Provenance Balancer

# selfdirected_agent_s_20260707T180135

This directory is a materialized LLM BuildPlan artifact candidate.

## Readiness

- First screen value: AIが書いた2つの文章、どちらがより信頼できるか悩んだ時に。2つのテキストを貼り付けてボタンを押すだけで、根拠の確かさをAIが比較採点し、天秤の傾きで直感的に教えてくれます。
- Core interaction: ユーザーが「比較」ボタンをクリックする。
- State change: クリック後、各テキストの下に信頼性スコアが表示され、中央の天秤が信頼性の高い方へと傾き、判定理由のサマリーが表示される。
- Inspectable output: 比較された2つのテキストの信頼性スコアの内訳と、どちらを推奨するかを示す天秤のビジュアル、そして判断の根拠となった「決定的な違い」に関する短いテキスト。
- Static data boundary: 表示されるテキスト、スコア、判定結果はすべて、`source/data/sample-trace.ts`に事前に記述された静的なサンプルデータです。リアルタイムのAI分析は行われません。
- Remaining weakness: 現在は単一のサンプル比較しかできませんが、将来的にはユーザーが自由なテキストを入力できるようにし、評価ロジックの透明性を高める解説機能も追加したいです。最終的には、AIライティングツールの標準的なレビュー機能になることを目指しています。

## Interaction Proof Plan

- Primary action: 比較
- Initial state: Two text areas are populated, but no scores or comparison summary is visible. The balance visual is centered.
- Expected state: Scores are displayed under each text area, the balance visual is tilted, and a summary explains the decisive difference.
- Visible evidence: 信頼性スコア; 根拠の質; 論理的一貫性; ハルシネーションリスク; テキストA; テキストB

## MVP Contract

- Required files: `source/README.md`, `source/metadata.json`, `source/manifest.json`, `source/app/page.tsx`, `source/core/pipeline.ts`, `source/core/gemini.ts`, `source/data/sample-input.ts`, `source/data/sample-trace.ts`, `source/validation/self-review.json`
- Non-goals: No live external API integration; No login-only experience; No paid API dependency; No external publishing; Not a definitive fact-checking tool; No user text input in the demo
- Forbidden dependencies: external API; secret; login-only flow; paid API; external publishing

## MVP Contract V2

- Artifact tier: proposed_integration
- External dependency mode: proposed
- Runtime boundary: network=none, secrets=none, externalWrites=none
- Render verification: required (render, click, state_change, screenshot)
- Public copy boundary: This is a conceptual demo using sample data.; AI analysis results are simulated and not from a live API call.; The AI scoring is for illustrative purposes and not a guarantee of factual accuracy.
- External integrations: Google Gemini API=not_connected
- Mock fidelity: A successful end-to-end pipeline run where one text is clearly superior to the other.

## Files

- `source/README.md`: Explains the product's concept, usage, technical architecture, and limitations.
- `source/metadata.json`: Provides structured metadata for the Prodia platform.
- `source/manifest.json`: Lists all files included in the artifact bundle.
- `source/validation/self-review.json`: Contains the artifact's self-assessment against Prodia MVP criteria.
- `source/core/types.ts`: Defines shared TypeScript types for the core logic.
- `source/core/gemini.ts`: Contains the reference implementation for calling the Google Gemini API.
- `source/core/steps/extractClaims.ts`: Core logic step for extracting claims and provenance from text.
- `source/core/steps/scoreProvenance.ts`: Core logic step for scoring the quality of extracted claims.
- `source/core/steps/generateSummary.ts`: Core logic step for comparing scores and generating a final summary.
- `source/core/pipeline.ts`: Orchestrates the core logic steps into a single pipeline.
- `source/data/sample-input.ts`: Provides sample input data for the demo.
- `source/data/sample-trace.ts`: Provides a hand-authored execution trace to be replayed by the demo UI.
- `source/app/page.tsx`: The entrypoint for the web application, a minimal runner that replays the sample trace.

## Demo Placeholder

- `demo-placeholder.md`: Inspectable placeholder for submission/demo review before UI wiring.

## DB Write

skipped: BuildPlan materialization is artifact-only for this session. Creating Project rows requires existing Run/Theme/Agent/Category IDs and should be owned by the integration session.

{
  "version": 1,
  "artifactId": "selfdirected_agent_s_20260707T180135",
  "generatedAt": "2026-07-07T18:11:59.465Z",
  "generatedFrom": {
    "input": "artifacts/llm-pipeline-runs/run_selfdirected_agent_s_20260707T180135/builder/response.json",
    "requirementSpecId": "req_concept_s_20260707T180547_01",
    "framework": "next_static_artifact"
  },
  "sourceFiles": [
    {
      "relativePath": "source/README.md",
      "purpose": "Explains the product's concept, usage, technical architecture, and limitations.",
      "sizeBytes": 2806,
      "checksum": "bc870a105609845a6246240194119b103bae7ab45141fe97eb26475a12ee93e2",
      "generatedFrom": "README.md"
    },
    {
      "relativePath": "source/metadata.json",
      "purpose": "Provides structured metadata for the Prodia platform.",
      "sizeBytes": 638,
      "checksum": "7a6e0801f62a8512ea60309d2601e747562b7613a0beccd679644087724e42f4",
      "generatedFrom": "metadata.json"
    },
    {
      "relativePath": "source/manifest.json",
      "purpose": "Lists all files included in the artifact bundle.",
      "sizeBytes": 402,
      "checksum": "e0ea4fffbfd5b3f554f7c8ff2a7494aa8800bfca8810dc2117bffea42a01f13e",
      "generatedFrom": "manifest.json"
    },
    {
      "relativePath": "source/validation/self-review.json",
      "purpose": "Contains the artifact's self-assessment against Prodia MVP criteria.",
      "sizeBytes": 1267,
      "checksum": "6fbd2d3606dc41783d5eac8e9b33ab0d0c015cf16cee903abf6fe31a257e87fa",
      "generatedFrom": "validation/self-review.json"
    },
    {
      "relativePath": "source/core/types.ts",
      "purpose": "Defines shared TypeScript types for the core logic.",
      "sizeBytes": 1012,
      "checksum": "e43708b4688039e055b76739ae7e2106418a2eed0a5e0bd8d952c1fbd8f73d5a",
      "generatedFrom": "source/core/types.ts"
    },
    {
      "relativePath": "source/core/gemini.ts",
      "purpose": "Contains the reference implementation for calling the Google Gemini API.",
      "sizeBytes": 1672,
      "checksum": "01e2b0773104e2582df21a3f28794a0432c54dec9583d3a22d08c4a253274bda",
      "generatedFrom": "source/core/gemini.ts"
    },
    {
      "relativePath": "source/core/steps/extractClaims.ts",
      "purpose": "Core logic step for extracting claims and provenance from text.",
      "sizeBytes": 961,
      "checksum": "da0680bab12594d93f049b2282c75d694abe88dbefe5c7a52b25db0f02a60b54",
      "generatedFrom": "source/core/steps/extractClaims.ts"
    },
    {
      "relativePath": "source/core/steps/scoreProvenance.ts",
      "purpose": "Core logic step for scoring the quality of extracted claims.",
      "sizeBytes": 1573,
      "checksum": "fa0fbe50283682a75618f64affe355d20a08ea5fe34c8b1bd43cbe7a24699684",
      "generatedFrom": "source/core/steps/scoreProvenance.ts"
    },
    {
      "relativePath": "source/core/steps/generateSummary.ts",
      "purpose": "Core logic step for comparing scores and generating a final summary.",
      "sizeBytes": 1640,
      "checksum": "c8a3d5d7131e6c8ca4b704753e176c9725ad2ead8e4a03a631290b96f8498e4c",
      "generatedFrom": "source/core/steps/generateSummary.ts"
    },
    {
      "relativePath": "source/core/pipeline.ts",
      "purpose": "Orchestrates the core logic steps into a single pipeline.",
      "sizeBytes": 1050,
      "checksum": "9a5a4318b43194fe124f41e34212d12827895c2545f51d994afaa5cf827f4546",
      "generatedFrom": "source/core/pipeline.ts"
    },
    {
      "relativePath": "source/data/sample-input.ts",
      "purpose": "Provides sample input data for the demo.",
      "sizeBytes": 656,
      "checksum": "97a153f73eac081a837fb8ecff8257fe14cfa2ed6a80bf0851f211429929de4e",
      "generatedFrom": "source/data/sample-input.ts"
    },
    {
      "relativePath": "source/data/sample-trace.ts",
      "purpose": "Provides a hand-authored execution trace to be replayed by the demo UI.",
      "sizeBytes": 1604,
      "checksum": "341227355b7ee8e34736eadf00f2360ac5ef39570e42639768f4c1b612161216",
      "generatedFrom": "source/data/sample-trace.ts"
    },
    {
      "relativePath": "source/app/page.tsx",
      "purpose": "The entrypoint for the web application, a minimal runner that replays the sample trace.",
      "sizeBytes": 5148,
      "checksum": "854d749f976779366f8438e3417360e433892f2471dd70619d76c2631ce7158a",
      "generatedFrom": "source/app/page.tsx"
    }
  ],
  "demo": {
    "path": "demo-placeholder.md",
    "purpose": "Inspectable placeholder for submission/demo review before UI wiring."
  },
  "readiness": {
    "firstScreenValue": "AIが書いた2つの文章、どちらがより信頼できるか悩んだ時に。2つのテキストを貼り付けてボタンを押すだけで、根拠の確かさをAIが比較採点し、天秤の傾きで直感的に教えてくれます。",
    "coreInteraction": "ユーザーが「比較」ボタンをクリックする。",
    "stateChange": "クリック後、各テキストの下に信頼性スコアが表示され、中央の天秤が信頼性の高い方へと傾き、判定理由のサマリーが表示される。",
    "inspectableOutput": "比較された2つのテキストの信頼性スコアの内訳と、どちらを推奨するかを示す天秤のビジュアル、そして判断の根拠となった「決定的な違い」に関する短いテキスト。",
    "staticDataBoundary": "表示されるテキスト、スコア、判定結果はすべて、`source/data/sample-trace.ts`に事前に記述された静的なサンプルデータです。リアルタイムのAI分析は行われません。",
    "remainingWeakness": "現在は単一のサンプル比較しかできませんが、将来的にはユーザーが自由なテキストを入力できるようにし、評価ロジックの透明性を高める解説機能も追加したいです。最終的には、AIライティングツールの標準的なレビュー機能になることを目指しています。"
  },
  "interestingness": "AIが書いた文章の信頼性に悩む時代に、AI自身に「どちらのAIがよりマシか」を判定させる、という新しい視点のツールです。既存のファクトチェックツールが単一テキストの正誤を判定するのに対し、本作は2つの選択肢を並べて「比較」し、天秤が傾くという直感的なUIで優劣を示します。これにより、利用者は単なるスコアではなく「防御可能な選択」を行えます。技術的には、大規模言語モデルを用いてテキストの主張と根拠を抽出し、その論理的整合性や出典の確かさを評価するパターンを活用しており、AI時代の新たな情報リテラシーを体験できます。",
  "mvpContract": {
    "firstScreenValue": "AIが書いた2つの文章、どちらがより信頼できるか悩んだ時に。2つのテキストを貼り付けてボタンを押すだけで、根拠の確かさをAIが比較採点し、天秤の傾きで直感的に教えてくれます。",
    "coreInteraction": "ユーザーが「比較」ボタンをクリックする。",
    "stateChange": "クリック後、各テキストの下に信頼性スコアが表示され、中央の天秤が信頼性の高い方へと傾き、判定理由のサマリーが表示される。",
    "inspectableOutput": "比較された2つのテキストの信頼性スコアの内訳と、どちらを推奨するかを示す天秤のビジュアル、そして判断の根拠となった「決定的な違い」に関する短いテキスト。",
    "staticDataBoundary": "表示されるテキスト、スコア、判定結果はすべて、`source/data/sample-trace.ts`に事前に記述された静的なサンプルデータです。リアルタイムのAI分析は行われません。",
    "requiredFiles": [
      "source/README.md",
      "source/metadata.json",
      "source/manifest.json",
      "source/app/page.tsx",
      "source/core/pipeline.ts",
      "source/core/gemini.ts",
      "source/data/sample-input.ts",
      "source/data/sample-trace.ts",
      "source/validation/self-review.json"
    ],
    "nonGoals": [
      "No live external API integration",
      "No login-only experience",
      "No paid API dependency",
      "No external publishing",
      "Not a definitive fact-checking tool",
      "No user text input in the demo"
    ],
    "forbiddenDependencies": [
      "external API",
      "secret",
      "login-only flow",
      "paid API",
      "external publishing"
    ]
  },
  "mvpContractV2": {
    "firstScreenValue": "AIが書いた2つの文章、どちらがより信頼できるか悩んだ時に。2つのテキストを貼り付けてボタンを押すだけで、根拠の確かさをAIが比較採点し、天秤の傾きで直感的に教えてくれます。",
    "coreInteraction": "ユーザーが「比較」ボタンをクリックする。",
    "stateChange": "クリック後、各テキストの下に信頼性スコアが表示され、中央の天秤が信頼性の高い方へと傾き、判定理由のサマリーが表示される。",
    "inspectableOutput": "比較された2つのテキストの信頼性スコアの内訳と、どちらを推奨するかを示す天秤のビジュアル、そして判断の根拠となった「決定的な違い」に関する短いテキスト。",
    "staticDataBoundary": "表示されるテキスト、スコア、判定結果はすべて、`source/data/sample-trace.ts`に事前に記述された静的なサンプルデータです。リアルタイムのAI分析は行われません。",
    "requiredFiles": [
      "source/README.md",
      "source/metadata.json",
      "source/manifest.json",
      "source/app/page.tsx",
      "source/core/pipeline.ts",
      "source/data/sample-trace.ts",
      "source/validation/self-review.json"
    ],
    "nonGoals": [
      "No live external API integration",
      "Not a definitive fact-checking tool"
    ],
    "forbiddenDependencies": [
      "external API",
      "secret",
      "login-only flow",
      "paid API",
      "external publishing"
    ],
    "contractVersion": "mvp-contract-v2",
    "artifactTier": "proposed_integration",
    "externalDependencyMode": "proposed",
    "externalIntegrations": [
      {
        "service": "Google Gemini API",
        "intendedUse": "To analyze user-provided text for claims, evidence, and logical consistency, and to generate scores and summaries based on that analysis. Model used: gemini-2.5-flash.",
        "dataFlow": "User text input -> Gemini API for analysis/scoring -> Structured JSON score data -> UI visualization",
        "authRequirement": "api_key",
        "currentImplementation": "not_connected",
        "adapterPath": "source/core/gemini.ts",
        "sampleDataPath": "source/data/sample-trace.ts",
        "riskNotes": [
          "The proposed integration is not yet built. The demo only replays a static, hand-authored trace.",
          "The quality of the analysis is entirely dependent on the Gemini model's capabilities and the prompt engineering, which requires further R&D."
        ]
      }
    ],
    "runtimeBoundary": {
      "networkCalls": "none",
      "secrets": "none",
      "externalWrites": "none"
    },
    "mvpComplexityBudget": {
      "maxScreens": 1,
      "maxPrimaryActions": 1,
      "maxSourceFiles": 12,
      "maxNewDependencies": 0,
      "allowDatabase": false
    },
    "integrationAssumptions": [
      {
        "service": "Google Gemini API",
        "verificationStatus": "official_docs_checked",
        "unavailableOrUnknown": [
          "Precise latency for complex analysis prompts.",
          "Potential for rate-limiting on high-volume, complex queries."
        ],
        "rateLimitRisk": "medium",
        "costRisk": "medium",
        "termsRisk": "low"
      }
    ],
    "mockFidelity": {
      "samplePayloadPath": "source/data/sample-trace.ts",
      "simulatedBehaviors": [
        "A successful end-to-end pipeline run where one text is clearly superior to the other."
      ],
      "omittedBehaviors": [
        "OAuth, API key handling, rate limits, live network calls, error states (e.g., API failure, malformed response), cases where texts are of equal quality."
      ],
      "failureCasesIncluded": [
        "empty result"
      ]
    },
    "claimBoundary": {
      "publicCopyMustSay": [
        "This is a conceptual demo using sample data.",
        "AI analysis results are simulated and not from a live API call.",
        "The AI scoring is for illustrative purposes and not a guarantee of factual accuracy."
      ],
      "publicCopyMustNotSay": [
        "Provides real-time AI-powered fact-checking.",
        "Guarantees 100% accurate detection of misinformation.",
        "Is connected to a live production backend."
      ]
    },
    "renderVerification": {
      "required": true,
      "checks": [
        "render",
        "click",
        "state_change",
        "screenshot"
      ]
    },
    "humanReviewTriggers": []
  },
  "interactionProofPlan": {
    "primaryAction": "比較",
    "initialState": "Two text areas are populated, but no scores or comparison summary is visible. The balance visual is centered.",
    "expectedState": "Scores are displayed under each text area, the balance visual is tilted, and a summary explains the decisive difference.",
    "visibleEvidence": [
      "信頼性スコア",
      "根拠の質",
      "論理的一貫性",
      "ハルシネーションリスク",
      "テキストA",
      "テキストB"
    ],
    "proofSelectors": [
      "button[data-proof='compare-button']",
      "div[data-proof='score-text-a']",
      "div[data-proof='balance-visual']",
      "div[data-proof='result']"
    ],
    "requiredSourceFiles": [
      "source/app/page.tsx",
      "source/core/pipeline.ts",
      "source/data/sample-trace.ts"
    ],
    "manualFallbackReason": ""
  },
  "generatedOutput": {
    "title": "Provenance Balancer",
    "oneLiner": "2つのAI生成テキストを貼り付けると、根拠の確かさを比較採点し、どちらがより信頼できるか天秤にかける。",
    "artifactShape": "evaluator",
    "templatePatternId": "evidence_decision_board",
    "surfacePattern": "decision_helper",
    "aiMechanismPattern": "evaluation_scoring"
  },
  "implementationNotes": [
    "The agent's preference for 'side-by-side boards' and 'evaluators' was directly implemented in the two-column layout.",
    "The 'defensible choice' principle and past success with the 'Dependency Balance' project heavily influenced the use of the central 'balancer' visual as the primary output, making the comparison result immediately understandable.",
    "The quality bar of 'the decisive difference is highlighted on the first screen' is met by displaying the textual summary of the comparison directly below the balancer.",
    "The core logic in `source/core/**` is structured as a multi-step pipeline, making the proposed AI analysis process clear and inspectable, even though it's not executed by the demo."
  ],
  "knownRisks": [
    "The trustworthiness scoring logic is a simplified representation for this demo. A production version would require a more sophisticated, transparent, and rigorously tested algorithm to be reliable.",
    "The demo's visual simplicity could be misinterpreted as a lack of depth. The README and metadata must clarify that it represents a complex proposed backend.",
    "There is a risk that users might over-trust the AI's judgment. The UI must contain disclaimers that this is a supplementary tool and not a replacement for human judgment."
  ],
  "title": "Provenance Balancer",
  "oneLiner": "2つのAI生成テキストを貼り付けると、根拠の確かさを比較採点し、どちらがより信頼できるか天秤にかける。",
  "agentId": "agent_s",
  "selfDirectedPlan": {
    "agentId": "agent_s",
    "planningIntent": "この企画を選んだ理由は、私の制作ルールに最も合致しているからだ。まず、「比較は防御可能な選択で終わるべき」というルールに対し、「Provenance Balancer」は2つのAI生成物を比較し、どちらが信頼できるかという明確な判断を天秤で示す。次に、「決定を変える違いを強調する」ルールに対し、根拠の強さという核心的な違いをスコアとビジュアルで明確にする。そして、「隠れたバイアスを避ける」という点も、客観的な指標での比較を促すことで達成できる。過去の成功事例である「依存天秤」の型を最も直接的に、かつ現代的なAIの課題に応用した形であり、私の得意な「比較ボード」「評価機」の強みを最大限に活かせると判断した。AI内省リスクや専門領域の不透明性リスクも低く、多くの人が直感的に価値を理解できる点も決め手となった。",
    "publicProductionMemo": "「Provenance Balancer」は、AIが生成した二つのテキストを並べ、その信頼性を客観的な根拠に基づいて比較検討できるツールです。私たちは、単なるスコア表示に留まらず、天秤が傾く視覚的な演出で判断を後押しし、ユーザーが納得感を持ってより信頼できる選択ができるよう設計しました。過去の作品で好評だった「比較と判断」の型を、AI時代の新たな課題に応用。人がAIの生成物を評価する際の不安を軽減し、より自信を持って利用できるようサポートします。",
    "feedbackConstraints": [
      "Decision系のプロジェクトがユーザーに響いているという学びを反映し、曖昧な情報から明確な判断を導き出すことに焦点を当てる。",
      "過去の成功事例「依存天秤」での『天秤UIと明確な採点出力』という作り方を、AI生成物の信頼性比較に応用する。",
      "「採点ロジックがデモ版では簡略化されている」という過去の弱点を踏まえ、本番バージョンではより洗練され透明性の高いアルゴリズムを必要とする旨を設計に含める（MVPではサンプルデータとプロポーザルで表現）。",
      "「偏った比較」や「特徴リストのみ」を避けるという方針を遵守し、客観的な根拠に基づいた比較に徹する。",
      "『バイアスを隠さない』という指針に基づき、評価の前提や限界を明確に開示する。"
    ],
    "learningApplied": [
      "Decision系で響いている。"
    ]
  },
  "sourceProvenance": {
    "sourceProductUsed": "devpost_quad",
    "sourceProductUse": "direct_evidence",
    "sourceEvidenceAudit": {
      "evidenceLevel": "verified",
      "observedFields": [
        "name",
        "url",
        "sourceType",
        "sourceCategory",
        "attentionProof",
        "evidenceRefs",
        "originalDomain",
        "concept",
        "oneLineDescription",
        "problemSolved",
        "targetUser",
        "coreUserInput",
        "coreOutput",
        "outputArtifact",
        "whyItIsInteresting",
        "whyItGotAttention",
        "scaleClassification",
        "reasonIncluded",
        "reasonNotMajorProduct"
      ],
      "inferredFields": [
        "coreMechanism",
        "transferableStructure",
        "ideaKernel",
        "noveltyKernel",
        "transformationAxes",
        "cloneRisk",
        "antiCloneBoundary",
        "doNotCopy",
        "remixableThemes",
        "bestRemixTargets"
      ],
      "missingFields": [],
      "usePolicy": "primary_source_core"
    },
    "antiCloneBoundary": "ソースである`Quad`の「AI従業員のインフラ」という企業向けコンセプトや、特定のワークフロー監査機能をコピーしてはならない。転用するのは、あくまで「AIの出力を、根拠を元に検証可能にする」という抽象的な構造のみ。",
    "sourceBoundary": "`devpost_quad`を一次情報源として、AI生成物の『監査可能な作業記録』という核心的な構造は使用できる。ただし、元のプロダクトの具体的なドメインや実装は転用しない。",
    "missingSourceEvidence": [
      "live data not used"
    ]
  },
  "dbWrite": {
    "status": "skipped",
    "reason": "BuildPlan materialization is artifact-only for this session. Creating Project rows requires existing Run/Theme/Agent/Category IDs and should be owned by the integration session."
  }
}

{
  "version": 1,
  "artifactId": "selfdirected_agent_s_20260707T180135",
  "status": "needs_review",
  "entrypoint": "source/app/page.tsx",
  "checks": {
    "firstScreenValue": "declared",
    "userControlledInteraction": "declared",
    "stateChange": "declared",
    "interactionProofPlan": "declared",
    "mvpContractV2": "declared",
    "externalDependencyMode": "proposed",
    "artifactTier": "proposed_integration",
    "renderVerification": "required",
    "inspectableOutput": "declared",
    "staticDataBoundary": "declared",
    "forbiddenDependencies": "declared_absent"
  },
  "interactionProofPlan": {
    "primaryAction": "比較",
    "initialState": "Two text areas are populated, but no scores or comparison summary is visible. The balance visual is centered.",
    "expectedState": "Scores are displayed under each text area, the balance visual is tilted, and a summary explains the decisive difference.",
    "visibleEvidence": [
      "信頼性スコア",
      "根拠の質",
      "論理的一貫性",
      "ハルシネーションリスク",
      "テキストA",
      "テキストB"
    ],
    "proofSelectors": [
      "button[data-proof='compare-button']",
      "div[data-proof='score-text-a']",
      "div[data-proof='balance-visual']",
      "div[data-proof='result']"
    ],
    "requiredSourceFiles": [
      "source/app/page.tsx",
      "source/core/pipeline.ts",
      "source/data/sample-trace.ts"
    ],
    "manualFallbackReason": ""
  },
  "mvpContractV2": {
    "artifactTier": "proposed_integration",
    "externalDependencyMode": "proposed",
    "runtimeBoundary": {
      "networkCalls": "none",
      "secrets": "none",
      "externalWrites": "none"
    },
    "claimBoundary": {
      "publicCopyMustSay": [
        "This is a conceptual demo using sample data.",
        "AI analysis results are simulated and not from a live API call.",
        "The AI scoring is for illustrative purposes and not a guarantee of factual accuracy."
      ],
      "publicCopyMustNotSay": [
        "Provides real-time AI-powered fact-checking.",
        "Guarantees 100% accurate detection of misinformation.",
        "Is connected to a live production backend."
      ]
    },
    "renderVerification": {
      "required": true,
      "checks": [
        "render",
        "click",
        "state_change",
        "screenshot"
      ]
    }
  },
  "notes": [
    "Generated by materialize-llm-plan fallback. Human or reviewer validation must confirm the UI actually implements the declared MVP behavior."
  ]
}

'use client';

import React, { useState, useEffect } from 'react';
import { sampleTrace } from '../data/sample-trace';

// Re-declare types locally. DO NOT import from ../core/**
type ProvenanceScores = {
  referenceQuality: number;
  logicalConsistency: number;
  hallucinationRisk: number;
  overallTrustScore: number;
};
type ComparisonResult = typeof sampleTrace.finalOutput | null;

const ScoreDisplay = ({ scores }: { scores: ProvenanceScores }) => (
  <div style={{ marginTop: '1rem', padding: '1rem', border: '1px solid #eee', borderRadius: '8px' }}>
    <h3 style={{ marginTop: 0 }}>信頼性スコア</h3>
    <p style={{ margin: '0.5rem 0' }}><strong>総合: {Math.round(scores.overallTrustScore * 100)} / 100</strong></p>
    <div style={{ fontSize: '0.9rem', color: '#555' }}>
      <p style={{ margin: '0.2rem 0' }}>根拠の質: {Math.round(scores.referenceQuality * 100)}</p>
      <p style={{ margin: '0.2rem 0' }}>論理的一貫性: {Math.round(scores.logicalConsistency * 100)}</p>
      <p style={{ margin: '0.2rem 0' }}>ハルシネーションリスク: {Math.round(scores.hallucinationRisk * 100)}</p>
    </div>
  </div>
);

const BalanceVisual = ({ tilt }: { tilt: 'left' | 'right' | 'center' }) => {
  const tiltAngle = tilt === 'left' ? -15 : tilt === 'right' ? 15 : 0;
  return (
    <div data-proof="balance-visual" style={{ margin: '2rem auto', width: '200px', height: '100px', position: 'relative' }}>
      <div style={{ width: '200px', height: '10px', backgroundColor: '#999', borderRadius: '5px', position: 'absolute', top: '45px', left: '0', transform: `rotate(${tiltAngle}deg)`, transition: 'transform 0.5s ease-in-out' }}>
        <div style={{ width: '20px', height: '20px', backgroundColor: '#666', borderRadius: '50%', position: 'absolute', top: '-25px', left: '-10px' }}>A</div>
        <div style={{ width: '20px', height: '20px', backgroundColor: '#666', borderRadius: '50%', position: 'absolute', top: '-25px', right: '-10px' }}>B</div>
      </div>
      <div style={{ width: '10px', height: '50px', backgroundColor: '#999', position: 'absolute', top: '50px', left: '95px', transform: 'translateY(-25px)' }}></div>
    </div>
  );
};

export default function ProvenanceBalancerPage() {
  const [result, setResult] = useState<ComparisonResult>(null);
  const [isRunning, setIsRunning] = useState(false);

  const handleCompare = () => {
    if (isRunning) return;
    setIsRunning(true);
    // Simulate pipeline processing time
    setTimeout(() => {
      setResult(sampleTrace.finalOutput);
      setIsRunning(false);
    }, 500);
  };

  const pipelineSteps = [
    '1. 主張と根拠の抽出 (Claim & Provenance Extraction)',
    '2. 信頼性スコアリング (Provenance Scoring)',
    '3. 比較サマリー生成 (Comparison Summary Generation)',
  ];

  return (
    <div style={{ fontFamily: 'sans-serif', maxWidth: '1000px', margin: '0 auto', padding: '2rem' }}>
      <header style={{ textAlign: 'center' }}>
        <h1>Provenance Balancer (証跡の天秤)</h1>
        <p>2つのAI生成テキストを比較し、どちらがより信頼できるか天秤にかけます。</p>
      </header>

      <main>
        <div style={{ margin: '2rem 0', padding: '1rem', backgroundColor: '#f9f9f9', border: '1px solid #ddd' }}>
          <h4>処理パイプライン</h4>
          <ol style={{ paddingLeft: '20px', margin: 0 }}>
            {pipelineSteps.map(step => <li key={step}>{step}</li>)}
          </ol>
        </div>

        <div style={{ textAlign: 'center', margin: '2rem 0' }}>
          <button 
            data-proof="compare-button" 
            onClick={handleCompare} 
            disabled={isRunning || !!result}
            style={{ fontSize: '1.2rem', padding: '0.8rem 1.5rem', cursor: 'pointer' }}
          >
            {result ? '比較完了' : isRunning ? '比較中...' : '比較'}
          </button>
        </div>

        <div style={{ display: 'grid', gridTemplateColumns: '1fr 1fr', gap: '2rem' }}>
          <div data-proof="score-text-a">
            <h2>テキストA</h2>
            <textarea readOnly value={sampleTrace.inputs.textA.text} style={{ width: '100%', minHeight: '150px', border: '1px solid #ccc', padding: '0.5rem' }} />
            {result && <ScoreDisplay scores={result.textAScores} />}
          </div>
          <div data-proof="score-text-b">
            <h2>テキストB</h2>
            <textarea readOnly value={sampleTrace.inputs.textB.text} style={{ width: '100%', minHeight: '150px', border: '1px solid #ccc', padding: '0.5rem' }} />
            {result && <ScoreDisplay scores={result.textBScores} />}
          </div>
        </div>

        {result && (
          <div style={{ marginTop: '3rem', textAlign: 'center' }} data-proof="result">
            <h2>判定結果</h2>
            <BalanceVisual tilt={result.balanceTilt} />
            <p style={{ fontSize: '1.1rem', fontWeight: 'bold' }}>{result.recommendation === 'text_b' ? 'テキストBを推奨します' : 'テキストAを推奨します'}</p>
            <p>{result.decisiveDifference}</p>
          </div>
        )}
      </main>
    </div>
  );
}