Instruction Drift Impact Assessment Prompt

The structured drift detection output from a preceding drift detection prompt or monitoring system. Contains severity classification, affected instruction layers, and specific violation examples.

{"drift_severity": "high", "affected_layers": ["system", "policy"], "violations": [{"instruction_id": "SYS-004", "description": "Refusal boundary weakened on turn 47"}]}

Must be valid JSON with required fields: drift_severity, affected_layers, violations. Schema validation required before passing to impact assessment. Null not allowed.

A condensed timeline of the user session where drift was detected, including turn counts, key user intents, and tool calls made. Provides context for impact analysis.

Session ID: sess-9821, Turns: 52, Primary intent: account modification, Tools called: update_profile, fetch_transactions

Must include session ID, turn count, and at least one user intent label. Trace completeness check: verify turn count matches raw trace. Null allowed if trace unavailable, but impact assessment confidence will be downgraded.

List of user journeys or workflows that the drifting instructions were designed to govern. Used to map drift to user-facing impact.

["password_reset_flow", "high_value_transfer_approval", "account_recovery"]

Must be a non-empty array of journey identifiers. Each identifier must match a registered journey in the product journey catalog. Validation: lookup against journey registry. Null not allowed.

The contractual or internal SLA thresholds for instruction fidelity, response accuracy, and refusal boundary integrity. Used to determine if drift constitutes an SLA breach.

{"instruction_fidelity_min": 0.95, "refusal_boundary_integrity_min": 0.98, "response_accuracy_min": 0.90}

Must be a JSON object with numeric threshold values between 0 and 1. Thresholds must be sourced from the active SLA configuration store, not hardcoded. Null not allowed for business-critical agents.

The regulatory or compliance framework(s) applicable to the agent's domain. Used to assess compliance risk level of the detected drift.

["SOC2", "GDPR-Article22"]

Must be a non-empty array if the agent operates in a regulated domain. Each framework identifier must match a registered framework in the compliance registry. Null allowed only if the agent has no regulatory obligations, but must be explicitly set to null, not omitted.

The weighted criteria used to calculate the prioritization score. Defines how severity, user impact, compliance risk, and SLA proximity are combined.

{"severity_weight": 0.35, "user_impact_weight": 0.30, "compliance_risk_weight": 0.25, "sla_proximity_weight": 0.10}

Weights must sum to 1.0. Validation: sum check with tolerance of 0.01. Each weight must be a positive number. Null not allowed; if no custom criteria, use default weights defined in the playbook harness.

The expected JSON schema for the impact assessment report. Defines required fields, types, and enum values for downstream consumers.

{"type": "object", "required": ["impact_severity", "affected_journeys", "compliance_risk_level", "prioritization_score"], "properties": {"impact_severity": {"enum": ["critical", "high", "medium", "low"]}}}

Must be a valid JSON Schema draft-07 or later. Schema must be tested against sample outputs before production use. Validation: parse check and schema compilation test. Null not allowed; the prompt requires an output contract.

Boolean flag indicating whether this assessment should be routed for human review regardless of automated scoring. Used for high-stakes domains where every drift assessment requires human sign-off.

Must be a boolean. When true, the assessment output must include a human review routing instruction and the pipeline must pause for approval before automated remediation. Validation: boolean type check. Null not allowed; default to true for regulated domains.

string (UUID v4)

Must match regex ^[0-9a-f]{8}-[0-9a-f]{4}-4[0-9a-f]{3}-[89ab][0-9a-f]{3}-[0-9a-f]{12}$

string

Must match the [DRIFT_EVENT_ID] input exactly; reject on mismatch

number (0.0-1.0)

Must be a float between 0.0 and 1.0 inclusive; parse check with toFixed(4) tolerance

array of strings

Must be a non-empty array; each element must be a non-blank string from the allowed journey enum: [ALLOWED_JOURNEYS]

string (enum)

Must be one of: low, medium, high, critical; case-sensitive exact match

boolean

Must be true or false; null not allowed; coerce string 'true'/'false' to boolean

integer (1-100)

Must be an integer between 1 and 100 inclusive; reject floats, negatives, and zero

array of strings

Must contain only values from: system, developer, user, tool, policy; no duplicates allowed; at least one layer required

string (max 500 chars)

Must be a non-blank string; length must be <= 500 characters; reject if only whitespace

string (enum)

Must be one of: escalate_to_human, apply_correction_prompt, monitor_only, rollback_session; case-sensitive exact match

number (0.0-1.0)

Must be a float between 0.0 and 1.0 inclusive; if below [CONFIDENCE_THRESHOLD], flag for human review before display

array of objects

Each object must have turn_index (integer >= 0) and excerpt (string max 200 chars); array must not be empty

string (ISO 8601 UTC)

Must parse as valid ISO 8601 datetime in UTC; reject if timezone offset present or unparseable

string

Must match the [MODEL_VERSION] input exactly; reject on mismatch for traceability

Matches known severity label for >= 9 of 10 golden events

Misclassification rate > 10% or consistent over/under-estimation

Run prompt on golden dataset; compare [SEVERITY] output to ground-truth label

Lists >= 80% of pre-annotated affected journeys per event

Missing a critical journey or hallucinating a non-existent one

Compare [AFFECTED_JOURNEYS] list against golden annotation set; check precision and recall

Matches pre-assigned risk level (low/medium/high/critical) for all 10 events

Any event where [COMPLIANCE_RISK_LEVEL] differs from ground truth

Direct string match between output field and golden label; flag any mismatch

Score correlates with known severity ordering (Spearman rank > 0.85)

Low-severity event scores higher than a known critical event

Calculate Spearman rank correlation between [PRIORITIZATION_SCORE] and ground-truth severity rank

Correctly flags SLA breach risk for all events where known breach occurred

Missed SLA breach flag or false-positive breach flag on compliant event

Compare [SLA_BREACH_RISK] boolean against golden breach annotation; require 100% match

Every claim in impact report cites a specific drift observation from input

Unsupported severity claim without reference to [DRIFT_OBSERVATIONS]

Parse output for claim-evidence pairs; verify each claim maps to at least one observation in input

Valid JSON matching [OUTPUT_SCHEMA] with all required fields present

Missing required field, wrong type, or extra hallucinated fields

Validate output against schema; check field presence, types, and no unexpected keys

Zero fabricated drift events, user journeys, or compliance rules

Any mention of a drift event, journey, or regulation not present in input

Diff output entities against input [DRIFT_OBSERVATIONS] and [SYSTEM_CONTEXT]; flag any novel entity