Long-Term Strategy¶

Audience: AutonoMath operator (Bookyou株式会社, BDFL solo). Quarterly review document. Status: Y0 launch is 2026-05-06. This file covers Y1-Y5 (2026-05 → 2031-05). Pricing baseline: ¥3/req tax-excluded (税込 ¥3.30), 50 req/月 anonymous free per IP, 100% organic acquisition, solo + zero-touch ops. No tier SKUs, no seat fees, no annual minimums. See pricing.md.

This document encodes the BDFL solo + zero-touch + organic-only constraints from feedback_zero_touch_solo, feedback_organic_only_no_ads, and project_autonomath_business_model. Any decision that violates those constraints (e.g. hiring a CS team, running paid ads, signing DPA negotiations) should be challenged against this doc before action.

1. Three scenarios (probability-weighted, not "guaranteed targets")¶

Scenario	Probability	Y5 ARR (FY2031 run-rate)	Trigger conditions
Best	30%	¥750M – ¥1.5B	V2 (Healthcare) + V3 (Real Estate) + V4 (English-language surface) all deliver ≥ 80% of Y1 SMB curve; ≥ 1 partner integration (freee / Money Forward / kintone / SmartHR / Claude Desktop) drives 30%+ traffic; "Japanese subsidy" as a search query becomes synonymous with AutonoMath in agent tool selection.
Base	50%	¥150M – ¥450M	V1 (subsidy + agri) holds; V2 launches but at 30-50% of V1 curve; organic SEO + GEO compounds without paid amplification; 1-2 vertical extensions max within 5 years.
Downside	20%	¥30M – ¥60M	V1 plateaus at SMB-only adoption; competitive aggregators (gBizINFO, jGrants direct, biz.stayway) regain share; we lose distribution to a freee/Money Forward integrated competitor; tax cliff dates (2026-09 invoice 2割特例 expire) create a one-shot demand spike followed by erosion.

Why probabilities, not promises. Past projections that promised ¥1B Y1 inflated investor / partner expectations and led to over-investment in directions that didn't compound. Probabilities force us to plan for the median, not the brochure.

Y1-Y5 ARR projection (3 scenarios, ¥3/req full metered)¶

            Y1 (2027)      Y2 (2028)        Y3 (2029)         Y4 (2030)        Y5 (2031)
Best 30%  : ¥30 - 45M  →  ¥120 - 240M  →  ¥330 - 600M  →  ¥600 - 1.0B   →  ¥750M - 1.5B
Base 50%  : ¥30 - 45M  →  ¥90 - 150M   →  ¥150 - 300M  →  ¥240 - 450M   →  ¥150 - 450M
Down 20%  : ¥18 - 30M  →  ¥30 - 60M    →  ¥30 - 60M    →  ¥30 - 60M     →  ¥30 - 60M

Y1 lower bound is anchored on (a) ¥3/req unit economics, (b) 100% organic = no marketing burn, (c) launch-day MAU of 5-10k and 12-month MAU growth to 15-25k authenticated + 100k+ anonymous, (d) average authenticated user @ 200-400 paid req/month, (e) stripe metered billing only — no contracts, no retention obligation. The math: 15k auth × 300 req/mo × ¥3 × 12 = ¥162M would be best-case Y1; we discount to ¥30-45M base because the curve is back-weighted (most users join in months 6-12 after SEO crawl).

Best-case Y5 (¥750M-1.5B) requires a 5x5 matrix maturity (5 verticals × 5 surfaces). It is not the current commitment — see Section 6.

2. Five moat priorities¶

Listed in descending catch-up impossibility for a future entrant. The ordering is: time-stamped data > human curation > customer relationship > regulatory expertise > breadth.

2.1 Time depth (catch-up: 36+ months, effectively impossible)¶

Source: am_amendment_snapshot already stores 14,596 amendment events with valid_from / valid_to semantics. Each row reflects an actual day-by-day legal change — not interpretable from the current law text alone.

Y1 target: 14,596 → 25,000+ (continue ingesting e-Gov diff feed)
Y3 target: 25,000 → 60,000+
Y5 target: 60,000 → 100,000+

A new entrant who launches in 2031 cannot reconstruct 2026-2031 amendment timestamps retroactively unless e-Gov publishes structured historical diffs (which they do not). This is the deepest moat AutonoMath has and the cheapest to maintain (passive ingestion).

2.2 Operator curation (catch-up: 18-24 months)¶

Source: hallucination_guard tier (currently 504 manually-rejected hallucinations + 181 exclusion / prerequisite rules). Each row is a real query that produced a wrong answer once, plus the corrected expected output. This is what stops us drifting into "looks fluent, is wrong" territory.

Y1 target: 504 → 1,500
Y3 target: 1,500 → 5,000
Y5 target: 5,000 → 10,000

Catch-up cost for a new entrant: equivalent to processing 100,000+ queries with manual correction overhead. We do this in-line as we audit (memory feedback_no_fake_data requires every claim to be 一次資料 verifiable).

2.3 Customer relationships (catch-up: 24-36 months)¶

Solo + zero-touch is not "no relationship" — it's "no human touch during the relationship". Customer relationship moat means:

Public testimonials (Y1 target: 10, Y3: 50, Y5: 200) — anonymized OK
Public case studies (Y1: 5, Y3: 20, Y5: 50) — Japanese SMB / 士業 stories
NPS public dashboard (Y2 onward, post-launch addition)
Documented "saved ¥X / found Y subsidy" before/after — embedded in docs blog

A future entrant must replicate trust signals; trust signals require time × people who have actually used the product. No shortcut exists.

2.4 Regulatory expertise (catch-up: 12-18 months)¶

Source: 景表法 / 個情法 / AI法 / インボイス compliance baked into product behavior:

Tier=X quarantine for non-verifiable claims
claim_strength tagging (avoiding 「最も」「絶対」「保証」 in search responses)
source_fetched_at honest semantics (出典取得, never 最終更新)
10-keyword pre-emptive block list for 詐欺 risk phrases
tokushoho / pepper / honesty compliance docs at /compliance/*

A new entrant can read law text and write a clone, but operationalizing it across 13,578 program rows + 5.26M facts without false-positive blocks of legitimate claims requires iteration that takes 12-18 months of real traffic.

2.5 Multi-domain breadth (catch-up: 18-30 months per new vertical × 5)¶

The 5x5 matrix (5 verticals × 5 surfaces) requires:

Vertical (V)	Status	Catch-up	Y target
V1 Subsidy / 補助金 / 農業 / 法令	✅ launch (2026-05-06)	—	continuous
V2 Healthcare / 医療 (program_healthcare)	T+90d (post-launch)	18-24m	Y1.5
V3 Real Estate / 不動産 (program_realestate)	T+200d	24-30m	Y2
V4 English-language surface	T+150d	12-18m	Y2
V5 Korean / Chinese surfaces	Y2-Y3 deferred	12-18m each	Y3-Y4

Surface variants (5): REST API, MCP stdio, LINE bot, embedded widget, 士業 affiliate. Each is wired today and can multiply into V2-V5 schemas at marginal cost (schema migrations 013+).

Note: V2-V5 are not Y1 commitments. They appear here so a Y2 review can decide go/no-go with a written baseline.

3. Year-by-year decision gates¶

Each gate has a quantitative tripwire. If we miss the lower bound for two consecutive quarters, we re-evaluate (downscale, pivot, or exit) — not optimize harder.

Y1 (2026-05-06 → 2027-05): "subsidy validation"¶

ARR target: ¥30M – ¥45M (lower-bound trip: ¥18M)
MAU target: 15k – 25k authenticated + 100k+ anonymous monthly
Subsidy validation gate: ≥ 20 published case studies, ≥ 5 testimonials, NPS ≥ 30
Decision at Y1+9mo (2027-02): green → Y2 plan; red → simplify to V1-only and lower the Y2 ceiling
Ops cost ceiling: ≤ ¥6h / week solo time on launch-day operations (memory feedback_organic_only_no_ads)

Y2 (2027-05 → 2028-05): "second vertical entry / runway decision"¶

ARR target: ¥90M – ¥150M (lower-bound trip: ¥60M)
V2 (Healthcare) launches, T+90d V2-launch metric: ≥ 500 program rows ingested, ≥ ¥5M added MRR within 6 months
Decision: Series A (if best-case curve), keep solo (if base), or sunset V2 (if downside)
Hire-1-ops gate: NOT yet — even at ¥150M ARR, solo + zero-touch holds. Only hire if MRR > ¥3M sustained 3 months AND solo time exceeds 20h/week (memory feedback_zero_touch_solo)

Y3 (2028-05 → 2029-05): "scale or maintain"¶

ARR target: ¥240M – ¥450M (lower-bound trip: ¥150M)
V3 (Real Estate) decision: launch (best/base) or defer (downside)
Hire-1-ops trigger: MRR > ¥3M sustained AND ≥ 20h/week solo time → hire 1 ops (specifically: data ingestion automation + Stripe billing edge cases). NOT a CS hire (memory bans CS team)
Series A consideration: only if V2 + V3 both > 30% of V1 ARR. Otherwise, stay bootstrapped.

Y4 (2029-05 → 2030-05): "exit-option year"¶

ARR target: ¥450M – ¥900M (lower-bound trip: ¥300M)
Series A / acqui-hire / IPO candidacy review. The 5 moats accumulated by Y4 (time depth ≥ 60k snapshots, curation ≥ 5,000 rows, ≥ 20 case studies, regulatory DB mature, ≥ 3 verticals live) make AutonoMath an attractive acqui-hire for a tax / accounting platform (freee, Money Forward) or an enterprise RAG vendor.
Decision: Sell at Y4 (lock-in 5y outcome) vs. continue to Y5 maturity.

Y5 (2030-05 → 2031-05): "maturity"¶

ARR target: ¥750M – ¥1.5B (best/base) or ¥30M – ¥60M (downside)
5x5 matrix at full coverage: 5 verticals × 5 surfaces = 25 product touchpoints
IPO / acquisition / private dividend: BDFL decides based on personal preference + Y4 review
Year-by-year gates retire here; subsequent decision-making is "annual continuation review" only

4. Quarterly review template¶

Every 3 months (Y1Q1, Y1Q2, ..., Y5Q4), review against:

quarter: 2027Q1   # example
date: 2027-04-30
arr_actual: ¥X.XM      # Stripe metered MRR × 12, last-30-day rolling
arr_lower_bound_for_quarter: ¥XM   # from year's lower-bound trip
arr_target_midpoint:        ¥XM   # from year's target range

scenario_classification: best / base / downside   # which 30% / 50% / 20% bucket are we in?

moat_metrics:
  amendment_snapshot_count:     N    # target: ≥ 14,596 + (months_since_launch × 800)
  hallucination_guard_count:    N    # target: monotonically increasing
  testimonials_published:       N    # target: ≥ Y1=10, Y3=50, Y5=200 (linear interp)
  case_studies_published:       N
  compliance_doc_completeness:  %

verticals_live: [v1_subsidy]   # [v1_subsidy, v2_healthcare, v3_realestate, v4_english, v5_kr_zh]
surfaces_live: [rest, mcp, line, widget, affiliate]

deviations:                  # any "off plan" decisions taken this quarter
  - desc: ...
    rationale: ...

next_quarter_actions:
  - ...                     # ≤ 5 actions, prioritized

This template is the ONLY mandatory recurring write — no other status reports, no monthly KPI emails, no weekly stand-ups (memory feedback_zero_touch_solo).

5. What we will NOT do (anti-goals)¶

These are explicit non-objectives — not "later", but "no":

✗ Paid advertising / SEM / display ads
✗ Outbound sales / cold email / cold calls
✗ Tier-based SaaS pricing (Free / Starter / Pro). The only free path is the anonymous 50/月 IP rate limit.
✗ DPA / MSA negotiations. Standard ToS is the only contract.
✗ Slack Connect / dedicated CS channels / phone support
✗ Onboarding calls
✗ Hiring sales reps
✗ Hiring a CS team
✗ White-label resale at "enterprise" tier (custom whitelabel widget at ¥30k/月 is fine, but no per-customer code branches)
✗ Ingesting from aggregator sites (noukaweb, hojyokin-portal, biz.stayway) — past 詐欺 incidents
✗ Reviving the "jpintel" brand (商標衝突 risk with Intel — project_jpintel_trademark_intel_risk)
✗ AnthropicAPI calls in our own server-side code path (per-request cost economics break — feedback_autonomath_no_api_use). Customer-side LLM only.

6. What "post-launch" means in this strategy¶

V2 Healthcare, V3 Real Estate, V4 English, V5 Korean/Chinese, and 5 partner integrations (freee / Money Forward / kintone / SmartHR / Claude Desktop) are post-launch P5-P6 items. They appear in the 5-year plan because:

The 5x5 matrix is plausible — schemas are pre-built (am_alias.language column, record_kind enum extension at migration 013) and src/jpintel_mcp/ is already vertical-agnostic.
Y2-Y4 best-case ARR depends on at least V2 landing.
Future readers (auditors, partners, acquirers) need to see the trajectory written down before they ask.

But none of V2-V5 are launch gates for 2026-05-06. The launch gate is V1 only — see analysis_wave18/audit_full/00_smart_merge_plan.md Section 6.

7. Honesty constraints (cross-link)¶

All ARR projections are probability-weighted ranges, never single-point promises (memory feedback_no_fake_data).
All claims of "X months catch-up" assume a competent + well-funded competitor; not a reassurance that we're safe (we are, in this period).
Operator cost ceiling = solo + ¥0 paid acquisition. Any deviation must be a written quarterly decision, not a default growth path.
This document supersedes _internal/strategy_*.md predecessors. Keep CHANGELOG current when major scenario probabilities shift.

8. References¶

docs/pricing.md — ¥3/req business model
evals/gold.yaml — 79-query precision baseline
docs/per_tool_precision.md — per-tool gate table
analysis_wave18/audit_full/00_smart_merge_plan.md (private) — launch plan + post-launch P5-P6
CLAUDE.md — non-negotiable constraints (¥3/req metered, no tiers, organic only, solo)

Last updated: 2026-04-25 (P4-F documentation pass, audit_full v8). Next mandatory review: 2027-04 (Y1 close).