Deng Killer

对话中的事实护盾：记录对话、抽取可核验事实主张，并以低打扰方式提示可能错误的陈述。

当前仓库是一个 monorepo，包含 iPhone 前端和后端核验服务 scaffold。

Structure

.
├── PRD.md
├── frontend/       # SwiftUI iPhone MVP
└── backend/        # FastAPI claim verification scaffold

Frontend

The iPhone MVP lives in frontend/.

Open the Xcode project:

cd frontend
open DengKiller.xcodeproj

Regenerate the Xcode project after editing frontend/project.yml:

cd frontend
tools/xcodegen/xcodegen/bin/xcodegen generate

Run core tests:

cd frontend
swift test

Run full iOS tests with an available simulator:

cd frontend
xcodebuild test -project DengKiller.xcodeproj -scheme DengKillerApp -destination 'platform=iOS Simulator,name=iPhone 17,OS=26.5' -derivedDataPath .build/DerivedData CODE_SIGNING_ALLOWED=NO

Backend

The backend scaffold lives in backend/. It provides a small FastAPI service for single-claim verification and keeps the PRD privacy boundary explicit:

• GET /health
• POST /v1/claims/verify
• Pydantic schemas aligned with the Swift VerificationRequest and VerificationResult models.
• A dual-mode verification pipeline: deterministic fixtures for local/CI and Brave Search plus OpenAI evidence judgement when deployment keys are set.
• Domain routing and source allowlists for technical, business, and civic facts.
• Optional structured source citations in verification results.
• Privacy guard that rejects full audio, full transcripts, full conversations, messages, and speaker history.

Set up the backend with uv:

cd backend
uv sync --extra dev

Run the API:

cd backend
uv run uvicorn app.main:app --reload

Run backend tests:

cd backend
uv run pytest

Local development loads backend/.env; production reads only injected environment variables. Real retrieval is enabled when BRAVE_SEARCH_API_KEY and OPENAI_API_KEY are configured with VERIFICATION_MODE=auto; Jina Reader then converts allowlisted public pages to clean Markdown before evidence judgement. Set VERIFICATION_MODE=fixture for deterministic offline behavior. The backend does not include persistent storage, accounts, or full conversation upload. See backend/README.md for configuration, architecture, and privacy rules.

TODO

Frontend work remaining from PRD.md:

• Polish real audio capture with volume/activity handling and more robust interruption recovery.
• Evaluate an on-device Whisper-style ASR option after the Apple Speech MVP.
• Improve sentence finalization beyond punctuation and Apple Speech final signals, especially for long pauses and multi-sentence ASR revisions.
• Keep full audio, full transcripts, session review, and verification results local to the phone by default.
• Add local deletion controls for single sessions and an optional auto-delete period.
• Improve low-interruption notifications: lightweight real-time state for high confidence/high importance claims, plus richer after-conversation review.
• Add UI support for evidence summaries, uncertainty explanations, and sources once the backend returns them.

Backend work remaining from PRD.md:

• Replace the rule-based mock verifier with real claim extraction, checkworthiness scoring, and initial classification.
• Extend source allowlists and golden evaluations beyond the initial technical, business, and civic coverage, especially for more regions and languages.
• Add lightweight RAG for curated stable sources alongside search-based retrieval.
• Add streaming verification progress, such as received, classifying, retrieving, judging_evidence, and completed.
• Add feedback and opt-in error-example endpoints without uploading full conversations by default.
• Add model-version metadata to the existing privacy-safe operational logs.
• Add optional storage for curated public sources, public evidence cache, verification metadata, and user-authorized high-confidence error examples.
• Add high-risk domain safeguards for political, medical, legal, and investment claims: evidence and limitations only, no final decision advice.

Privacy Boundary

Deng Killer must default to protecting the user's full conversation:

• Do not upload full audio by default.
• Do not upload full transcripts by default.
• Do not upload speaker history or full conversation context by default.
• Backend verification should receive only one selected claim plus minimal necessary context.
• Logs and analytics must not contain private transcript or audio content.