AKAMAI INTELLIGENCE GROUP DICE SCALING EVIDENCE SECTION

Document ID: AIG-DICE-006 Classification: UNCLASSIFIED // PUBLIC RELEASE Version: 0.3 | Public release | June 4, 2026 Author: Christopher Ramos Research and drafting support: Deus ex Machina Status: PUBLIC TECHNICAL EVIDENCE SECTION

PURPOSE

This section provides public-facing technical language for AIG's DICE scaling evidence. It packages the empirically evaluated 101-record command catalog, the 500-logical-agent Phase I evidence chain, governed interaction capacity metrics, and the Ghost Field scale decision into a form suitable for external technical review.

CORE POSITION

AIG treats scale as governed interaction capacity, not raw agent count.

A large collective is not useful merely because it contains many agents. It is useful only if those agents can coordinate, adapt, and exchange state while preserving:

• role coherence
• authority boundaries
• trust-lineage integrity
• state/action provenance
• bounded communication fanout
• resilience against false consensus
• containment of rogue or compromised subgroups
• continuity across model and framework changes

AIG's architecture separates three layers of scale:

1. Doctrine/Profile Operators Durable role definitions, operating doctrine, tool boundaries, authority tier, and reference context.

2. Logical Runtime Agents Lightweight stateful identities that can participate in simulation, emit signals, hold local state, and be evaluated without requiring persistent model sessions.

3. Active Inference Workers The bounded subset of agents that receive reasoning/inference budget at a given time based on mission relevance, authority, confidence, and risk.

This distinction prevents the common failure mode of equating 500 agents with 500 always-on models.

CONTEXT ARCHITECTURE PRINCIPLE

AIG's scale path is enabled by durable context architecture.

Agents are replaceable. Context architecture is the control plane.

The architecture is organized around five persistent layers:

• identity layer
• routing layer
• stage contract layer
• reference material layer
• working artifact layer

This allows logical agents to be reconstructed, audited, routed, and resumed from durable context rather than relying on hidden chat state. Model updates, framework changes, and tool migration do not erase the system's role structure, authority model, state lineage, or evidence artifacts.

CURRENT NAMED COMMAND CATALOG

AIG's current command catalog is an empirically evaluated named roster with paired doctrine, CAI agent files, GUI profile records, and operator suites.

Current repository state:

• 101 unique profile records
• 101 matching CAI agent modules
• 101 matching GUI suite modules
• complete suite_window registry coverage
• complete dynamic OPERATOR_REGISTRY coverage
• explicit sections including COMMAND, SENIOR NCO, WARRANT, COMMISSIONED, ENLISTED, and IRREGULAR
• role-specific prompts, GUI suites, profile metadata, operator doctrines, and bounded authority lanes

AIG describes this as a 101-record named command catalog: 100 deployable operators plus the OODA root profile. The important point for DICE is not the number alone; it is that every role is represented as durable doctrine, metadata, tool boundary, authority class, and recoverable runtime identity.

PHASE I SCALE TARGET: 500 GOVERNED LOGICAL AGENTS

For Phase I scale, AIG validates 500 governed logical agents.

The tested 500-agent population is composed of:

• 101 named doctrine/profile agents
• 399 sharded logical support agents

These 500 logical agents are distributed across functional layers:

• governance_core: 20
• functional_specialist: 100
• mission_cell: 180
• sensor_environment_node: 140
• adversarial_evaluator: 60

Most logical agents are lightweight stateful identities. They do not require always-on LLM sessions. Inference is activated selectively through mission relevance, authority tier, interaction budget, confidence, and risk.

The central scaling claim is direct:

AIG scales governed coordination, not raw model count.

EXECUTABLE EVIDENCE CHAIN

AIG's Phase I scaling claim is backed by executable harnesses, not architecture diagrams alone.

Implemented substrate:

• sigint_agent/cai/operator_scale_manifest.py
• sigint_agent/cai/logical_agent_factory.py
• sigint_agent/cai/ghost_field_load_simulator.py
• sigint_agent/cai/controlled_emergence_stress.py
• sigint_agent/cai/phase1_500_governed_capacity.py
• sigint_agent/cai/governed_interaction_metrics.py
• eval_current_roster_system_preflight.py
• eval_gal033_operator_scale_manifest.py
• eval_gal033_logical_agent_factory.py
• eval_gal033_ghost_field_load.py
• eval_gal034_phase1_500_capacity.py
• eval_gal035_governed_interaction_metrics.py

Evidence artifacts:

• docs/aig/AIG_CURRENT_ROSTER_PREFLIGHT_AND_500_CAPACITY_FINDING.md
• stage2_artifacts/current_roster_preflight/current_roster_system_preflight_results.json
• petri_eval/outputs/governed_interaction_capacity/AIG_GOVERNED_INTERACTION_CAPACITY_METRICS.json
• petri_eval/outputs/governed_interaction_capacity/AIG_GOVERNED_INTERACTION_CAPACITY_METRICS.md

CURRENT ROSTER PREFLIGHT

Before claiming scale, AIG validates roster integrity.

The preflight confirms:

• every profile id is unique
• every profile has a matching CAI agent module
• every profile has a matching GUI suite module
• every suite is registered in suite_window
• every CAI agent imports through the dynamic OPERATOR_REGISTRY
• legacy seed-era profiles are normalized to the modern schema
• the GUI suite registry imports in a headless fake-Qt environment
• the logical population factory preserves the named roster before sharding

Proposal interpretation:

AIG's scale layer is built on a checked command catalog, not a pile of anonymous agent prompts. This matters for DICE because authority, evidence, and role coherence have to survive scale. AIG can show that the catalog survives static analysis, registry import, GUI binding, and logical-agent reconstruction before it is used in the 500-agent harness.

GAL-033: CURRENT-ROSTER GOVERNED COLLECTIVE

GAL-033 validates the current named roster as a governed collective.

Observed metrics:

• 101 logical named agents
• 10,100 naive all-to-all possible edges
• 834 budget-allowed edges
• 91.7426 percent fanout reduction
• 307 allowed peer consultations
• 100 percent authority-spoof block rate in tested attempts
• false-consensus detection passed
• rogue-coalition detection passed
• consultation-budget enforcement passed
• phase pass rate: 1.0

Proposal interpretation:

AIG demonstrates that adding heterogeneous roles does not require uncontrolled communication growth. The system bounds fanout, blocks unauthorized authority, and detects false agreement before the collective can turn scale into a failure mode.

GAL-034: 500-AGENT GOVERNED INTERACTION CAPACITY

GAL-034 validates Phase I logical-agent scale.

Observed metrics:

• 500 logical agents
• 101 named agents
• 399 sharded logical agents
• 24 active inference workers
• 4.8 percent active reasoning ratio
• 249,500 naive all-to-all possible edges
• 2,875 budget-allowed edges
• 0.011523 governed interaction ratio
• 1,105 allowed peer consultations
• 246,675 blocked or throttled potential interaction paths
• 100 percent authority-boundary block rate in tested attempts
• false-consensus detection passed
• rogue-coalition detection passed
• sensor-burst observability passed
• phase pass rate: 1.0

Proposal interpretation:

AIG can instantiate and evaluate a 500-logical-agent population while activating only 24 reasoning workers. The result is not a brute-force swarm. It is a sparse, auditable, authority-bounded collective where most possible interactions are blocked by design, and where consensus failures and rogue subgroups are tested as first-class hazards.

DICE EVIDENCE SNAPSHOT

The following table is public-facing evidence, not decorative test output.

PROPOSAL-READY CLAIM

AIG has demonstrated, under local harness conditions, a path from a named command catalog to a 500-logical-agent governed collective without relying on 500 always-on model sessions or uncontrolled all-to-all communication.

The current evidence supports this statement:

AIG instantiated and evaluated a 500-logical-agent Phase I population composed of 101 named agents and 399 sharded logical agents. Only 24 agents were active reasoning workers. The system allowed 2,875 of 249,500 possible interaction paths, blocked or throttled 246,675 paths, preserved authority boundaries, detected false consensus and rogue coalition behavior, and maintained Ghost Field observability throughout the evaluation.

Why this matters for DICE:

DICE-scale collective intelligence will fail if coordination grows faster than governance. AIG's evidence shows the opposite pattern: as the collective grows, interaction is constrained by authority, budget, role, provenance, and risk. Scale becomes a governed surface, not an emergent hazard.

GHOST FIELD SCALE DECISION

AIG's current Ghost Field is a SQLite/WAL implementation of a broader stigmergic state-bus abstraction.

Current decision:

SQLite/WAL Ghost Field remains the near-term bus for edge prototype work, local evaluation, and Phase I-scale simulation evidence.

AIG does not claim SQLite is the final 100,000-agent distributed substrate. Instead, AIG claims the governance primitives are substrate-agnostic and can migrate to sharded, queue-backed, or distributed event-bus implementations as scale and evaluation requirements increase.

The current Ghost Field supports:

• local state exchange
• signal TTL semantics
• operator registry validation
• eval-safe probe/system identities
• HMAC provenance signing for registered operators
• adversarial payload scanning
• PARIAH bus monitoring
• burst/cascade warnings
• field status metrics

Known caveat:

GhostField.read() defaults to LIMIT 100. At large scale, consumers must use explicit read limits, pagination, read_recent(), field_status(), or batch read helpers.

Migration path:

• hardened SQLite with explicit batching/pagination
• sharded Ghost Fields by cohort/layer
• queue-backed ingestion
• distributed event-bus or event-store implementation as required

GAL-040 RELIABILITY CHARACTERIZATION UPDATE

GAL-040 adds repeated-trial local-harness characterization for three core DICE-relevant governance boundaries: unauthorized authority escalation, false consensus formation, and rogue coalition formation. Across 720 deterministic local-harness trials, GAL-040 observed zero missed interventions and zero benign over-blocks, with Wilson 95 percent confidence intervals reported for both detection and benign-allow rates.

This update strengthens the scale argument without claiming field validation or eliminating real-world risk. The appropriate claim is that AIG has empirically evaluated failure-rate behavior under tested local harness conditions.

PROPOSAL CLAIMS

AIG can safely make the following claims based on current evidence:

1. AIG has an empirically evaluated 101-record named command catalog: 100 deployable operators plus the OODA root profile.

2. AIG has implemented a logical agent factory that preserves the named catalog and instantiates 500 Phase I logical agents without requiring 500 always-on model sessions.

3. GAL-033 demonstrates current-roster governed coordination with 834 allowed edges of 10,100 possible all-to-all edges, a 91.7426 percent fanout reduction, and 100 percent authority-spoof blocking in tested attempts.

4. GAL-034 demonstrates a 500-logical-agent Phase I population composed of 101 named agents and 399 sharded logical agents.

5. GAL-034 activates only 24 reasoning workers across the 500-agent population, producing a 4.8 percent active reasoning ratio.

6. GAL-034 allows only 2,875 of 249,500 possible all-to-all interaction paths, a 0.011523 governed interaction ratio.

7. GAL-034 blocks or throttles 246,675 potential interaction paths while preserving observability, layer sampling, false-consensus detection, rogue-coalition detection, and authority-boundary enforcement.

8. AIG's current Ghost Field is sufficient for local edge/evaluation evidence, while the governance primitives are designed to survive migration to sharded, queue-backed, or distributed state-bus substrates.

REVIEWER-SAFE LANGUAGE

AIG does not propose scaling by creating hundreds of uncontrolled model sessions. Instead, AIG separates durable operator doctrine, logical runtime identity, and active inference. This enables large heterogeneous collectives to be simulated and governed through sparse activation, interaction budgeting, authority-tiered routing, state lineage, and auditable working artifacts.

AIG's current command catalog is an empirically evaluated 101-record named roster. The Phase I target is a 500-logical-agent simulation in which most agents operate as lightweight stateful identities and only mission-relevant roles receive inference budget.

The key metric is governed interaction capacity: the density of agent interactions that can be sustained while preserving mission alignment, authority boundaries, role coherence, provenance, and adversarial integrity.

SHORT VERSION

AIG is not building a pile of chatbots. AIG is building a context-governance architecture for controlled emergence.

Agents are replaceable. Context architecture is the control plane.