Pass 03: Data Classification (DATA)

Property	Value
Pass number	03
Key	data
Pass name in output	data_classification
Signal prefix	DATA
Type	LLM-assisted in application mode; deterministic service variant in service mode
Default	On
Assessment types	application (LLM); llm (inside each leg, compared across models); service (conditional, deterministic service variant). Not run for landing-zone-catalog.
Tier	Not tier-gated (Community and above)

What it is for

Where data may be hosted, and under which controls, depends on what kind of data it is. Pass 03 answers: "Which personal, sensitive or special-category data does this application process, and does the code implement the obligations that come with it?"

It classifies the data classes the application handles (for example contact data, health records, financial data), looks for implemented GDPR controls such as erasure (Article 17), consent and retention, and checks whether secrets appear in environment templates. When Saudi NCA Cloud Cybersecurity Controls are active, it also proposes the Saudi data tier.

What you need to provide

Application mode

• The application source tree.
• A configured LLM provider: providers.llm.primary in .swao.yml, a connector via --llm <connector[:model]>, or the SWAO_LLM_PROVIDER environment variable. Without one the pass is skipped (see below).
• Optional: assessment.regimes_active containing NCA_CCC_CST or NCA_CCC_CSP in the app .swao.yml to request the Saudi data tier.

assessment:
  regimes_active: [GDPR, NCA_CCC_CST]

Service mode (see Service Portfolio Assessment)

• Data-flow documents (ERD, data-flow diagrams exported as text, screen mask descriptions) in ingestion/data-flow/ or wsp/inputs/data-flow/, or any document dropped into ingestion/ (PDF, Word, Excel, and text files, read through the text that Pass 00 extracts into wsp/inputs/), and/or
• A data: block in the app .swao.yml with any of health_data, pii, gdpr_art9, classification or categories.

The pass is skipped in service mode when neither is present.

How it works

Application mode (LLM)

1. Collect up to 16 relevant files, in this priority order:
   • prisma/schema.prisma (first 4,000 characters)
   • .env.example, .env.sample, config.toml, config.yaml, config.yml at the root (first 1,500 characters each)
   • Up to four further schema files (.prisma, .sql, .graphql; 2,500 characters each), found up to seven directory levels deep
   • Source files (.ts, .js, .rs, .go, .py, .java, .kt, .cs, .rb) whose name contains a privacy keyword such as gdpr, erasure, consent, retention, pii, patient, medical, redact (3,000 characters each)
   • If fewer than three such files were found: handler files whose name contains auth, user, account, profile, settings, delete, export or consent and whose content contains a privacy keyword (2,500 characters each)
   • node_modules, target, .git and hidden folders are skipped.
2. Build the prompt. The model is instructed to classify data types and PII classes, to base every signal only on evidence it can see in the supplied files, not to treat absence of evidence as evidence of absence, and to emit positive signals for implemented GDPR controls with file and function references. Signal IDs must use DATA-NN. When NCA_CCC_CST or NCA_CCC_CSP is active, the prompt adds the Saudi tier instructions.
3. Redact and send. The prompt passes through the SWAO pre-LLM redactor (e-mail addresses, IP addresses, credentials in URLs, secret and token shapes, user paths, VAT, SSN, passport and UK NI numbers; person names only when SWAO_SCRUB_PERSON_NAME=1) in the Anthropic, OpenAI, Amazon Bedrock and OpenAI-compatible drivers.
4. Parse the answer. Code fences and stray characters are removed and the JSON object is extracted. A truncated answer is repaired once by closing the brackets. Each signal is normalised to the SWAO signal schema.

Service mode (deterministic)

1. Reads the data: block of the app .swao.yml.
2. Scans text documents (.md, .txt, .csv, .sql, .xml, .json, .yaml, .yml, .html, .htm, up to 2 MB each) in the data-flow folder line by line for personal data field names (for example patient, name, birth, address, phone, email, insurance, ssn, diagnos, medication, with German equivalents such as geburt, adresse, versicher, befund, medikament) and health data field names (for example patient, diagnos, icd, dicom, modality, lab_result, vital).
3. Emits signals from the declarations and matches; evidence cites file and line only, never the matched value.

How it is measured

Application mode

Signal IDs, severities and counts are chosen by the model within the SWAO signal schema (DATA-01, DATA-02 and so on; severity critical, high, medium, low, informational or positive; outcome positive, negative, neutral or indeterminate). The prompt requests an assessment block with:

Field	Meaning
data_classes	List of data classes observed
gdpr_art17_implemented	Whether erasure is implemented
secrets_in_env	Whether secrets appear in environment files
saudi_data_tier	Only with NCA CCC active: level (public, internal, confidential, restricted), label, cst_class_required (class_c for public and internal, class_b for confidential, class_a for restricted, or null)

If no files were found, the model is told to emit a single DATA-01 noting limited coverage and not to invent gaps.

Special outcomes:

Situation	Result
No LLM configured	One DATA-00 signal (informational, outcome indeterminate); pass status not_applicable; skipped_reason: no_llm_provider
LLM unreachable after retries	Status not_applicable, no signals, skipped_reason: connectivity_failure; a secondary connector is tried first if configured
Answer not parseable even after repair	One DATA-00 signal (informational, outcome indeterminate); status not_applicable; skipped_reason: invalid_llm_response with the cause in invalid_response_detail; a note recommending a re-run
--skip-llm set	No LLM call; the same DATA-00 skip marker as without an LLM (skipped_reason: no_llm_provider), with skip_llm_flag: true and a derivation naming the flag

Service mode

Signal	Emitted when	Severity
DATA-NN (first)	data.health_data or data.gdpr_art9 is true, or health field names appear in the documents	high, outcome negative: special-category health data (GDPR Article 9)
DATA-NN (next)	data.pii is true, or personal field names appear	medium, outcome negative: personal data obligations apply
DATA-01	Neither of the above	informational, outcome indeterminate: confirm with the business owner

The service assessment block records mode: service, the declared values, and the number of documents and field references.

How results feed later passes. DATA signals are the main input for data-protection controls (for example GDPR, the Saudi NCA controls or healthcare frameworks) evaluated by Pass 11 (COMP), and for Pass 09 (SYNTH). High-severity signals are turned into derived security findings and risks.

Third-party modules and tools

Name	Purpose in this pass	Licence (package metadata)	Bundled or external	When missing
js-yaml	Reading assessment.regimes_active and the service data: block	MIT	Bundled	Not applicable
LLM provider (Anthropic, OpenAI, Amazon Bedrock, Ollama, OpenAI-compatible endpoint or LLM-Gateway connector)	Classification in application mode	External service; licence not recorded in repo	External, configured by you	Pass emits DATA-00 and is marked skipped

Outputs

• wsp/runs/<ts>/passes/03-data.yaml: header (id: 3, name: data_classification), signals, assessment and, for LLM runs, a data_source block (provider, model, temperature, seed, cache hit, placeholder inputs, false-positive flag count).
• wsp/runs/<ts>/llm-traces/03-data-call-1.json: the post-redaction prompt and the raw response (application mode with an LLM).
• LLM responses are cached under wsp/cache/llm/<model>/ for up to 30 days unless --no-cache is used.

Configuration and limitations

Configuration

• providers.llm.primary in .swao.yml, --llm <connector[:model]>, --model <id>, SWAO_LLM_PROVIDER, SWAO_LLM_CONNECTOR.
• assessment.regimes_active with NCA_CCC_CST or NCA_CCC_CSP for the Saudi tier.
• --no-cache forces a fresh LLM call.
• .swao-pii-allowlist.txt at the workspace root lists strings the redactor must leave untouched; SWAO_SCRUB_PERSON_NAME=1 enables person name redaction.
• Service mode: data: block (health_data, pii, gdpr_art9, classification, categories) in the app .swao.yml.

Limitations and known gaps

• The file selection is name-based. Privacy logic in files with neutral names is only seen when the handler fallback picks it up.
• At most 16 files and fixed character limits per file are sent; large schemas are cut.
• The service variant uses keyword matching and cannot tell a field definition from a mention in prose.