Sample End-to-End Run Record

August 14, 2026 · View on GitHub

Validation proof: a full end-to-end run on the real 2023 CUMCM Problem C — "Automatic Pricing & Restocking of Vegetable Items", exercising the whole pipeline: modeling+coding role → paper role → Git handoff → independent review.

  • Collaboration repo: math-team/ (Gitee)
  • Split: member-a (modeling+coding) delivered Q1–Q4; member-c (paper) wrote/reviewed with vision QA
  • All data is real from the problem's attachments; every number in the paper traces back to results/ and the figures

1. Problem & data

  • Problem: 2023 CUMCM Problem C (vegetable pricing & restocking), 4 sub-problems
  • Data size:
    • Attachment 1: 251 items / 6 categories
    • Attachment 2: 878,503 rows of sales flow (2020-07-01 ~ 2023-06-30, includes returns and discount flag)
    • Attachment 3: 55,982 rows of wholesale price
    • Attachment 4: loss rates (6 sub-categories + 251 items)

2. Collaboration chain (A → C)

member-a (modeling+coding)
  ├─ 题目分析报告.md      ← read problem, split sub-problems, modeling constraints
  ├─ 术语表格.md
  ├─ 问题1~4 scripts       ← actually run
  ├─ results/             ← distribution/relationship/price-volume/restock plan/sensitivity CSV + reproducibility manifest.json
  └─ figures/             ← raw/process/result, covering Q1–Q4
        │  git add member-a && push
        ▼
member-c (paper) git pull
  ├─ W1 evidence outline
  ├─ 论文.tex + 论文.pdf   ← XeLaTeX (TinyTeX)
  ├─ 代码/                ← appendix B core code (read-only copy from member-a)
  ├─ 评审记录.md           ← W1/W2/independent review/multiple fix iterations
  └─ ai use disclosure

Throughout, each committed only their own member-* folder — no interference; the paper's evidence all comes from member-a's real deliverables.

3. Member A (modeling+coding) results

Problem 1 — distribution & relationships

  • Long-tail concentration: Top5 items = 25%, Top25 = 64%, Top50 = 84% of volume
  • Daily-item sales right-skewed (mean ≈10.1 kg / median 5.2 kg); log-normal AIC beats gamma (304,372.5 < 660,894.3)
  • Day-of-week factor: Fri 18.2%, Sat 17.6% ≫ weekdays 12.4%–12.8%
  • Seasonal: peaks in Jul–Sep (summer) and Jan (CNY)
  • Category correlation: mostly weak–moderate positive; eggplant is relatively independent (Pearson 0.07–0.31)
  • Discounts identify price-sensitive samples (5.4% discounted, ~7–8 off)

Problem 2 — category price-volume + next-week restock/pricing

  • Constant-elasticity demand function (price elasticity + day-of-week/month/holiday factors), 2SLS correction
  • Elasticity: foliage -0.364, chili -0.647
  • Next week (Jul 1–7) per-category daily restock volume + mark-up pricing strategy
  • Roll backtest profit improvement +56.2% (8289.98 vs 5589.42)

Problem 3 — single-item restocking (27–33 items)

  • Selected 33 of 49 sellable items, each ≥2.5 kg
  • Selection criterion = profit potential per unit of display space (space-constraint driven)
  • Demand-satisfaction lower-bound sensitivity scan + satisfaction-revenue trade-off
  • Jul 1 single-day profit optimization

Problem 4 — data collection recommendations

Proposed 7 types of data (next-day actual purchase price, true demand, promo/display data, dynamic loss, weather/traffic, return reasons, customer profiles) with rationale, based on gaps actually encountered during modeling.

4. Member C (paper) results

  • Paper.pdf / Paper.tex: XeLaTeX compiles, body + appendix 33 pages (appendix B core code ~6 pages)
  • 18 formal figures: 6 each for Q1/Q2/Q3 (raw/process/result); Q4 presented as a table
  • Numbering continuity audit: Fig 1–18, Table 1–12, Eq 1–10, Listing 1–5 all continuous; cite ↔ bibliography 6↔6, no dangling references
  • W1 evidence outline / W2 paper final-check: passed

Independent review & multiple fix rounds (bounded iteration + vision sub-agent)

  1. First review PASS (all low)
  2. Enhancement round + re-review FAIL: caught and deleted a fabricated number "0.78" (replaced with honest interval (0.7,0.8)); fixed self-contradictory elasticity tiers and misattributed improvements; re-review PASS
  3. Code-appendix round: caught two reproducibility defects (return-signature mismatch, undefined rho); re-review PASS
  4. Numbering continuity: found and fixed an appendix table missing \label that caused a table-number jump 10→12
  5. Figure round: vision sub-agent flagged Figure 2 as visually misleading (a clipped histogram masked the "log-normal beats gamma" conclusion); redrawn as dual panels and re-validated PASS

Note: the reviewing main model itself cannot read images (deepseek-v4-flash has no image input), so visual QA of all 18 figures was performed by the opencode-go/mimo-v2.5 vision sub-agent, and this was explicitly disclosed in the paper.

5. What this plugin pack contributed

CapabilityWhere it showed up
3-stage workflow & gates (M1/P1/P2, W1/W2)member-a modeling/coding delivery, member-c writing/final-check
3-folder collaboration (member-a/b/c, no interference)A delivered → C pulled and wrote; conflict-free
Vision sub-agent (mimo-v2.5 visual review)visual QA of 18 figures; caught & fixed the Figure 2 visual misdirection
Bounded iteration + independent review gateblocked fabricated 0.78; fixed numbering jump and visual misdirection
Reproducibility manifestresults/复现清单.json, results traceable
Git collaboration + proxy (sslBackend=openssl)clone/pull/push all the way

6. Artifact index

  • Analysis/build: member-a/问题1~4_求解.py, member-a/results/*.csv, member-a/figures/*
  • Paper: member-c/论文.pdf, member-c/论文.tex, member-c/代码/
  • QA: member-c/评审记录.md, member-c/W1 证据大纲.md, member-c/ai使用声明.md

Note: member-b's earlier Q3 delivery was superseded by member-a's full Q1–Q4; this paper's evidence relies on member-a's deliverables, as honestly documented in the review record — no fabricated changes to others' contributions.