Corroborate.ai is an enterprise-grade knowledge arbitration engine designed to solve the critical problem of LLM hallucinations and unauditable "black box" memory. Unlike traditional memory systems (like Mem0 or Zep) that rely on a single model call to decide what is "true," Corroborate.ai treats truth as a function of context.
This repository contains the Android Reference Client, a high-fidelity implementation that demonstrates the Confidence Auction mechanics, multi-jurisdictional arbitration, and regulatory traceability required for high-stakes AI applications in insurance, legal, and banking sectors.
Current AI memory systems suffer from three single points of failure:
- Arbitration Risk: One LLM call decides facts without a deterministic audit trail.
- Jurisdictional Blindness: No notion of "simultaneous truths" (e.g., a pricing rule valid in Germany but illegal in the US).
- Regulatory Liability: Inability to perform cascading erasures (GDPR Art. 17) or provide signed provenance for EU AI Act compliance.
Corroborate.ai mitigates these risks by replacing LLM-based consolidation with a Deterministic Scorer Ensemble.
- Business Value: Reduced legal liability, guaranteed regulatory compliance, and high-fidelity knowledge retrieval that scales sub-linearly with data volume.
- Target Users: AI Engineers, Compliance Officers, Enterprise Architects.
The Android client interacts with a regionalized cloud stack ensuring strict data residency.
graph TD
User((User/Agent)) -->|Query/Ingest| Mobile[Android Client]
Mobile -->|REST API| Gateway[API Gateway / Auth]
Gateway -->|Tenant Route| Stack[Regional Cloud Stack]
subgraph "Regional Cloud Stack"
Worker[Ingestion Worker]
Auction[Confidence Auction Engine]
Graph[(Neo4j - Claim Graph)]
Vector[(Qdrant - Semantic Index)]
Audit[(Merkle Audit Log)]
end
Mobile -.->|Verify| Auction
Worker -->|Conflict Check| Graph
Worker -->|Embed| Vector
Auction -->|Sign Record| Audit
The following diagram illustrates the request-response lifecycle during a claim resolution.
sequenceDiagram
participant A as Android App
participant G as API Gateway
participant R as Resolution Engine
participant S as Scorer Ensemble
participant L as Merkle Audit Log
A->>G: POST /v1/claims/resolve {query, region, tenant}
G->>R: Route to regional partition
R->>R: Retrieve Semantic Candidates (top-k=20)
R->>S: Compute Scores (Sr, St, Sc, Sa, Sh)
S-->>R: Score Vector [0,1]^n
R->>R: Geometric Mean Fusion
alt Confidence Delta < 0.10
R-->>A: 409 Ambiguous (Return Candidates)
else High Confidence Winner
R->>L: Append Signed Record
R-->>A: 200 OK (Claim + Provenance + AuditRef)
end
Component Responsibilities:
-
Scorer Ensemble: deterministic calculation of Source Reliability (
$S_r$ ), Recency Decay ($S_t$ ), Corroboration ($S_c$ ), and Regional Authority ($S_a$ ). - Contradiction Guard: Logic gate preventing silent resolution of contradicting claims with low confidence deltas.
- Merkle Audit Log: Append-only log for EU AI Act compliance.
Kotlin 2.1.10: Primary language.
Jetpack Compose: Modern declarative UI.
Material 3: Enterprise-grade component library.
Retrofit 2.11.0: Type-safe HTTP client.
Kotlinx Serialization: JSON mapping.
OkHttp 4.12.0: Underlying networking stack.
- Storage: S3 for raw encrypted episodes.
- Database: Neo4j (Graph) for claim edges; Qdrant (Vector) for semantic search.
- Security: Merkle-root anchoring for tamper-evident logs.
app/src/main/java/com/example/corroborate/
├── data/
│ ├── api/ # Retrofit interfaces & Request/Response models
│ ├── model/ # SQL-mapped Kotlin data classes (Tenant, Claim, etc.)
│ └── repository/ # Repository pattern abstractions for UI/API separation
├── ui/
│ ├── theme/ # Material 3 styling, colors, and typography
│ ├── viewmodel/ # MVVM state holders for Resolve, Ingest, and Audit screens
│ └── MainActivity.kt # Root navigation and screen orchestration
├── util/
│ └── ConfidenceScorer.kt # CORE ENGINE: Deterministic scoring formulas
└── util/
- Android Studio Ladybug (2024.2.1) or newer.
- JDK 17 or 21.
- Android SDK 37 (UpsideDownCakePrivacySandbox).
Create a local.properties file in the root directory (never commit this!):
# .env.example equivalent for Android
API_BASE_URL="https://api.example.ai/v1/"
TENANT_ID="example-tenant-id"
MERKLE_LOG_ANCHOR="secure-log-endpoint"- Clone the repository:
git clone https://github.com/alfinohatta/corroborate-ai.git - Open in Android Studio.
- Sync Gradle:
File > Sync Project with Gradle Files. - Build:
Build > Make Project.
The app implements the Erasure Cascade (§3.3) workflow. When a user requests deletion:
- Linked episodes are hard-deleted.
- Claims with independent corroboration are PII-stripped.
- Claims without independent corroboration are hard-deleted.
- AGENT: Basic Resolve/Ingest permissions.
- VERIFIER: Human-in-the-loop verification access (§4.1).
- ADMIN: Configuration and erasure permissions.
We welcome enterprise contributions. Please follow our workflow:
- Fork the repo and create a feature branch.
- Implement your changes (ensure 100% test coverage for
util/ConfidenceScorer). - Submit a PR with a detailed impact analysis on the Auction mechanics.
This project is licensed under the License. See LICENSE for details.
- Q: Why geometric mean over arithmetic mean?
- A: Geometric mean punishes a near-zero score on any dimension (e.g. wrong region) much harder, preventing "wrong region" facts from winning purely on recency.
- Q: How do I trigger the Contradiction Guard?
- A: In the Resolve screen, type 'ambiguous' to simulate a scenario where two claims share a < 0.10 confidence delta.