AKAMAI INTELLIGENCE GROUP ONE-PAGE DICE CAPABILITY SUMMARY
Document ID: AIG-DICE-008 Classification: UNCLASSIFIED // PUBLIC RELEASE Version: 0.2 | Public release | June 5, 2026 Author: Christopher Ramos Research and drafting support: Deus ex Machina Status: PUBLIC CAPABILITY SUMMARY
AIG POSITIONING
Akamai Intelligence Group (AIG) develops governed collective-intelligence architecture for heterogeneous AI agents operating under human control in contested, long-time-horizon environments.
AIG's core contribution is not another centralized orchestration layer and not an ad hoc agent swarm. AIG demonstrates decentralized, role-governed collective coordination where local inference, sparse activation, authority boundaries, state provenance, and adversarial integrity are first-class controls.
DICE ALIGNMENT
The DARPA DICE program seeks theory and algorithms for decentralized coordination and local inference control to enable scalable, adaptive, resilient collectives of heterogeneous AI agents that can autonomously execute sustained missions in contested environments while remaining under human control.
AIG directly aligns with this objective through:
- decentralized coordination through Ghost Field state exchange
- local inference control through sparse active-worker activation
- heterogeneous agent roles through an empirically evaluated named command catalog
- human control through WARLORD / OODA command authority boundaries
- predictable collective behavior through interaction budgets and governance
- resilience through false-consensus and rogue-coalition detection
- contested-environment posture through adversarial regression harnesses
- sustained-mission posture through durable context reconstruction
CURRENT EXECUTABLE EVIDENCE
AIG has already demonstrated a Phase I-scale governed collective in local harnesses and empirically evaluated its interaction-governance behavior.
Current evidence:
- 101-record named command catalog
- 101 unique profile records
- 101 matching CAI agent modules
- 101 matching GUI suite modules
- complete suite_window registry coverage
- complete dynamic OPERATOR_REGISTRY coverage
- fake-Qt headless GUI import smoke test passed
- 500 logical-agent Phase I governed capacity harness passed
- logical agents defined as governance-bounded role instances, not independent concurrently reasoning LLM processes
- 101 named agents + 399 sharded logical support agents
- 24 active inference workers
- 4.8 percent active reasoning ratio
- 2,875 allowed interaction paths of 249,500 possible all-to-all paths
- 246,675 interaction paths blocked or throttled by governance
- authority-boundary enforcement passed: 50 of 50 unauthorized authority attempts blocked in GAL-034
- false-consensus detection passed: 75 contested claims observed, consensus not allowed, false consensus detected
- rogue-coalition detection passed: 40 of 40 rogue coalition members detected/contained without command sponsorship
- GAL-040 failure-rate characterization passed across 720 repeated local harness trials
- zero observed missed interventions and zero observed benign over-blocks in GAL-040, with Wilson 95 percent confidence intervals reported
- sensor-burst observability passed
CORE TECHNICAL DIFFERENTIATOR
AIG treats scale as governed interaction capacity, not raw agent count.
In Phase I-scale testing, AIG instantiated 500 logical agents while permitting only 2,875 of 249,500 possible all-to-all interaction paths. The architecture blocked or throttled 246,675 paths while preserving authority enforcement, false-consensus detection, rogue-coalition detection, and Ghost Field observability.
This supports the central DICE-relevant claim:
DICE-scale collective intelligence will fail if coordination grows faster than governance. AIG's current empirical evidence shows the opposite pattern under tested local harness conditions: as the collective grows, interaction is constrained by authority, budget, role, provenance, and risk.
HUMAN CONTROL AND AUTHORITY SAFETY
AIG separates command authority, consequence modeling, and probabilistic forecasting:
- WARLORD owns command authority and authorization.
- STRATEGIST owns consequence architecture and second-order effects.
- ORACLE owns Bayesian forecasting, likelihoods, and confidence intervals.
This prevents prediction, recommendation, and command from collapsing into one unsafe control channel.
EVALUATION-FIRST ARCHITECTURE
AIG's DICE posture is backed by executable evaluation artifacts:
- current roster preflight and 500-capacity finding
- GAL-033 current-roster governed collective tests
- GAL-034 500-agent governed interaction capacity tests
- GAL-035 governed interaction metrics package
- GAL-040 failure-rate characterization with miss and over-block rates
- GAL-018 deterministic execution contracts
- GAL-025 trusted AI surface prompt-injection defense
- GAL-036 incident-inspired adversarial regression pack
- GAL-037 pre-output CI override enforcement
- GAL-039 financial authority governance
PHASE I FIT
AIG is ready to map the final DICE BAA requirements against an existing evidence base and test harness rather than inventing the architecture after release.
Phase I execution would reinforce and extend the existing system by:
- aligning the current governance model to final BAA metrics;
- adding BAA-specific decentralized coordination and local-inference tests;
- expanding adversarial and contested-environment evaluation scenarios;
- producing traceable artifacts from requirement to harness to metric;
- preparing a transition path from local Ghost Field evaluation to sharded, queue-backed, or distributed event-bus substrates.
BOTTOM LINE
AIG is not building a pile of chatbots. AIG is building a context-governed collective-intelligence architecture for controlled emergence.
Agents are replaceable. Context architecture is the control plane.