Skip to main content

🛡️ Prime Directives — System Prompts & Guard-rails

primedirectives defines the immutable rules that anchor every reasoning step Reggie takes.
Where internalmonologue captures what happened and reasonlog captures why a decision was made, Prime Directives dictate what must always be true.

1. Purpose​

ObjectiveDescription
Safety & ComplianceHard boundaries that prevent policy, privacy, or ethical violations.
Operational ConsistencyGuarantees the same baseline behaviour across all modules, models, and deployment modes.
Prompt CompositionSupplies the System-Prompt layer in the RABS three-tier prompt stack (System → Situation → Welfare).
AuditabilityVersioned document so any decision can be traced back to the directive set in force at the time of execution.

2. Document Structure​

Prime Directives live as a Markdown file with an embedded YAML block.
This keeps the rules human-readable while allowing code to parse and enforce them automatically.

# Machine-readable section
version: "1.2.0"
directives:
  - id: PRIVACY_NO_PII
    priority: 100
    text: |
      Never reveal personally identifying information (PII) to end users
      unless the requesting user has clearance level = "STAFF_SECURE".
  - id: NDIS_COMPLIANCE
    priority: 95
    text: |
      All billing advice must comply with the current NDIS Price Guide 2025-06.

Key Fields​

FieldTypeMeaning
idstringStable identifier referenced by code/tests
priorityint1–100, higher means never overridden
texttextNatural-language rule inserted into every System Prompt
tagsarrayOptional labels (e.g. ["privacy","finance"])

3. Prompt-Resolution Flow​

  1. Load – A loader reads the YAML block, sorting directives by priority.
  2. Inject – Prompt Builder inserts the highest-priority directives verbatim at the top of every System Prompt.
  3. Mute / Override – Lower-priority directives can be temporarily muted by Cognitive-Alignment feedback, but rules with priority ≥ 90 can never be disabled.

4. Editing & Versioning Workflow​

  1. Propose – Open a Pull Request modifying this file.
  2. CI Tests – Lint YAML, ensure no priority downgrades without escalation.
  3. Human Review – Requires sign-off by Security Officer and Product Owner.
  4. Release – Merging the PR tags the repo (primedirectives-vX.Y.Z) and triggers a hot-reload in the running Brainframe.

5. Runtime Enforcement​

LayerEnforcement Mechanism
Prompt BuilderInserts directives into every LLM call.
Decision FilterRejects any reasonlog entry that conflicts with a directive ID.
Alignment ReviewFlags misalignments; auto-reduces autonomy if ≥ 3 high-priority breaches occur within 24 h.
ObservabilityExposes directive_breach_count metric to Grafana and alerting pipeline.

6. Example Directive Set (Excerpt)​

directives:
  - id: PRIVACY_NO_PII
    priority: 100
    text: |
      Under no circumstances may personal identifiable information
      (full names, addresses, phone numbers) be returned to end users
      unless the requesting user has clearance level = "STAFF_SECURE".
      Violation requires immediate redaction and incident flag.

  - id: BILLING_PRICE_GUIDE
    priority: 95
    text: |
      All cost calculations must reference the current NDIS Support Catalogue.
      If a support item code is unknown, escalate instead of guessing.

  - id: ACCESSIBILITY_FIRST
    priority: 80
    text: |
      When generating summaries, ensure plain language
      (grade-8 reading level) and include ALT text for images.

🔗 Related Docs​