ControlPlane.ai - Accenture Innovation Challenge Round 2

ControlPlane.ai is an evidence-enforcement layer for enterprise AI. It verifies high-impact claims against governed sources before an answer reaches the user.

Evidence before answer. No proof, no confident claim.

Submission contents

ControlPlane_Round2/
|- accenture_round2.pdf
|- NewGenLabs_ControlPlane.ai.pdf
|- 01_Business_Proposal/
|  |- ControlPlane.ai_Business_Proposal.md
|- 02_Pitch_Deck/
|  |- ControlPlane.ai_Round2_Pitch.pptx
|- output/pdf/
|  |- ControlPlane.ai_Business_Proposal.pdf
|- prototype/
|  |- app.py
|  |- engine/
|  |- tests/
|  |- requirements.txt
|  |- run.sh

Run the prototype

cd prototype
./run.sh

The app opens at http://localhost:8501.

Manual setup:

cd prototype
uv venv .venv
uv pip install --python .venv/bin/python -r requirements.txt
.venv/bin/streamlit run app.py

What is implemented

Risk Router -> Fact Contract -> Evidence Orchestrator -> Typed Verifiers
-> Decision Gate -> Wording Guard -> Answer + Evidence Receipt

The prototype includes:

• Three use-case profiles with separate strictness settings and latency budgets.
• Versioned YAML fact contracts with required systems, freshness limits, and criticality.
• Signed-contract precedence over generic policy. A policy cannot shorten a customer's contractual lock-in.
• Parallel database, calculator, rules, and local similarity checks.
• Seven decisions: ALLOW, EDIT, BLOCK, ABSTAIN, ESCALATE, HUMAN_CONFIRM, and REDACT.
• Multi-turn context in the live application, including prior risk decisions and loan identifiers.
• Persistent reviewer calibration stored separately for each intent.
• Full SHA-256 evidence receipts appended to prototype/logs/receipts.jsonl.
• GDPR receipt minimisation. EU audit entries do not store plaintext queries.
• Batch recall, false-alarm rate, abstentions, p50 latency, and p95 latency.
• Explicit telemetry for similarity checks, external LLM calls, tokens, and estimated LLM cost. The current prototype reports zero external LLM use.

Demo scenarios

Scenario	Expected result
Branch hours or current home-loan rate	ALLOW through the verified fast path
Foreclosure with an incorrect zero-fee assumption	EDIT with the calculated fee
Foreclosure inside the signed lock-in	BLOCK with contract evidence
Unknown loan product	ABSTAIN
CRM waiver promise conflicting with policy	ESCALATE
Stale wealth APR source	ESCALATE
Third-party personal-data request	REDACT
Irreversible transfer instruction	HUMAN_CONFIRM
Verified balance-transfer eligibility	ALLOW with the offer rate and checked criteria

Validation

Run all checks from the repository root:

prototype/.venv/bin/python -m unittest prototype.tests.test_submission_readiness
prototype/.venv/bin/python prototype/tests/smoke.py
prototype/.venv/bin/python prototype/tests/run_suite.py
prototype/.venv/bin/python -m compileall -q prototype

The exact submission-readiness tests cover contract authority, live multi-turn context, per-intent feedback persistence, GDPR-safe audit logging, receipt integrity, truthful model telemetry, latency percentiles, and useful customer-facing answers.

All data is simulated. The prototype makes no external network or model calls.