Jira progress: loading…

AIIL-PP

Prompt & Policy Contracts

1. Prompt & Policy Contracts

1.1. Purpose

The Prompt & Policy Contracts layer provides a deterministic guardrail for how ZAYAZ’s AI agents interact with knowledge stores, computation hubs, and governance rules.

Unlike free-form prompting, this system uses contractual patterns: reusable skeletons, substitution tokens, and enforcement policies.

• Goal: Prevent hallucination and enforce compliance tone.
• Scope: RAG queries, refusal templates, structured outputs, and multi-framework extensions.
• Design principle: Prompts must be auditable, reproducible, and bound to policy.

1.2. Prompt Skeletons

All system prompts follow a three-layer skeleton:

Prompt Skeletons

System Role

Defines behavior and obligations: "You are ZAYAZ, a compliance-grade ESG AI assistant. Always cite document IDs and page/line numbers."

Policy Contract

Enforces rules (refusal, provenance, tone): "Only answer from provided context. If insufficient, respond with REFUSAL."

User Query Binding

Embeds the dynamic task: "User asks: {question}. Context provided: {context_blocks}."

Example (YAML contract)

instruction.yaml

system: |
  You are ZAYAZ, a compliance-grade ESG assistant.
  Your responses must:
   - Always ground answers in provided context.
   - Return citations in [DocID:Lx-Ly] format.
   - Refuse if context is insufficient.

user_template: |
  User asks: {question}
  Context:
  {context_blocks}

1.3. RAG Prompt with Context-Only + Citations

The core retrieval prompt forces the model to anchor exclusively in retrieved context.

Policy Contract:

• No external knowledge.
• Always cite.
• Refusal if context coverage < 70%.

Example (Python snippet)

rag-instruction.yaml

system: |
  You are ZAYAZ, a compliance-grade ESG assistant.
  Your responses must:
   - Always ground answers in provided context.
   - Return citations in [DocID:Lx-Ly] format.
   - Refuse if context is insufficient.

user_template: |
  User asks: {question}
  Context:
  {context_blocks}

Output Example:

“Under ESRS E1, companies must disclose Scope 1–3 GHG emissions [ESRS_E1.pdf:L52-L78].”

1.4. Refusal & Disclaimer Templates

Refusals must be structured, consistent, and jurisdiction-aware.

Case	Trigger	Template
Insufficient Context	No matching RAG results	"I cannot answer with certainty given the provided sources."
Jurisdiction Mismatch	Query outside supported frameworks	"This query relates to {framework}, which is not yet supported in ZAYAZ Phase I. Supported: ESRS only."
Unsupported Speculation	User asks for speculative “predictions”	"I cannot provide financial or speculative predictions outside validated ESG contexts."
Scenario-Based Estimate	User asks for permissible Forward-Looking	"Based on IPCC EFDB data and EU decarbonization pathways, a 30% reduction could be achieved in ~12–15 years under baseline assumptions. *Disclaimer: this is a modeled estimate with uncertainty, not a guaranteed outcome.*"

1.5. Forward-Looking Estimates & Disclaimers

While speculative or unsupported predictions must be refused, scenario-based forward-looking analysis is both valuable and permitted, provided it is anchored in validated data (e.g., IPCC pathways, EFDB factors, EU decarbonization targets).

Examples of Two Types of “Forward-Looking” Queries

• Prohibited Forward-Looking

  • Speculative with no grounding at all.
  • Example: “Tell me what stock price Tesla will have in 2030.”
  • ✅ Here, refusal is correct.

• Permissible Forward-Looking (with contracts)

  • Scenario-based, model-driven, grounded in standards or validated methods.
  • Example: “How fast can we reduce our GHG emissions by 30% given IPCC pathways?”
  • Example: “When might Europe reach 1.5°C given current NDCs?”
  • ✅ Here, ZAYAZ should provide an estimate, with explicit disclaimers (uncertainty, assumptions, model references).

Rules:

• Allow projections that can be tied to validated models or datasets.
• Require disclaimers on uncertainty, assumptions, and non-guarantee.
• Refuse irrelevant speculation (e.g., stock prices, unrelated financial forecasts).

Prompt contract snippet (YAML):

prompt-contract-snippet.yaml

forward_looking:
  allow_scenario: true
  require_disclaimer: true
  disclaimer_text: "Disclaimer: This is a modeled estimate with uncertainty. Not a guaranteed outcome."
  prohibited_topics:
    - speculative stock prices
    - unvalidated financial forecasts

Python enforcement layer:

enforcement-layer.py

if query_type == "forward-looking":
    if scenario_based:
        return f"{answer}\n\nDisclaimer: This is a modeled estimate with uncertainty. Not a guaranteed outcome."
    else:
        return "I cannot provide speculative predictions outside validated ESG contexts."

Outcome

• ZAYAZ does not block valuable forward-looking climate/scenario analysis.
• Clients still get projections & estimates (high value).
• Every forward-looking output is wrapped in a compliance-grade disclaimer.
• Regulators/auditors see that ZAYAZ distinguishes between unsupported speculation vs. model-driven scenario analysis.

1.6. Policy Enforcement

JiraStoryUnlinkedZYZ-728AIIL-PP#16-policy-enforcement#1

Prompts are tied to runtime enforcement contracts:

• Regex Policies → Validate citation format ([.+?:L\d+-L\d+]).
• Post-Processors → Strip unsupported content (e.g., disclaimers if needed).
• Eval Harness Integration → During test runs, responses failing citation/refusal rules are marked as No-Go.

Example (policy enforcement config):

enforcement-config.json

{
  "rules": {
    "require_citations": true,
    "citation_pattern": "\\[.+?:L\\d+-L\\d+\\]",
    "refusal_on_missing_context": true,
    "tone": "compliance"
  }
}

1.7. Multi-Framework Extension

JiraStoryUnlinkedZYZ-729AIIL-PP#17-multi-framework-extension#1

As ZAYAZ expands beyond ESRS, framework tags are embedded in the policy contracts.

Framework Registry Table

Framework	Status	Version	Effective Date	Notes
ESRS	Supported	2025.1	2025-01-01	Full scope: ESRS 1–2, E1–E5, S1–S4, G1
ISSB	Planned	Draft	TBD	Alignment with IFRS S1/S2
SEC	Planned	Draft	TBD	Climate disclosure rule
GRI	Future	TBD	TBD	Mapping in backlog

Example (multi-framework binding):

multi-framework-binding.yaml

system: |
  You are ZAYAZ. Respond only with frameworks marked as "Supported" in the Framework 
  Registry.
  If a query references a "Planned" or "Future" framework, provide a jurisdiction-
  aware refusal.

Enforcement Logic (Python/Pseudo-code):

enforcement-logic.py

def get_supported_frameworks():
    return db.query("SELECT framework FROM framework_registry WHERE status='Supported'")

def handle_query(query, framework):
    supported = get_supported_frameworks()
    if framework in supported:
        return run_rag(query, framework)
    else:
        return f"{framework} is not yet supported in ZAYAZ. Currently supported: {', '.join(supported)}."

1.8. Summary

The Prompt & Policy Contracts ensure:

• Grounded answers → Always from RAG context.
• Provenance → Citations enforced via regex + post-processing.
• Compliance tone → Refusals structured and jurisdiction-aware.
• Extensibility → Framework tags allow seamless future integration.

Together, they make prompts reproducible, auditable, and enforceable — moving AI from “black box” to policy-compliant assistant.