コンテンツにスキップ

STATE — Wave 46 tick#4 (永遠ループ) audiences cost-saving PR

Status: PROPOSED · Generated: 2026-05-12 · Branch: feat/jpcite_2026_05_12_wave46_audiences_cost_saving

PR#: pending (open after admin merge of tick#3 #151 already landed; this is the next consumer of the canonical doc docs/canonical/cost_saving_examples.md).

1. Scope (3 audience pages)

Wave 46 tick#3 landed pricing + compare cost-saving framing on 2026-05-12 (PR #151). tick#4 extends the same "純 LLM vs jpcite ¥3/req" framing to the 3 highest-traffic audience pages so the saving figures users see on each audience landing match the canonical doc.

Pages touched (3 files):

  • site/audiences/ma_advisor.html
  • site/audiences/cpa_firm.html
  • site/audiences/shindanshi.html

Out of scope (deliberate): the other 14 audience pages (construction.html / tax-advisor.html / shihoshoshi.html / vc.html / journalist.html / manufacturing.html / real_estate.html / shinkin.html / shokokai.html / smb.html / subsidy-consultant.html / dev.html / admin-scrivener.html / index.html) keep the existing baseline framing for now and will be cleaned in subsequent ticks of this loop. Reason: dual-CLI lane atomicity (mkdir guard) + smaller verified diff.

2. Per-page saving table

Headline cost-saving figures advertised on each audience landing (consistent with docs/canonical/cost_saving_examples.md token-cost model: 純 LLM token 単価 ¥300/1M, jpcite ¥3/req tax-excluded).

audience unit 純 LLM コスト jpcite ¥3/req 節約 / unit rough off
ma_advisor (M&A advisor) 1 deal × 50 req ¥15,000/deal ¥150/deal ¥14,850/deal ~99%
cpa_firm (公認会計士事務所) 月次 100 社 × 5 req = 500 req/月 ¥150,000/月 ¥1,500/月 ¥148,500/月 ~99%
shindanshi (中小企業診断士) 月次 30 顧問 × 10 req = 300 req/月 ¥90,000/月 ¥900/月 ¥89,100/月 ~99%

2.1 Per-page task breakdown (top of the new Cost saving calculator table)

ma_advisor.html (1 deal × 50 req, sums to ¥150 jpcite / ¥15,000 LLM):

# task 純 LLM ≈ jpcite 節約
1 法人 360° (対象会社 surface) ¥4,500/deal 15 req ¥45 ¥4,455/deal
2 DD deck 作成 ¥2,700/deal 9 req ¥27 ¥2,673/deal
3 補助金返還義務 chain ¥3,000/deal 10 req ¥30 ¥2,970/deal
4 業法 fence 検証 ¥2,400/deal 8 req ¥24 ¥2,376/deal
5 事業承継 制度 mapping ¥2,400/deal 8 req ¥24 ¥2,376/deal
Σ (1 deal × 50 req) ¥15,000/deal ¥150/deal ¥14,850/deal (~99% off)

cpa_firm.html (月次 100 社 × 5 req = 500 req/月):

# task 純 LLM ≈ jpcite 節約
1 措置法 42-4 lookup ¥30,000/月 100 req ¥300 ¥29,700/月
2 IT 導入補助金 会計処理 dual check ¥40,000/月 100 req ¥300 ¥39,700/月
3 被監査会社 DD question deck ¥35,000/月 100 req ¥300 ¥34,700/月
4 行政処分歴 filter ¥30,000/月 100 req ¥300 ¥29,700/月
5 監査調書 template scaffold ¥15,000/月 100 req ¥300 ¥14,700/月
Σ (月次 100 社 × 5 req) ¥150,000/月 ¥1,500/月 ¥148,500/月 (~99% off)

shindanshi.html (月次 30 顧問 × 10 req = 300 req/月):

# task 純 LLM ≈ jpcite 節約
1 月次 saved search (30 顧問) ¥24,000/月 90 req ¥270 ¥23,730/月
2 eligibility chain ¥18,000/月 60 req ¥180 ¥17,820/月
3 renewal forecast ¥24,000/月 60 req ¥180 ¥23,820/月
4 補完制度 探索 ¥12,000/月 45 req ¥135 ¥11,865/月
5 申請 kit scaffold ¥12,000/月 45 req ¥135 ¥11,865/月
Σ (月次 30 顧問 × 10 req) ¥90,000/月 ¥900/月 ¥89,100/月 (~99% off)

3. Replaced surface text (audit log)

For each page, the following 4 elements were replaced consistently:

  • <meta name="description"> — old ROI / delta framing → "節約 ¥… (約 99% off)"
  • <meta property="og:description"> — same swap
  • JSON-LD FAQ "コストは?" answer — same swap, new saving line + 99% off note
  • Main <section class="features" aria-labelledby="roi-title"> → renamed aria-labelledby="cost-saving-title". The <h2> heading is now "Cost saving calculator (... 純 LLM vs jpcite ¥3/req)". The 5-task table columns become task / 純 LLM コスト (baseline) / jpcite ¥3/req / 節約 ¥/.... The trailing footnote replaces ROI multiplier language with the ¥300/1M token unit-price assumption and the "削減保証ではない" disclaimer.

Out of scope on these pages (preserved untouched):

  • hero / breadcrumb / fence section
  • MCP install snippet (Claude Desktop config)
  • CTA buttons (?src=audiences_* query params remain)
  • legal-note footer (業法 fence 文言 unchanged)
  • JSON-LD Service / Offer / FAQPage other Q&A nodes (only the cost answer was edited)
  • common JSON-LD @graph org + price block

4. Tests

New: tests/test_audiences_cost_saving.py — 10 tests, all PASS locally (/Users/shigetoumeda/jpcite/.venv/bin/pytest, Python 3.13.12):

  1. test_canonical_doc_exists
  2. test_no_roi_arr_yarn_framing (ROI / ARR / 射程 0-grep gate on 3 pages)
  3. test_each_page_has_cost_saving_section
  4. test_each_page_saving_amount_present_and_in_doc
  5. test_each_page_links_to_canonical_doc
  6. test_brand_consistency (legacy 税務会計AI / AutonoMath / zeimu-kaikei.ai 0)
  7. test_unit_price_constant (¥3 + 従量 wording present)
  8. test_structural_anchors_intact (hero/fence/cost-saving/install/cta)
  9. test_html_parses_clean (stdlib html.parser, 0 errors per file)
  10. test_each_page_h2_count_preserved (exactly 4 h2 per page)

Result: 10 passed in 0.94s.

5. Verify gates (バグなし)

  • grep -cE "(ROI|ARR|射程|roi-title)" site/audiences/{ma_advisor,cpa_firm,shindanshi}.html → 3 × 0 (no leak)
  • grep -cE "(zeimu-kaikei|税務会計AI|AutonoMath|autonomath\.ai)" site/audiences/{ma_advisor,cpa_firm,shindanshi}.html → 3 × 0 (no legacy brand leak)
  • stdlib html.parser on all 3 files → 0 errors, h2 = 4 each, h3 = 0
  • pytest tests/test_audiences_cost_saving.py → 10/10 PASS

6. Files in this PR

  • site/audiences/ma_advisor.html (modified)
  • site/audiences/cpa_firm.html (modified)
  • site/audiences/shindanshi.html (modified)
  • tests/test_audiences_cost_saving.py (new, ~220 LOC)
  • docs/research/wave46/STATE_w46_tick_audiences_pr.md (this file, new)

No file rm/mv (破壊なき整理整頓ルール 遵守). No main worktree usage (dual-CLI lane atomic via mkdir /tmp/jpcite-w46-audiences-cost.lane). No LLM API import added.

7. Loop continuation

The remaining audience pages (14 files) can be cleaned in subsequent ticks of this loop, each chunk ≤ 3 files to keep PR diff reviewable and lane atomicity intact. Priority order suggested for the next tick:

  1. tax-advisor.html (税理士) — case 1 base = ¥15,300/月 already canonical
  2. subsidy-consultant.html (補助金コンサル) — case 6 base = ¥8,160/月
  3. shihoshoshi.html (司法書士) — derive from case 6 + 法務 weighting

After all 17 audience pages, follow up with industry/*.html and roles/*.html sweeps if those exist.