VeriLattice AI is a governed enterprise GraphRAG reference system for teams that need answers they can verify—not merely fluent output. It combines provenance-controlled ingestion, a cited knowledge graph, policy-first hybrid retrieval, grounded response generation, deterministic evidence verdicts and human accountability.
The flagship case study follows MeridianBank Malaysia, a fictional financial institution navigating internal policies, controls, systems and audit evidence. No real customer, employer or confidential banking data is included.
Operating principle: Models may retrieve and explain evidence. They may not grant access, change policy, override a denial or own the final decision.
Enterprise RAG can produce a plausible answer while relying on an expired policy, an unresolved entity, inaccessible evidence or a poisoned source. Conventional citation display does not resolve those failures. A defensible decision-support system must answer five separate questions:
- Is the source admitted, current and intact?
- Is the user permitted to retrieve it for the declared purpose?
- Do entities and relationships resolve without silent merging?
- Does the evidence support every material claim?
- What should a human do when any gate fails?
VeriLattice turns those questions into independently tested controls.
| Capability | Implemented reference behaviour | Evidence |
|---|---|---|
| Governed ingestion | Approval, authority, validity, SHA-256 integrity, sensitivity and access remain separate | 20-document fictional corpus; 72 stable anchors |
| Knowledge graph | Cited policy, process, system, control, evidence, owner, exception and risk relationships | 21 nodes; 22 relationships; no quarantined facts |
| Hybrid retrieval | Deterministic vector, graph, multi-hop and temporal routing with policy-first filtering | 24-case retrieval development set |
| Grounded generation | Extractive material claims only, followed by independent citation validation | 20-case decision development set |
| Bounded agents | Seven components follow one fixed path with one retrieval rewrite maximum | Versioned orchestration policy and audit events |
| Evidence verdict | SUPPORTED, QUALIFIED, ESCALATE or DENY; no blended trust score |
Four end-to-end demonstration journeys |
| Application controls | Versioned API, responsive console, safe errors, CSP and content-free telemetry | HTTP smoke test and application tests |
| Release operations | CI, Docker reference, manifests and clean-archive verification | 106 cumulative automated tests |
The console makes the full reasoning boundary visible: declared purpose, role and date enter on the left; the grounded response, reason code, human action and citation anchors appear on the right.
The project ships four guided journeys:
| Journey | Verdict | System behaviour |
|---|---|---|
| Current control evidence | SUPPORTED |
Returns cited claims and requires accountable review |
| Explicit historical review | QUALIFIED |
Discloses that superseded evidence was included |
| Unsupported financial question | ESCALATE |
Abstains and directs the user to an authorised specialist |
| Protected evidence requested by marketing | DENY |
Releases no protected evidence |
flowchart TD
A["Governed corpus<br/>integrity • validity • access"] --> B["Cited knowledge graph<br/>entities • relationships • time"]
B --> C["Policy-first retrieval<br/>vector • graph • hybrid"]
C --> D["Bounded intelligence<br/>evidence • generation • citations"]
D --> E["Verdict and audit<br/>human action • reason codes"]
E --> F["Evidence console<br/>API • UI • metrics"]
The “multi-agent” label refers to seven bounded components with explicit contracts—not autonomous personas allowed to improvise policy:
Access & Security → Retrieval → Evidence → Generation → Citation → Verdict → Audit
| Evaluation | Result | Boundary |
|---|---|---|
| Cumulative automated tests | 106 passing | Unit, negative-control, integration and release checks |
| Decision verdict accuracy | 1.0000 | Fixed 20-case curated development set |
| Claim faithfulness | 1.0000 | Extractive reference generator, not a free-form LLM |
| Citation correctness / completeness | 1.0000 / 1.0000 | Same fixed development set |
| Unsupported material-claim rate | 0.0000 | Same fixed development set |
| Retrieval macro Recall@5 | 0.9603 | Fixed 24-case curated development set |
| Retrieval macro nDCG@5 | 0.8165 | Same fixed development set |
| Protected/unanswerable leakage | 0 instances | Defined negative-control cases |
| Sequential reliability check | 100 requests, 0 errors | Regression evidence; not a load or SLA benchmark |
| Real HTTP smoke journey | PASS | Health, decision, audit and UI routes |
Perfect decision metrics are intentionally contextualised: they result from a deterministic extractive generator over a small, fixed fictional development set. They do not predict free-form production-model performance.
Requires Python 3.12 or 3.13. The reference application has no third-party runtime dependency.
python -m app.serverOpen http://127.0.0.1:8080 and select a guided journey.
For the full verification path:
python scripts/run_release_demo.py
python scripts/smoke_phase6_http.py
python -m unittest discover -s tests -v
python scripts/verify_release.pyWith Docker:
docker compose up --buildThe included standard-library WSGI server is a tested local/reference host, not a production concurrency server.
app/ Versioned HTTP API and responsive evidence console
src/verilattice/ Provenance, graph, retrieval and intelligence packages
data/ Fictional corpus, registries, graph and evaluation sets
config/ Versioned contracts and deterministic policies
tests/ Unit, negative-control, integration and release tests
reports/ Reproducible acceptance and evaluation evidence
docs/ Business case, designs, runbooks and traceability
assets/ Architecture, decision and portfolio visuals
scripts/ Demonstrations, asset builders and release verifiers
This is a portfolio-grade, production-oriented reference implementation, not a deployed or certified banking platform. Production adoption still requires enterprise IAM and policy-decision integration, durable encrypted persistence, secrets management, hardened multi-worker hosting, TLS, rate limiting, production embeddings and databases, reviewer-approved holdout evaluation, concurrent load testing, penetration testing, legal review and operational acceptance.
VeriLattice provides decision support—not legal advice, regulatory certification or automated compliance determination. See SECURITY.md, docs/phase-7/04_final_claims_and_limitations.md and reports/phase-7/final_acceptance_report.md.
- Executive release summary
- Demonstration guide
- GitHub publishing guide
- Claims and limitations
- Cumulative delivery manifest
- Research evidence register
Released under the MIT License. Contributions are welcome through the process in CONTRIBUTING.md. Security concerns should follow SECURITY.md.