Alert overload
Evidence arrives faster than analysts can connect it.
NTU InnovateX 2026 / Track 2
Trust the evidence.
Control the consequence.
A permissioned AI workforce for security decisions that remain explainable, human-governed and tamper-evident.
[ scroll to inspect ]
[ 01 ]
Autonomy, with evidence.
Scroll-driven identity01 / The problem
Security teams do not need another black box.
They need help with the volume—without surrendering control of the consequence.
Opaque judgment
A convincing answer can still be unsupported or inconsistent.
Unsafe autonomy
A wrong remediation can create a second incident.
Mutable history
A database record can be rewritten after the decision.
02 / The system
Bounded intelligence
Reasoning where it helps.
Rules where it matters.
Models can recommend. Deterministic policy and people decide what is allowed.
Policy gate01Detection
02Correlation
03Investigation
04Triage
05Remediation
06Verifier
Explain the evidence
Typed stages expose outputs, citations and confidence.
Constrain the action
Structural checks remain authoritative after the model finishes.
Hold the switch
Dangerous, uncertain or degraded outcomes require approval.
Prove the record
Evidence and output digests are anchored on Sepolia.
03 / How it works
One incident.
Four layers of trust.
Raw evidence stays private. Each layer adds context or control, but no layer inherits the authority of the next.
Evidence
GUIDE rows become a bounded, label-free context.
Agents
Six specialists reason in a fixed, inspectable order.
Authority
Verifier, policy and human approval govern the outcome.
Proof
Only hashes, identities, approvals and timestamps go on-chain.
04 / Controlled interaction
Run the incident.
Watch authority hold.
Validated static replay
Detection
AI reasoningSeverity 0.78 · suspicious device and account surfaced.
Correlation
Deterministic control26 alerts collapsed into 18 reproducible clusters.
Investigation
Grounded reasoningSimilar cases and graph paths added without exposing the label.
Triage
AI judgmentTruePositive selected with cited evidence and rationale.
Remediation
Simulated actionMonitor and watchlist recommended with rollback plan.
Verifier
Independent challengeUncertainty triggers escalation; autonomy remains blocked.
Tamper evidence
The operator controls the database.
Not the anchored digest.
Verification recomputes both hashes from the underlying rows. It never trusts a saved “valid” flag.
WAITING
Evidence
0xc0e05df5fc…2789c307—Agent output
0x12e19e6002…23def7f3e—05 / Measured evidence
Built, measured
and deliberately honest.
Real numbers on Microsoft GUIDE data—not projections and not metrics computed over the hand-picked presentation cases.
incidents
Microsoft GUIDE working set
evidence rows
processed into inspectable context
alert reduction
870 alerts → 276 clusters
rehearsal checks
validated replay sweep
automated tests
420 Python + 29 Solidity
on-chain identities
registered on Sepolia
The claim we will not make
The agents do not currently beat the classifier.
Measured live-model macro F1 was 0.4084 versus the LightGBM baseline at 0.4669.
The value is controlled reasoning.
Explanation, deterministic correlation, policy, human approval and tamper evidence.
Replay is a safety feature.
Live variance was measured. The pitch uses validated, byte-identical responses.
06 / Technical documentation
Inspect the boundaries.
The system is designed so its most important claims can be checked independently.
01Agent contracts
+
Six typed stages use schema validation, grounding checks, one bounded retry and visible degradation.
Detection → Correlation → Investigation → Triage → Remediation → Verifier02Retrieval
+
BM25 and 384-dimensional vectors are fused with Reciprocal Rank Fusion. Queries require no API call.
5,000 indexed documents · RRF k=6003Graph correlation
+
Union-Find creates deterministic alert clusters; capped BFS and confidence-weighted Dijkstra add context.
96,351 nodes · 26,194 edges04Policy engine
+
Catalogue, risk, citation, verifier and degraded-stage checks determine whether autonomy is allowed.
High risk, uncertainty or degradation → human05Proof lifecycle
+
Canonical JSON is hashed with keccak256. The contract records identity, approval and final state.
Proposed → Approved → Finalized06Privacy boundary
+
Raw evidence, prompts, rationales and comments remain in SQLite; only digests and metadata leave it.
Sensitive evidence never goes on-chain07Failure handling
+
A failed stage degrades visibly to deterministic output and forces human review.
Replay miss → visible fallback → approval requiredNetworkSepoliachainId 11155111
AgentSphere SecureOps
Autonomous security needs
enforceable accountability.
Reason broadly. Act narrowly. Prove what happened.
Repository
View AgentSphere ↗
Source and documentation
Contact
nikhilranjanmurmu75@gmail.com
Submission contact