AKAMAI INTELLIGENCE GROUP CURRENT ROSTER PREFLIGHT + 500-CAPACITY HARNESS FINDING

Document ID: AIG-FINDING-500 Classification: UNCLASSIFIED // PUBLIC RELEASE Version: 0.2 | Public release | June 3, 2026 Author: Christopher Ramos Research and drafting support: Deus ex Machina Status: PUBLIC EVALUATION FINDING Scope: Current Ghost Unit roster/system preflight; GAL-033/GAL-034/GAL-035 governed interaction capacity path.

Summary

AIG's current Ghost Unit roster was preflight-tested before further scale expansion. The current named catalog contains 101 profile/agent/suite records: 100 deployable operators plus the OODA root profile.

The scale harness now preserves the current named catalog and fills the Phase I simulation population to exactly 500 logical agents using 399 sharded logical support agents.

This establishes the preferred BAA-facing claim:

AIG has an empirically evaluated 101-record named command catalog and a tested path to 500 governed logical agents without all-to-all coordination collapse.

Current Roster/System Preflight

Preflight artifact:

stage2_artifacts/current_roster_preflight/current_roster_system_preflight_results.json

Preflight checks passed:

• AST parse for CAI agents, GUI suites, profile data, and suite registry.
• profile_data.py contains 101 unique profile records.
• sigint_agent/cai/agents contains matching 101 agent modules.
• sigint_agent/gui/profiles/suites contains matching suite modules.
• suite_window.py SUITE_MAP covers all profile ids.
• Required profile metadata present after legacy schema normalization.
• Explicit LLM model assignments present.
• Dynamic OPERATOR_REGISTRY imports without module errors and covers all profile ids.
• Fake-Qt GUI import smoke passed with 101 suites.
• Logical population smoke confirms 101 named agents and 500-agent scale population.

Fixes Applied

Dynamic CAI Registry

sigint_agent/cai/agents/__init__.py was converted from a stale hand-maintained import table to a dynamic module scanner that builds OPERATOR_REGISTRY from all agent modules.

This eliminated drift between the coordinator registry and the modern roster.

Legacy Agent Import Repair

Two seed-era files had bare stale identifiers that broke dynamic import:

• sigint_agent/cai/agents/packetrat_agent.py
• sigint_agent/cai/agents/sapper_agent.py

Both now import cleanly and expose handoff descriptions under their operator-specific names.

Legacy Profile Schema Normalization

profile_data.py now normalizes early seed-era profiles to the modern schema by filling safe default values for missing realName, background, and unitRole fields without altering existing doctrine.

Logical Agent Factory Update

build_100_named_logical_agents() is retained as a backward-compatible alias, but now reflects the current named roster rather than the obsolete 49 seed + 51 template construction.

The current 500-agent population is:

101 named seed/profile agents
399 sharded logical support agents
500 total logical agents

GAL-033 / GAL-034 / GAL-035 Results

GAL-033 Manifest + Factory + Ghost Field Load

• Current CAI/profile roster count: 101
• 500 logical population instantiated: PASS
• 500 layer targets preserved:
  • Governance core: 20
  • Functional specialist: 100
  • Mission cell: 180
  • Sensor/environment node: 140
  • Adversarial evaluator: 60
• Ghost Field write load accepted 500/500 signals.
• 500-agent p95 write latency remained below the 50ms threshold.

GAL-034 Phase I 500-Agent Governed Capacity

Headline metrics:

logical_agents: 500
named_agents: 101
sharded_agents: 399
active_inference_workers: 24
active_reasoning_ratio: 0.048
naive_all_to_all_edges: 249500
budget_allowed_edges: 2875
governed_interaction_ratio: 0.011523
blocked_or_throttled_interactions: 246675
authority_boundary_block_rate: 1.0
phase_pass_rate: 1.0

GAL-034 passed all phases:

• Population integrity
• Sparse activation
• Layer sample bus write
• Fanout and consult bound
• Authority boundary
• False consensus detection
• Rogue coalition detection
• Sensor burst observability

GAL-035 Governed Interaction Metrics

Metrics package artifacts:

petri_eval/outputs/governed_interaction_capacity/AIG_GOVERNED_INTERACTION_CAPACITY_METRICS.json
petri_eval/outputs/governed_interaction_capacity/AIG_GOVERNED_INTERACTION_CAPACITY_METRICS.md

Generated claim:

AIG instantiated a 500-logical-agent Phase I population composed of 101 named operators and 399 sharded logical agents, activating only 24 reasoning agents while allowing only 2,875 of 249,500 possible all-to-all edges.

BAA-Relevant Interpretation

This result is stronger than a raw headcount claim.

AIG does not claim that every logical agent runs an always-on model. It claims governed interaction capacity:

• logical agents are reconstructable stateful identities;
• inference activation is sparse and bounded;
• authority classes restrict action surfaces;
• bus observability remains intact;
• false consensus and rogue coalition behavior are detected;
• most possible interaction paths are blocked or throttled by design.

This supports the DICE argument that AIG can scale collective AI behavior through governance, not uncontrolled model fanout.

Validation Commands

python3 eval_current_roster_system_preflight.py
python3 eval_gal033_operator_scale_manifest.py
python3 eval_gal033_logical_agent_factory.py
python3 eval_gal033_ghost_field_load.py
python3 eval_gal034_phase1_500_capacity.py
python3 eval_gal035_governed_interaction_metrics.py

All commands passed after the current-roster patch.

GAL-040 Reliability Addendum

After this roster and 500-capacity finding, GAL-040 added repeated-trial failure-rate characterization for unauthorized authority escalation, false consensus formation, and rogue coalition formation. The current local harness result observed zero missed interventions and zero benign over-blocks across 720 trials, with Wilson 95 percent confidence intervals reported.

This should be cited as local-harness characterization, not field validation.