コンテンツにスキップ

English V4 Launch Plan (P6-E, dd_v8_03)

Target slot: T+150d post AutonoMath launch (2026-05-06 + 150 = 2026-10-03). Implementation budget: 10 working days spread across the slot. Owner: B6 subagent (this file is the spec, not the rollout log).

Why a separate plan: English V4 is a Y2 ARR uplift bet (¥100-150k/mo). Cohort = foreign-born startup founders, visa sponsors, cross-border M&A teams that already pay for English Stripe / GitHub / Notion docs but have no Japanese counterpart for 補助金 / 融資 / 税制 lookup.

The launch is scaffolded today (2026-04-25) so that activation is a single ~10-day push when the slot opens, not a multi-week dig-out from scratch. Today's deliverables: i18n module + style guide + this plan.

Scope decision (locked)

  • In: read-side localization for all 16 autonomath tools + 7 jpintel one-shot tools, 50+ key i18n catalog (expand to ~200), response_language parameter, llms-full.en.txt, docs/index.en.md, site/index.en.html, cohort outreach prep.
  • Out: machine translation (every English string is hand-curated), English-only program data (we do not re-translate program content), English support tickets / sales motion (zero-touch ops principle).
  • Constraint: ja default preserved. Every existing caller that does not pass response_language keeps Japanese output — no breaking change.

Day-by-day timeline

D1 — am_alias.language migration (collection-CLI)

Migration: scripts/migrations/040_am_alias_language.sql.

-- 040_am_alias_language.sql
-- P6-E: tag aliases by language so English fallback can resolve to
-- the official English transliteration of authorities, laws, programs.
ALTER TABLE am_alias ADD COLUMN language TEXT
    NOT NULL DEFAULT 'ja'
    CHECK (language IN ('ja','en','kana','romaji'));

CREATE INDEX ix_am_alias_language ON am_alias(canonical_id, language);

-- Backfill: alias_kind='english' rows → language='en'.
UPDATE am_alias SET language = 'en' WHERE alias_kind = 'english';
-- Backfill: alias_kind='kana' rows → language='kana'.
UPDATE am_alias SET language = 'kana' WHERE alias_kind = 'kana';
-- Backfill: rows whose alias matches a Latin-only regex → 'romaji'
-- (handled by the post-migration python script `scripts/i18n/backfill_romaji.py`
-- to keep regex out of SQLite for portability).
  • Owner: collection CLI worker (NOT this subagent — out of scope per the task constraints in feedback_data_collection_tos_ignore).
  • Verification: SELECT language, COUNT(*) FROM am_alias GROUP BY 1 should show ja >> en > kana > romaji after backfill.
  • Backout: ALTER TABLE am_alias DROP COLUMN language; (sqlite 3.35+; we are on 3.43, safe.)

D2-D3 — i18n message catalog expansion (~200 keys)

Source of truth: src/jpintel_mcp/i18n/__init__.py (scaffolding ~50 keys landed today). D2 expands by category:

Category Key prefix Count target Source of strings
Envelope (16 tools × 4 statuses) envelope.<status>.<tool> 64 envelope_wrapper.DEFAULT_EXPLANATIONS
Error user_message error.<code> 25 cs_features.USER_MESSAGES
Onboarding tips tips.<n> 30 cs_features.onboarding_tips_for_age_days
Suggestion templates suggest.<intent> 40 cs_features.derive_suggestions
Input warnings warn.<rule> 20 cs_features.derive_input_warnings
Meta block field labels meta.<field> 15 (REST + MCP envelope)
Discovery one-shots oneshot.<tool> 14 server.py 7 one-shots × ja+en hint
Total ~208

D3 adds the t() callsite swaps: every f"..." literal in envelope_wrapper.py and cs_features.py becomes t(key, lang).

D4-D5 — response_language parameter wiring (23 tools)

Tools that gain the parameter:

  • 16 autonomath tools in src/jpintel_mcp/mcp/autonomath_tools/ (the package's __all__ list — search_tax_incentives, search_certifications, list_open_programs, enum_values_am, search_by_law, active_programs_at, related_programs, search_acceptance_stats_am, intent_of, reason_answer, get_am_tax_rule, search_gx_programs_am, search_loans_am, check_enforcement_am, search_mutual_plans_am, get_law_article_am).
  • 7 one-shot discovery tools in server.py (smb_starter_pack, subsidy_combo_finder, deadline_calendar, dd_profile_am, similar_cases, regulatory_prep_pack, subsidy_roadmap_3yr).

Signature delta (backward compatible, ja default):

def search_tax_incentives(
    ...,
    response_language: Literal["ja", "en"] | None = "ja",
) -> dict:
    ...
    explanation = t(f"envelope.{status}.search_tax_incentives",
                    response_language or "ja")

Where to inject: envelope_wrapper.build_envelope already centralises explanation/suggestions/meta — add a single lang kwarg there and let each tool pass it through.

D6 — llms-full.en.txt

scripts/regen_llms_full_en.py (new) mirrors regen_llms_full.py but: - writes to site/llms-full.en.txt - pulls English aliases for program names from am_alias WHERE language='en' - prepends an English overview header - preserves the same compact-program inventory structure so AI agents can switch between the two with one URL change

D7 — docs/index.en.md + mkdocs nav

docs/index.en.md (new) — English landing page mirroring docs/index.md. Mkdocs material supports i18n plugin; we ship a hand-translated single page first, then evaluate whether the full mkdocs-static-i18n plugin is worth the build-time cost.

mkdocs.yml nav addition (locked in this PR):

nav:
  - English:
      - Overview: index.en.md
      - Getting Started: getting-started.en.md
      - API Reference: api-reference.en.md
      - Pricing: pricing.en.md

D8 — site/index.en.html

Already mature (site/en/{about,getting-started,index,pricing,tokushoho}.html landed pre-launch). D8 task is a refresh pass to align with the post-launch copy in site/index.html + add a top-of-page banner pointing at the new English MCP docs.

D9 — Cohort outreach prep (5 personas, English)

5 persona briefs translated to English (under docs/_internal/personas.en/ — _internal/ excluded from public mkdocs build per mkdocs.yml):

  1. Foreign-born startup founder (visa-aligned 経営管理 holder) — needs 創業補助金 in English to file with their Japanese lawyer.
  2. Visa sponsor (HR ops at multinational) — needs subsidy / certification info in English to brief the founder team.
  3. Cross-border M&A team (overseas PE) — needs 行政処分 + 採択 stats in English for due diligence.
  4. Foreign accountant / tax advisor in Japan — needs 税制特例 sunset dates in English to advise English-speaking clients.
  5. Foreign-language LLM agent (Claude / GPT in non-Japanese context) — needs llms-full.en.txt to ground its answers without hallucination.

D9 also drafts a single English launch tweet + LinkedIn post; no paid ads, no cold outreach (organic-only principle).

D10 — Launch + monitoring

  • Flip mkdocs.yml nav to expose English entries.
  • Publish llms-full.en.txt.
  • Tweet + LinkedIn post.
  • Watch for 7 days: error rate per response_language segment, English cohort signup count vs Japanese baseline, English MCP call share.
  • If English share < 2% by D17, freeze further investment. If > 5%, fund Phase 2 (machine translation pilot for the program-content layer).

Backward compatibility checklist

  • response_language defaults to "ja" everywhere.
  • t(key) defaults to lang="ja".
  • Existing tests (tests/test_envelope_cs_features.py, etc.) keep passing without modification.
  • OpenAPI schema diff: only additive (new optional parameter).
  • No DB write paths touched on jpintel.db side.

Risks & mitigations

Risk Mitigation
English strings drift from JP semantics Style guide + tests asserting key parity
Catalog explodes to thousands of keys Cap at ~200 in P6-E; never auto-translate program content
mkdocs-static-i18n adds build minutes Stay on hand-curated single page until D9 outreach lands real demand
English cohort never materialises D17 hard freeze; sunk cost is 10 days, recoverable
MT quality complaints Never ship MT — every English string is hand-curated

Do not do

  • Do not machine-translate the program rows (programs.title_jp, summary_jp). The corpus is legally sensitive (景表法 / 消費者契約法) and a bad translation creates 詐欺 risk.
  • Do not create English-only support / sales channels. Zero-touch principle holds.
  • Do not add an English tier badge or English-only price tier. The pricing model stays ¥3/req metered, language-agnostic.
  • Do not rename internal jpintel_mcp paths to add _en suffixes. PyPI package and import path stay legacy.

References

  • Memory: feedback_autonomath_no_ui — value lives in API/MCP/static docs.
  • Memory: feedback_zero_touch_solo — no English sales / CS team.
  • Memory: project_autonomath_business_model — ¥3/req metered, no tiers.
  • Source: src/jpintel_mcp/i18n/__init__.py — catalog scaffolding.
  • Source: docs/i18n_style_guide.md — English tone + bilingual conventions.
  • Source: src/jpintel_mcp/mcp/autonomath_tools/envelope_wrapper.py — JP message authority (D2-D3 mirrors this).