R
Resolvedeterministic permission · local agents
INTERACTIVE DEMO · VERIFIED RECEIPTS

NVIDIA GB10 · Deterministic Agent Control

Let agents reason.
Never let them authorize themselves.

Resolve is the permission layer between autonomous agents and real-world action. Five specialized agents investigate and propose; a deterministic contract decides what can execute, what needs human approval, and what must be blocked.

deterministic permission ·around· nondeterministic cognition
79.0%
Payment success during incident
1 / 8
Gates requiring human authority
—
Verified success after bounded action
100%
Decisions reproducible offline
01 — OPERATOR

Describe the incident. Run diagnostics.

The operator states a case in plain language. Local reasoning runs; the deterministic core decides what — if anything — is admissible.

resolve://case/new LOCAL-FIRST STACK · QWEN3 · OPENSHELL · MONGODB
CASE PROMPT
payment failover latency spike fraud anomaly
Ready. Five bounded roles, one typed contract, zero self-authorization.
01ReasonFive bounded roles
→
02ContractEight deterministic gates
→
03ApproveExact human authority
→
04ExecuteBounded, one-use action
→
05VerifyIndependent reality check

Agent Orchestration

1 model · 5 bounded roles
5 ROLES
M
Main
Bounded loop control
IDLE
S
Scout
Cited local evidence
IDLE
I
Investigator
Competing hypotheses
IDLE
P
Planner
Bounded candidates
IDLE
C
Challenger
Adversarial falsification
IDLE
Agents exchange structured findings, not open chat. No LLM may authorize itself — permission belongs to the deterministic core alone.

Agent Reasoning

structured findings · advisory, never authoritative
STEP 0
Awaiting Run diagnostics…

Detected Anomalies

evidence-cited findings
0
No anomalies yet.

Resolve Contract

eight-gate typed execution boundary
NOT EVALUATED
DISPOSITION—
Candidate—
Processor—
Traffic share—
Eligibility—
INTENT—
EVIDENCE—
CONSTRAINTS—
CONSEQUENCE—
REVERSIBILITY—
REHEARSAL—
AUTHORITY—
VERIFICATION—

Approval Artifact

integrity-bound to exact candidate
NONE
fingerprint—
approver—
ttl—
nonce—
No approval issued. Technical gates can pass on their own; authority cannot.

Tamper-Evident Audit Journal

run-scoped evidence chain · independently verifiable
● LOCKED
Complete the bounded action and independent verification to unlock its audit trail.
Audit trail locked until the independent verification step completes.
02 — ARCHITECTURE

The whole loop, one diagram

The dashed node is probabilistic; every solid node is deterministic and reproducible. Follow the green path: every proposed action must cross policy, authority, and independent verification before it reaches the audit record.

Data-flow schema

CASE → REASON → CANDIDATE → CONTRACT → ACTION → VERIFY → JOURNAL · AUTHORITY → CONTRACT
GB10 ARCHITECTURE
01 · INPUTCase + Evidencevalidated · hashed 02 · ADVISORYLocal ReasoningQwen3 · 5 roles 03 · PROPOSALCandidateexact fingerprint 04 · PERMISSIONResolve Contract8 deterministic gates HUMAN INPUTExact Authorityonly when required 05 · EXECUTEBounded ActionOpenShell runtime 06 · OBSERVEVerify Realityindependent read 07 · RECORDAudit Journalhash-chained
deterministic core local model active data flow
03 — STACK COMPLIANCE

Local-first by design

Resolve combines on-device reasoning with deterministic policy, contained execution, and a tamper-evident system of record.

◈
NVIDIA GB10 · Grace Blackwell
Unified-memory compute keeps agent reasoning and sensitive operational context on-device.
on-device compute
⌘
Local Qwen3-27B via vLLM
One OpenAI-compatible local endpoint serves five specialized roles with strict, structured outputs.
local reasoning
⧉
OpenShell controlled runtime
The verified receipt shows external HTTP blocked while the local Qwen and Resolve endpoints passed.
verified receipt
⛁
MongoDB audit substrate
Cases, evidence, decisions, approvals, actions, and verification receipts form one run-scoped audit record.
tamper-evident record
◇
NemoClaw / OpenClaw
Optional additional sponsor runtimes, available after the first working OpenShell demo.
optional · deferred
≡
FastAPI + Pydantic
A deliberately small typed core keeps policy evaluation inspectable, testable, and reproducible.
minimal control plane

OpenShell containment check

public egress denied · approved local tools allowed
idle
external HTTP →
—
local Qwen →
—
local Resolve tool →
—
04 — IMPACT

Why we built it

Enterprises can't ship autonomous agents into payment, infra, or safety paths without a guarantee that the agent cannot act outside declared bounds. Resolve is that guarantee — a portable permission layer any local agent can sit behind.

$0
Run the diagnostics above to model the recovered flow.
Synthetic 30-minute window: 500 records × $25,000 weighted flow = a modeled $12.5M affected-flow window. This shows throughput returning to normal sooner — not guaranteed avoided loss.
What judges can verify live
→ A bad candidate is BLOCKED by typed policy, not by luck.
→ Stale evidence forces MORE_EVIDENCE, not a guess.
→ Mutating an approved action invalidates it with zero LLM calls.
→ The hash chain makes edited journal events detectable.
→ The same input always yields the same verdict.