Add schema doc

SCHEMA.md

@@ -0,0 +1,95 @@
+# 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<struct>
+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.