Compliance

Staff Guidance Check

The Staff Guidance Check automatically verifies whether a climate statement meets the requirements of the New Zealand Climate Standards (NZ CS 1, NZ CS 2, NZ CS 3). This guide explains how the check works, what the results mean, and how findings are determined.

What Does It Check?

The platform maintains a catalog of requirements drawn from the NZ Climate Standards. Each requirement specifies what a climate disclosure should contain — from specific data fields (like Scope 1 emissions) to narrative quality (like describing how climate risks are integrated into strategy).

When you trigger a check, the system evaluates a company's climate statement against every applicable requirement and produces a finding for each one — telling you whether the requirement was met, partially met, not met, or not applicable.

How the Check Works

The check uses a two-layer approach to combine objective data validation with qualitative assessment:

Climate Statement

Layer A: Rule-Based

Deterministic checks

Checks if required data fields exist
Validates values are non-empty, numeric, within range
Objective, repeatable, instant results

Layer B: Semantic

AI-powered assessment

Assesses quality and completeness of narrative
Evaluates materiality and entity-specificity
Considers context, coherence, and clarity

Holistic Review

Statement-level fair presentation assessment

Fair Presentation

Could disclosures mislead users?

Coherence

Do sections connect logically?

Financial Linkage

Are financial statement ties identified?

Compliance Report with Findings

Understanding Finding Statuses

Each requirement receives one of these statuses after evaluation:

PASS

Requirement is fully satisfied. Evidence was found and meets the standard.

FAIL

Requirement not satisfied. No evidence found, or evidence does not meet the standard.

PARTIAL

Some evidence found but incomplete. The disclosure addresses the requirement partially.

NOT REQUIRED

Requirement exempted due to adoption provisions or conditions that do not apply to this entity.

UNCLEAR

Cannot be determined automatically. Requires manual review by a human assessor.

Check Modes

Each requirement is assigned a check mode that determines how it gets evaluated:

RULE

Checked only by deterministic rules. The result is binary — the data field either exists and meets the criteria, or it does not.

LLM

Checked only by AI semantic analysis. The requirement is qualitative and cannot be verified by simple data presence.

HYBRID

Checked by both layers. The rule check runs first. If the result is PARTIAL or FAIL, the semantic layer provides a second opinion.

Severity Levels

Not all requirements carry equal weight. Each is classified by severity to help you prioritize:

Blocker

Critical requirement. Failure represents a significant compliance gap that must be addressed.

Major

Important requirement. Failure indicates a notable gap that should be addressed in future disclosures.

Minor

Nice-to-have requirement. Improving this area would strengthen the overall disclosure quality.

Thematic Areas

Requirements are organized into thematic areas that correspond to the major pillars of the NZ Climate Standards:

Governance

Board oversight, climate expertise, committees, management roles, remuneration linkage

Strategy

Transition plans, physical and transition impacts, scenario analysis, time horizons

Risk Management

Risk identification tools, enterprise integration, climate risk register

Metrics & Targets

GHG emissions (Scope 1/2/3), intensity metrics, targets, base year data

General

Fair presentation, reporting period, adoption provisions, assurance

Score Calculation

Two scores summarize the overall result of a check run:

Deterministic Score

Score = (PASS count in RULE & HYBRID)

(Total applicable RULE & HYBRID requirements)

× 100

Measures objective, data-level compliance. NOT_REQUIRED findings are excluded from the denominator.

Semantic Score

Score = (PASS count in LLM & HYBRID)

(Total applicable LLM & HYBRID requirements)

× 100

Measures narrative quality and disclosure completeness. NOT_REQUIRED findings are excluded from the denominator.

Score Thresholds

0–49%

50–79%

80–100%

Red zoneAmber zoneGreen zone

Adoption Provisions

New Zealand entities may claim adoption provisions that exempt them from certain requirements during initial reporting periods. The check automatically accounts for these — requirements that are gated by a claimed adoption provision receive a “Not Required” status rather than being assessed.

AP1

First-time adopter: comparative period disclosures not required

AP2

Scope 3 measurement relief for initial reporting periods

AP3

Transition metrics exemption for first reporting period

Note

The reporting period index (years since 2023) also affects applicability. Some requirements only become applicable after the first reporting year. For example, a company filing its FY2024 statement (period index 1) may be exempt from comparative period disclosures.

Worked Example

Here is a simplified example of how findings might look for a company's climate statement:

Requirement	Mode	Severity	Status
Scope 1 emissions disclosed	RULE	BLOCKER	PASS
Scope 2 emissions disclosed	RULE	BLOCKER	PASS
Scope 3 categories reported	HYBRID	MAJOR	PARTIAL
Board oversight description	LLM	MAJOR	PASS
Transition plan disclosed	LLM	BLOCKER	FAIL
Comparative period data	RULE	MINOR	NOT REQ

Passed

Failed

Partial

Not Required

Technical Reference

Detailed explanation of the algorithms, data flow, and evaluation logic used in the Staff Guidance Check.

Requirements Catalog

The check runs against a structured YAML requirements catalog containing over 100 individual requirements. Each requirement specifies:

idUnique identifier for the requirement
nameHuman-readable description
refsNZ Climate Standards references (e.g., NZ CS 1 para 14)
thematic_areaCategory: GOVERNANCE, STRATEGY, RISK_MANAGEMENT, METRICS_TARGETS, or GENERAL
check_modeRULE, LLM, or HYBRID
severityBLOCKER, MAJOR, or MINOR
jsonpathsPaths to check in the statement JSON (for RULE/HYBRID modes)
rule_checksValidation rules: present_nonempty, equals, min_length, numeric_positive, year_range, present_if_exists
llm_rubricAssessment criteria for semantic evaluation (for LLM/HYBRID modes)
applies_ifConditional applicability based on adoption provisions, period index, or data presence

Deterministic Evaluation (Layer A)

For RULE and HYBRID requirements, the deterministic checker performs these steps:

Applicability check: Evaluate applies_if conditions against the check context. If not applicable, return NOT_REQUIRED.
Gating check: If the requirement is gated by an adoption provision that the entity has claimed, return NOT_REQUIRED with the gating reason.
JSONPath resolution: Query the climate statement JSON using the requirement's JSONPath expressions. Record which paths resolve to data.
Rule evaluation: For each resolved path, apply rule checks (present_nonempty, numeric_positive, year_range, etc.). Record pass/fail per rule.
Status determination:
- All paths found AND all rules pass → PASS
- No paths found AND no rules pass → FAIL
- Some paths found OR some rules pass → PARTIAL

Semantic Evaluation (Layer B)

For LLM and HYBRID requirements, the semantic checker performs these steps:

Evidence extraction: Relevant sections of the climate statement JSON are extracted based on the requirement's thematic area and JSONPaths.
Prompt construction: A structured prompt is built containing the requirement name, NZ CS references, severity level, LLM rubric, entity context (type, financial year, adoption provisions), and the extracted evidence.
AI assessment: The prompt is sent to an LLM which evaluates the disclosure quality. The model returns a structured response including status (PASS/FAIL/PARTIAL/UNCLEAR), confidence score, rationale, missing evidence items, and follow-up questions.
HYBRID merge logic: For HYBRID requirements, if the deterministic layer returned PARTIAL or FAIL, the semantic result takes precedence. If the deterministic layer returned PASS, the finding stays as PASS.

Holistic Review

After all individual findings are collected, a holistic review assesses the statement as a whole. This produces three types of flags:

Fair Presentation Risks

Issues where disclosures could mislead readers. Assesses materiality, completeness, accuracy, and whether important information has been omitted or obscured.

Coherence Issues

Missing logical connections between disclosure sections. For example: risks should connect to metrics, metrics to targets, targets to capital deployment, and scenario analysis to strategy and transition plans.

Financial Statement Linkage

Identifies where climate disclosures create obligations that should be reflected in financial statements. For example, a net-zero commitment may imply capital expenditure plans or asset impairments.

Score Formulas

The two summary scores are calculated as follows:

// Deterministic Score

Deterministic Score = (PASS findings where mode = RULE or HYBRID)

/ (Applicable findings where mode = RULE or HYBRID) × 100

// Semantic Score

Semantic Score = (PASS findings where mode = LLM or HYBRID)

/ (Applicable findings where mode = LLM or HYBRID) × 100

// Applicable = Total minus NOT_REQUIRED

Applicable = Total Requirements − NOT_REQUIRED count

Check Context

Each check run is executed with a context object that influences requirement applicability:

Field	Description	Example
reportingPeriodIndex	Years since 2023 (first CRD year is 2024 = index 1)	1
adoptionProvisionsUsed	List of adoption provisions claimed by the entity	["AP1", "AP2"]
entityType	Regulated entity type classification	REGISTERED_BANK
entityName	Name of the reporting entity	Acme Bank Ltd
financialYear	Financial year of the statement	FY2024

Rule Check Types

The deterministic layer supports the following rule check types:

Rule	Description
present_nonempty	The field exists and contains a non-empty value (not null, empty string, or empty array)
equals	The field value equals an expected value
min_length	The field value (string or array) meets a minimum length
numeric_positive	The field contains a number greater than zero
year_range	The field contains a year within a plausible range (e.g., 2000–2100)
present_if_exists	The field must be present only if a related field exists (conditional presence)