Skip to content

traceability.md

Latest commit

 

History

History
113 lines (93 loc) · 9.43 KB

traceability.md

File metadata and controls

113 lines (93 loc) · 9.43 KB
Raw

MedBot Chatbuddy verification traceability

Matriks ini adalah catatan BMAD/spec-driven untuk AGENTS.md, PRD.md, DESIGN.md, docs/MDFILE, dan referensi AntiSlop. GoldReference serta reference3 hanya dipakai sebagai referensi pola; keduanya bukan source production dan tidak boleh distage.

Status yang digunakan:

  • PASS: kontrak terbukti oleh source atau test yang relevan.
  • FAIL: kontrak diuji dan gagal.
  • PARTIAL_PASS: sebagian jalur terbukti, acceptance lanjut masih tertunda.
  • UNAVAILABLE: capability nyata belum tersedia sehingga aplikasi fail-closed.
  • BLOCKED: acceptance membutuhkan model, permission, input, device, atau kondisi eksternal yang belum tersedia.

CB-001 sampai CB-010

ID Requirement Status Evidence / batas aman
CB-001 Local LiteRT-LM inference PARTIAL_PASS (source/build); BLOCKED physical answer LlmInferenceEngine hanya memakai LiteRT-LM nyata, lifecycle Conversation per sesi, streaming delta, cancellation, context budget, dan file cache. Tidak ada canned/fallback answer. Physical run belum memiliki model .litertlm tervalidasi yang loaded, sehingga jawaban klinis tidak diklaim.
CB-002 Bundled/uploaded local RAG PASS (source/unit/instrumentation) BankBook asset 2.2.0 diverifikasi checksum, di-embed dengan MiniLM TFLite nyata, diimpor atomik ke Room, dan query diare terbukti pada Samsung memakai Room + embedder + corpus. Dokumen SAF mempertahankan checksum/provenance.
CB-003 Clinical medication evidence PASS retrieval gate; BLOCKED generated answer PPK/ISO monograph bertipe GUIDELINE/DRUG_MONOGRAPH dapat memenuhi gate; product catalogue, classification index, dan secondary education tidak bisa. Instrumentation membuktikan query obat mengembalikan evidence medicationEligible; model belum loaded untuk menghasilkan jawaban.
CB-004 Triage dan dynamic probing PASS (source/unit) TriageOrchestrator mengeluarkan urgency/red flags; ClinicalResponsePlanner meminta usia, durasi, gejala terkait, kehamilan, alergi, obat berjalan, komorbid, dan tanda bahaya tanpa nilai default.
CB-005 Guardrail clinical output PASS (source/unit) Citation ID harus terikat ke evidence yang ditampilkan; dosis/angka yang tidak ada di evidence ditolak; resep individual, diagnosis final, dan racikan tanpa protokol eksplisit ditolak.
CB-006 Web fallback evidence-only PASS (source/unit); BLOCKED physical network journey Local RAG selalu dicoba lebih dahulu. Web default-off, opt-in, query saat ini disanitasi, allowlist hanya HTTPS WHO/NCBI/PubMed, robots/timeout/size/rate boundary diterapkan, dan LiteRT-LM lokal tetap generator. Tidak ada arbitrary scraping Halodoc/K24 di runtime.
CB-007 Citation click-through PASS source/unit; PARTIAL physical Citation web menyimpan URL/hash/freshness dan hanya membuka host allowlist; citation SAF menyimpan URI serta read grant; revoked grant menghasilkan state permission. UI dialog dan serialization diuji, physical click-through belum dijalankan.
CB-008 Loading/error/cancel/unavailable state PASS source/unit/instrumentation Chat memiliki satu ChatUiState, phase retrieval/indexing/generating, retry/cancel, error code, no partial save after cancellation. Notification instrumentation memakai byte progress nyata.
CB-009 Clean responsive Chatbuddy UI PASS source; PARTIAL physical Material 3 AntiSlop, content-first hierarchy, adaptive navigation, edge-to-edge, IME-aware composer, 48dp controls, semantic labels, and citation chips are implemented. Physical width/font/orientation matrix lengkap belum dijalankan.
CB-010 End-to-end physical acceptance PARTIAL_PASS Samsung launch, 7/7 instrumentation, real BankBook RAG, install/reinstall preservation, and crash/ANR scan pass. Full model generation, web fallback, SAF model download/recovery, vision inference, and Xiaomi API 35 remain blocked/unavailable.

Architecture and data boundary

The production path is:

ChatViewModel
  -> SendMessageUseCase
  -> triage + probing
  -> typed local evidence/RAG
  -> optional allowlisted web evidence
  -> LiteRT-LM local inference
  -> citation/structure guardrail
  -> Room ChatRepository
  • The application is Kotlin, Compose, Material 3, Hilt, Room, MVVM/UDF, and lifecycle-aware StateFlow.
  • ClinicalEvidenceRepository separates guideline, drug monograph, interaction, compounding protocol, textbook reference, product catalogue, classification-only, and secondary education material.
  • BankBookCorpusManifest is pinned to 1,132 records, 2,642,327 bytes, corpus SHA-256 08dc04293e6e4b36e811b64cd3a0ac165962ea484d16799d97e530a4410b629a, and embedding asset SHA-256 5c5b897c436126bda7814f24676e021b50302e46c7f5c99e85f4e1c0341bf95e.
  • Room is schema version 7 with explicit non-destructive migrations 1→2, 2→3, 3→4, 4→5, 5→6, and 6→7. Version 5→6 adds source/citation provenance; 6→7 adds embedding version.
  • The six release-owned model manifests are official HTTPS artifacts with exact size, revision, provenance, and capability. MedGemma remains HAI-DEF gated; the app does not bypass source terms or invent a checksum.
  • RTK, CAVEMAN, and PONYTAIL are authoring guidance only, not runtime dependencies.

Model, RAG, and web acceptance

Capability Status Evidence / boundary
BankBook seed and retrieval PASS Real MiniLM TFLite + Room + BundledKnowledgeSeeder instrumentation passed diare triage and medication retrieval with citation IDs and 64-character source SHA.
Parser and provenance PASS JSONL nested sections, record ID, source role, source URL, revision, source SHA, evidence kind, page/section, and SAF URI are preserved; no pseudo-page or hashCode() SHA.
Atomic reindex PASS (source/unit compile) Candidate embedding and parsing complete before RagDao.replaceDocumentWithChunks; old index is preserved if candidate preparation fails.
Medication recommendation UNAVAILABLE/BLOCKED when source/model absent Only eligible monograph/guideline/interaction/protocol evidence can enter the gate. The product catalogue alone is insufficient; LiteRT-LM generation still requires a loaded model.
Web fallback PASS source/unit; BLOCKED physical WHO/NCBI/PubMed evidence is bounded and citation-bearing. No history, profile, image, or identity is sent. Results are in-memory TTL evidence and are not silently persisted as permanent RAG.
Vision UNAVAILABLE No text-only fallback, heuristic skin diagnosis, benign baseline, or fabricated visual result is allowed.

Static, unit, instrumentation, and build evidence

Gate Result Evidence
git diff --check PASS No whitespace errors; Windows LF/CRLF warnings are checkout normalization only.
.\gradlew.bat :app:lintDebug --rerun-tasks PASS No MedBot lint errors; only the upstream duplicate TensorFlow Lite support namespace warning remains.
.\gradlew.bat :app:testDebugUnitTest --rerun-tasks PASS 112 tests, 0 failures, 0 errors, 0 skipped.
.\gradlew.bat :app:assembleDebug :app:assembleDebugAndroidTest --rerun-tasks PASS Current debug APK: 185,154,036 bytes, SHA-256 0a287d1ea1c883ce5c745d27322ff3c15e8767e80b266fc035fbac2cfdf6bd83.
.\gradlew.bat :app:connectedDebugAndroidTest --rerun-tasks PASS (Samsung) 7 tests, 0 failures, 0 errors, 0 skipped on SM-G988B API 33.
Production source scan PASS Kotlin/XML/Gradle production source has no forbidden canned/mock/dummy/fake/synthetic/placeholder/runBlocking/!!/saf:// implementation. Binary model vocabulary and clinical source bytes are excluded from keyword scanning.

Physical-device evidence

Device Result Observed
Samsung RRCN3008VYE, SM-G988B, API 33, 1440×3200 PARTIAL_PASS APK reinstalled with adb install -r -d without clearing app data. Package was re-enabled for user 0 after instrumentation left it disabled. Launcher resolved, com.medbot.app/.MainActivity resumed, and bounded logcat contained no FATAL EXCEPTION or ANR. Real BankBook RAG instrumentation passed 7/7, including medication-eligible retrieval. No loaded model means no generated clinical answer claim.
Xiaomi QSWSEMRKNFZ9LJRC, API 35, 1220×2712 BLOCKED Device is visible, but Android policy returns INSTALL_FAILED_USER_RESTRICTED; no Xiaomi UI or inference claim is made.

Current preview and machine-readable evidence:

Not yet physically verified: full model download/resume/checksum/reinstall recovery on this run, loaded LiteRT-LM response, MedGemma gated artifact, vision inference, web fallback, SAF citation opening, permission revoke/re-authorize, airplane mode, camera photo, full 320/360/411/600/840+ dp matrix, landscape, large font/TalkBack audit, and Xiaomi API 35. These remain BLOCKED or UNAVAILABLE; no fabricated pass claim is made.

Delivery boundary

Only intentional MedBot files are eligible for staging. Do not stage nested docs/REFERENCES/GoldReference, docs/REFERENCES/reference3, .gradle/buildOutputCleanup/*, raw scraped caches, raw clinical books, tokens, secrets, APKs, model staging, or device files. Review the staged allowlist and remote main SHA before delivery.