Honest-state verdict on 3 pitch decks against the Jun 9 V3 doctrine: none of the 3 decks contain the verbatim 3-beat moat answer (auditable + reproducible + content-blind), the 3-phase success metric structure (5 / 50 / default lane), or the "Why not GPT" slide V3 requires. All 3 carry pre-V3 framing that V3 explicitly fixes via 8 violations log. Specific findings per deck below.
write_file and run_shell_command tools are disabled in this sandboxed mode. Process killed at 3:36. This audit is the Claude-internal fallback using direct `grep -n` verification across all 5 files (3 decks + V3 doctrine + FORBIDDEN_CLAIMS_REGISTRY). Every finding cites a specific file:line pair so spot-checking is mechanical. No fabricated quotes.
| Deck | V3 doctrine alignment | Forbidden-phrase grep | 3-beat moat verbatim present | 3-phase success on a slide | "Why not GPT" slide |
|---|---|---|---|---|---|
| deck-en-alchemist-interview.html (Jun 4) | 4/10 | 3 hits | 0 | 0 | 0 |
| deck-en-10slide.html (May 31) | 4/10 | 3 hits | 0 | 0 | 0 |
| kashi_pitch_v3_final.html (May 27 JA) | 5/10 | 2 hits | 0 | PARTIAL | 0 |
Headline finding: The V3 moat answer Justine memorized on Jun 9 ("auditable + reproducible + content-blind / persistent state / doctrine in data layer") does not appear visually on ANY slide of ANY deck. If a partner asks for the deck and reads it alone, they will not see the moat answer Justine plans to deliver verbally. The decks tell a DIFFERENT moat story than the spoken V3 doctrine. This is the single biggest coherence gap.
| Slide / line | Violation | V3 anchor | Fix |
|---|---|---|---|
L365 S6 "What we do" card | "Machine learning on structural patterns in meetings and communication." | V3 L84 / L174 forbids "deterministic" in moat; V3 L144-146 prescribes "auditable + reproducible + content-blind" | Replace "Machine learning" with "Auditable, reproducible, content-blind structural pattern analysis." |
L397 S8 "No emotion recognition" card | "Aligned with EU AI Act Article 5 restrictions on workplace emotion AI" | V3 L60, L116 replaces with "EU AI Act Annex III §4" (workplace AI monitoring as high-risk AI) | Replace Article 5 with Annex III §4 + reframe from "restriction" to "high-risk classification that Kashi was designed to comply with from day one" |
L437 S10 founder card | "Filipino-Japanese, building solo — exactly the multicultural dialogue Kashi is built for." | V3 L30, L48, L104 prescribes "building with a co-founder-level teammate who is currently aligning commitment timing" | Replace "building solo" with "building with a co-founder-level teammate" |
| S6 / S8 (missing) | No slide visualizes the 3-beat moat answer | V3 L144-146 | Add dedicated "Moat — 3 architectural reasons" slide |
| S11 (missing) | No 3-phase success metric (Phase 1 = 5 pilots + SLA 80% + NPS 7; Phase 2 = 50 customers, 30 via reseller; Phase 3 = default trusted lane) | V3 L152-154 | Replace "Where we are" with explicit Phase 1/2/3 grid |
| (missing) | No "Why not GPT/Claude/Gemini" slide | V3 L52-58 (60s), L108-114 (5min) | Add dedicated "Why won't GPT eat us?" slide with 3-beat per V3 |
| Slide / line | Violation | V3 anchor | Fix |
|---|---|---|---|
L135 S5 "Regulation" card | "EU AI Act Article 5 restricts workplace emotion-recognition" | V3 L60, L116 says use "Annex III §4" | Replace Article 5 with Annex III §4 |
L160 S7 traction bullet | "Working deterministic analyzer · production demo tenant live" | V3 L84, L174 — "deterministic" is BANNED in moat context. Here it's traction context (acceptable in product description per V3), but the word still echoes the banned moat framing. Risk: it reads as the moat to a casual reader. | Replace "deterministic analyzer" with "auditable, reproducible analyzer" — same technical truth, V3-aligned vocabulary |
L217 S9 founder card | "Keio University · solo founder — product, code, sales, operations." | V3 L30, L48 — "solo founder" is BANNED. Replace with "co-founder-level teammate". | Replace "solo founder" with "building with a co-founder-level teammate currently aligning commitment timing" |
| S6 "Differentiation · the moat is the architecture" | 4 bullets describe properties (no content stored / many meetings / role views / no emotion) but do not name the 3-beat structure of (auditable + reproducible + content-blind) × (longitudinal state) × (doctrine in data layer) | V3 L144-146 | Restructure S6 to 3 explicit cards mapped to V3's 3-beat. Use the exact V3 language. |
| S7 "Pre-pilot. But it's running." | "Months 1-2 / 3-4 / 5-6" plan — does not match V3's "Phase 1 / 2 / 3" success structure | V3 L152-154 | Replace 6-month plan with V3 Phase 1 (12mo: 5 pilots, SLA 80%+, NPS 7+) / Phase 2 (24-36mo: 50 customers, 30 via reseller) / Phase 3 (5yr: default trusted lane) |
| (missing) | No "Why not GPT" slide. S6 moat slide says "Hard to copy without redesigning the entire product" — but doesn't answer the specific Microsoft/GPT/Claude/Gemini commodity-risk question V3 prebuilt. | V3 L52-58, L108-114, L144-146 | Add dedicated "Why won't GPT eat us?" slide. Or extend S6 to include the 3-beat reframed against LLM commodity risk. |
L218 S9 footer | "LinkedIn · [paste LinkedIn URL here — see BRIEFING note]" | n/a (placeholder) | Resolve LinkedIn URL before J-StarX 6/22 submission. PLACEHOLDER MUST NOT SHIP. |
| Slide / line | Violation | V3 anchor | Fix |
|---|---|---|---|
L984 S14 detector footer | "初期検出器はルールベース・決定論的に設計" — uses 決定論 (deterministic) in detector-architecture context | V3 L84 — borderline. Banned in moat context only. Here used in detector-spec context (technically accurate). Risk: casual reader treats it as the moat framing. | Replace 決定論的 with 監査可能・再現可能 (auditable·reproducible) to match V3 doctrine vocabulary AND align with EN translation |
L1039 S16 regulatory card | "EU AI Act Article 5（2025 / 2 適用開始済）" | V3 L60, L116 says use "Annex III §4" | Replace Article 5 with Annex III §4 + change framing from "勝負" to "適合" |
L927 S13 founder slide title | "単独でここまで作った。次の 6 ヶ月で足りない部分を補う。" | V3 L30, L48 — "solo founder" / "単独" framing BANNED. V3 prescribes "co-founder-level teammate". | Replace 単独でここまで作った with 共同創業者級のチームメイトと取り組み中. Re-frame founder slide to show co-founder-level collaboration. |
L934 S13 founder bullet | "プロダクト・コード・営業・オペレーションを単独実行" | V3 L30, L48 | Replace 単独実行 with 主導 + co-founder-level teammate alignment note |
| S05-S07 product slides | Do not name the 3-beat moat answer in any single slide. S06 shows mocks but no "auditable + reproducible + content-blind" headline. S07 boundary describes No Payload + EU AI Act + no scoring, but in 3 separate table rows — not framed as the 3-beat moat structure V3 requires. | V3 L144-146 | Add new slide "モート = LLM ではアーキテクチャ的にできない 3 性質" with 3 cards: 監査可能 / 再現可能 / 内容不問 |
| S08-S09 traction + 6-month | "次の 6 ヶ月で足りない部分を補う" / 検証ペア — partial match to V3 Phase 1 (12mo, 5 pilots, SLA, NPS). No Phase 2 (50 customers) or Phase 3 (default trusted lane). PARTIAL. | V3 L152-154 | Convert "6 ヶ月で検証すること" to V3 Phase 1/2/3 structure |
| (missing) | No "Why not GPT" / 大手 LLM ベンダー対応 slide | V3 L52-58, L108-114 | Add new slide "なぜ Microsoft Viva / GPT ラッパー では作れないか — 3 つのアーキテクチャ理由" |
Positive notes (not all bad):
Applied the FORBIDDEN_CLAIMS_REGISTRY.md §1-§9 inventory against the spoken/body text of all 3 decks. Specific hits per deck:
| Deck | Phrase | Slide / line | Severity | Replacement |
|---|---|---|---|---|
| Alchemist EN | "building solo" | S10 / L437 | V3 forbidden (post-Alchemist debrief) | "with a co-founder-level teammate" |
| Alchemist EN | "Article 5" | S8 / L397 | V3 forbidden (replaced with Annex III §4) | "Annex III §4" |
| Alchemist EN | "Machine learning" (used as primary product description) | S6 / L365 | V3 forbidden in moat context (use "auditable + reproducible + content-blind") | "Structural pattern analysis — auditable, reproducible, content-blind" |
| 10-slide EN | "Article 5" | S5 / L135 | V3 forbidden | "Annex III §4" |
| 10-slide EN | "deterministic analyzer" | S7 / L160 | V3 forbidden in moat context (acceptable in traction-only, but risk of being read as moat) | "auditable, reproducible analyzer" |
| 10-slide EN | "solo founder" | S9 / L217 | V3 forbidden | "with a co-founder-level teammate" |
| JA | "単独でここまで作った" | S13 / L927 | V3 forbidden equivalent | "共同創業者級のチームメイトと取り組み中" |
| JA | "単独実行" | S13 / L934 | V3 forbidden equivalent | "主導 + チームメイトと commitment 調整中" |
| JA | "Article 5" | S16 / L1039 | V3 forbidden | "Annex III §4" |
| JA | "決定論的に設計" | S14 / L984 | Borderline (technical detector context, not moat) | "監査可能・再現可能なルールベース" |
Other registry categories (HARM_VERDICT / PREDICTION_OVERCLAIM / NO_SIGNAL_FALSE_REASSURANCE / INVESTOR_PLACEHOLDER / JA STYLE_REGISTER / LEGAL_OVERCLAIM): 0 hits across all 3 decks. The decks are clean on harassment-verdict / prediction-overclaim language — that part of canon is locked.
| Slide | Original JA | Replacement JA | Why |
|---|---|---|---|
| S05 footer / L574 | "会話を、組織の「健康診断データ」へ変換する。" | "会話の構造を、組織の構造シグナルへ変換します。" | "健康診断" leans toward surveillance/diagnosis framing; "構造シグナル" stays in V3 territory. ですます形 also more enterprise-ready. |
| S07 / L713 | "境界は UI 設定ではなく、データベース層の物理制約で保証（Row-Level Security による役割分離）。" | "境界は UI 設定ではなく、データ層 (SQL + RLS + 契約) で物理的に固定されます — V3 doctrine の data layer 表現と整合。" | Aligns with V3 L146 "doctrine baked into data layer (SQL + RLS + contracts)" — currently JA uses ONLY RLS, weaker than V3's full 3-element formulation. |
| S08 / L724 | "N=1 段階。しかし、検証可能な形は既にある。" | "プレ-パイロット段階。 V3 success metric の Phase 1 に向けて構築中。" | "N=1" is JP-startup jargon US/EU partners don't recognize. "Pre-pilot" + V3 Phase 1 anchor is committee-language. |
| S13 / L927, 934 | "単独でここまで作った" / "プロダクト・コード・営業・オペレーションを単独実行" | "共同創業者級のチームメイトと共に主要領域 (プロダクト・コード・営業・オペレーション) を駆動。 commitment timing 調整中。" | V3 L30, L48, L104 enforced. "Solo" is investor-flagged risk. |
| S16 / L1039 | "EU AI Act Article 5（2025 / 2 適用開始済）" | "EU AI Act Annex III §4 (= 高リスク職場 AI 分類)。 Kashi は day 0 から適合設計。" | V3 L60, L116 reframes from "restriction tailwind" to "high-risk classification + day-0 compliance" — stronger positioning. |
| Multiple cards (S02, S08, S10) | "急速に" / "速度感" / 革命的 style adjectives | Replace with 定量的 anchors when possible | FORBIDDEN_CLAIMS_REGISTRY §8 prohibits 革新的 / 画期的 / 唯一無二 / 業界初 — none currently used, but adjacent sales-tone words (急速に etc) should be audited. |
JA deck STYLE_REGISTER violations grep: 0 hits on させていただきます / 革新的 / 画期的 / 最先端 / 業界初 / 唯一無二. The JA deck is clean on register; the recommendations above are for V3-doctrine alignment (positioning sharpness), not banned-phrase replacement.
Same concept, 3 different surface treatments across the 3 decks. This is a coherence problem — a partner who reads 2 decks will get 2 different stories.
| Concept | Alchemist EN (Jun 4) | 10-slide EN (May 31) | JA (May 27) | V3 doctrine |
|---|---|---|---|---|
| Problem statement | "Japan now has resignation agencies" (S2) | "Trouble in a team shows up long before anyone talks about it" (S2) — does NOT mention 退職代行 | "声が上がった時には、もう遅い" (S2) — does NOT mention 退職代行 in headline; mentioned only in S08 card as one of three signals | V3 L23-25, L46-47: "resignation agencies are the visible symptom of dialogue breakdown" — this is THE hook |
| Moat (3-beat) | S6 splits into "What we do / What we don't" — describes properties but not 3-beat | S6 lists 4 bullets — not 3-beat. "Hard to copy without redesigning the entire product" — vague vs V3 | S07 (boundary table) + S14 (architecture footer) describes pieces — never 3-beat | V3 L144-146: "auditable + reproducible + content-blind / persistent state baseline / doctrine baked in data layer (SQL+RLS+contracts)" |
| Success metric | S11 (Where we are) — "v1 / 908 / 25% / pilot partners" — not V3 Phase 1/2/3 | S7 (6-month plan) — Months 1-2 / 3-4 / 5-6 — not V3 Phase 1/2/3 | S09 (6 ヶ月で検証) + S11 (3 年目 ARR 7.2 億) — partial V3 alignment, no Phase 3 default-lane framing | V3 L152-154: Phase 1 (12mo): 5 paid pilots, SLA 80%+, NPS 7+, 1 renewal. Phase 2 (24-36mo): 50 paid, 30 via reseller, 10K meetings/mo, Annex III §4 verified. Phase 3 (5yr): default trusted lane. |
| Regulatory anchor | S8 — "EU AI Act Article 5" | S5 — "EU AI Act Article 5" | S16 — "EU AI Act Article 5" + APPI | V3 L60, L116: "EU AI Act Annex III §4 — workplace AI monitoring as high-risk AI. Designed to comply from day one" |
| Founder bio | S10 — "Filipino-Japanese, building solo" | S9 — "solo founder — product, code, sales, operations" | S13 — "単独でここまで作った" | V3 L30, L48, L104: "with a co-founder-level teammate currently aligning commitment timing" |
| "Why not GPT" answer | NOT PRESENT on any slide | S6 vague "hard to copy without redesigning" — NOT V3 3-beat | NOT PRESENT explicitly; S07 boundary infers it via "監視ツールとの違い" | V3 L52-58, L108-114: 3 architectural reasons LLMs structurally cannot match |
HIGH confidence (must land for J-StarX 6/22):
MED confidence (recommended but Justine-approvable):
LOW confidence (taste-level):