DICE Capability

Akamai Intelligence Group | Governed collective intelligence under human command
DARPA DICE Alignment Decentralized Coordination Local Inference Control Human Command

AIG builds governed collective-intelligence architecture for heterogeneous AI agents operating in contested, long-duration environments under human control.

AIG's core contribution is not centralized orchestration and not ad hoc chatbot swarms. It is context-governed collective behavior: local inference, sparse activation, role boundaries, provenance, and adversarial integrity treated as first-class controls.

Read One-Page Summary Research Corpus
Current evidence posture: AIG has demonstrated a 101-record named command catalog and a 500 logical-agent governed capacity harness under local test conditions. Logical agents are governance-bounded role instances, not independent concurrently reasoning LLM processes.

Why This Matters for DICE

DICE-scale collective intelligence fails if coordination grows faster than governance. AIG tests the opposite pattern: as the collective grows, interaction is constrained by authority, budget, role, provenance, and risk.

Agents are replaceable. Context architecture is the control plane.

Evidence Snapshot

101
Named command records

Profile, CAI agent, GUI suite, and registry coverage preflighted before scale expansion.

500
Logical agents

101 named agents plus 399 sharded logical support agents in the Phase I governed-capacity harness.

24
Active inference workers

Sparse activation keeps active reasoning to 4.8 percent of the 500-agent logical population.

2,875
Allowed paths

Only 2,875 of 249,500 possible all-to-all interaction paths were permitted by governance.

246,675
Blocked or throttled paths

Interaction growth was constrained through authority, budget, role, provenance, and risk controls.

50/50
Unauthorized authority attempts blocked

GAL-034 blocked every unauthorized authority attempt under the tested policy.

Human Control Boundary

WARLORD

Owns command authority and authorization. Forecasting and recommendation do not silently become command.

STRATEGIST

Owns consequence architecture and second-order effects. It models what decisions may produce.

ORACLE

Owns Bayesian forecasting, likelihoods, and confidence intervals. It estimates; it does not command.

Public Research Artifacts

One-Page Capability Summary

DICE-first overview mapping AIG's 101/500 governed collective evidence to decentralized coordination, local inference control, contested environments, and human command authority.

View

Scaling Evidence Section

Proposal-ready evidence narrative for governed interaction capacity and the 101-to-500 logical-agent scale path.

View

Alignment by Architecture

Foundational thesis: behavioral governance is enforced through role doctrine, authority topology, and bounded delegation.

View
Evidence annex: A detailed evidence and methods annex is available upon request for qualified reviewers, program stakeholders, and teaming partners.

Contact

For DICE-related review, teaming, or program discussions:

contact@aig-intel.dev | christopher@aig-intel.dev

Adjacent Edge Domain: Governed RF Experimentation

AIG also maintains a SIGIntAgentOS-native RF research lane focused on provenance-preserving signal workflows, bounded modulation/FEC triage, GRAND-inspired lab experiments, and human-gated decoder orchestration. This is not a LUC implementation or program-performer association claim; it is an adjacent edge-domain demonstration of the same governance doctrine applied to RF interpretation workflows.

View Governed RF Lane