded.json` — 63-metric results matching abstract scope
+- `references/generation_examples/generated_paper_round4.tex` — 4-ceiling-break paper
+- `references/generation_examples/verification_report_round4.json` — 181/205 verified (rate 0.88)
+- `references/generation_examples/self_review_round4.json` — **Reject/Overall=5** (best generation score)
 
 ## Code changes committed alongside