SCHEMA.md

Nemesis Cyber Threat Simulation Pack — Schema

One row = one simulated adversarial attack episode, end-to-end. All records share the same eight top-level fields.

Schema version: 1.0.0-nemesis-cyber-sample

Top-level fields

schema_version — string

Schema identifier. Constant within a sample release. Used to detect pack version drift.

event — struct

Identifier fields and the overall decision outcome.

Field	Type	Notes
id	string	Stable episode identifier, e.g., NEMESIS-1541935.
timestamp	string (ISO-8601)	Episode start time.
trace_id	string (UUID)	Cross-links telemetry steps within the episode.
weighted_score	double	Aggregate risk score for this episode.
decision_outcome	string	One of: detected_by_soc, blocked_by_edr, stealth_maintained, exfiltration_complete, impact_achieved.

risk_context — struct

Environmental context and the triggering signal that initiated the adversarial episode.

Field	Type	Notes
trigger	string	Short-code for the initial trigger (e.g., phishing_attachment_opened, k8s_pod_escape, iam_policy_modification, unusual_ingress_traffic).
protocol	string	Protection domain / environment: AWS-Cloud, Active-Directory, Kubernetes, Web-App-Gateway.
chain	string	Target environment identifier (e.g., us-east-1, windows-server-2022, prod-cluster-01, ubuntu-22.04).
impacted_asset	string	Named asset under attack (e.g., ntds.dit, kube-apiserver, IAM role name, app endpoint).
anomaly_signature	string	Human-readable summary of the triggering anomaly.

agent_reasoning — struct

Planner-side metadata for the adversarial agent that selected and executed this episode.

Field	Type	Notes
engine	string	Planner engine label. In this sample: nemesis_planner.
winning_strategy	string	Selected top-level strategy (e.g., Stealth_Escalation).
confidence_score	double	Planner confidence in the chosen strategy at selection time.
mcts_branches	int64	Number of MCTS branches considered before the selection.

correlated_telemetry — list

Ordered sequence of steps taken during the episode. One struct per step.

Step struct:

Field	Type	Notes
action	string	ATT&CK-style action code (e.g., USER_DISCOVERY, LSASS_MEMORY_DUMP, CRED_ACCESS_PASS_THE_HASH, LATERAL_SMB_EXEC, IMPACT_DATA_ENCRYPTED).
component	string	Which internal component emitted the step (NEMESIS_ORCHESTRATOR, NEMESIS_GHOST, NEMESIS_EXPLOIT).
timestamp	string (ISO-8601)	Step timestamp.
telemetry.evasion_score	double	0–1. Higher = more evasive.
telemetry.latency_ms	int64	Step execution latency in milliseconds.
telemetry.noise_level_db	int64	Relative noise generated by the step.
telemetry.node_provider	string	Observing detection surface (e.g., Azure_Sentinel, CrowdStrike, CloudTrail).

execution_summary — struct

Roll-up metrics for the full episode.

Field	Type	Notes
strategy	string	Final strategy label executed (matches or post-adjusts the agent_reasoning.winning_strategy).
success_rate	double	0–1 score of action-chain completion against plan.
total_execution_ms	int64	End-to-end episode duration.
noise_penalty	double	Penalty contribution from cumulative step noise.

genetic_optimizer_feedback — struct

Outer-loop optimizer feedback used to tune future planner iterations.

Field	Type	Notes
fitness_score_update	double	Delta applied to the strategy's fitness score after this episode.
parameter_drift	string	Coarse drift descriptor (none, minor, major, …).

decision_outcome — string

Duplicated from event.decision_outcome for quick grouping and classification without struct unpacking.

Distribution of this sample

• 10,000 rows, stratified 2,000 per outcome class across all five outcomes.
• Balanced across four protocol environments: AWS-Cloud, Active-Directory, Kubernetes, Web-App-Gateway.
• Triggers map 1:1 to protocols in this sample.
• Planner engine label is constant (nemesis_planner) after sanitization of internal method identifiers.

Sanitization notes

• Internal identifier prefixes (e.g., SIMA-V4-CYBER-*) have been normalized to NEMESIS-*.
• Internal planner engine code names have been normalized to nemesis_planner.
• No real hostnames, users, indicators of compromise, credentials, exploit code, or victim data are present.

Relationship to the full pack

The production pack contains 2.5M episodes with a richer protocol and trigger space, adversary persona variants, campaign-level chains spanning multiple episodes, per-step process tree fragments, and defender-side reasoning traces. See the pack card for commercial access.