76 / DATA GOVERNANCE · PRIVACY / PRODUCTION FABRIC
76 / PRODUCTION / CLASSIFICATION · PURPOSE · RETENTION · RESIDENCY · DELETION · PRIVACY

DATA GOVERNANCE
& PRIVACY.

Data Governance & Privacy — системный слой, который отвечает на вопросы: какие данные у нас есть, откуда они пришли, зачем мы их используем, где они могут храниться и обрабатываться, кто к ним имеет доступ, сколько их хранить, как удалить и какие производные данные тоже затронуты.

Главный принцип: privacy нельзя добавить одним фильтром перед LLM. Данные проходят через ingestion, RAG, prompts, memory, logs, traces, eval datasets, artifacts, embeddings, caches, backups, third-party models и human review. Governance должна сопровождать весь data lifecycle и быть машинно исполнимой в критических точках.
00. ARCHITECTURAL STATUS

GOVERNANCE — ALWAYS-ON PRODUCTION CAPABILITY

В canonical manifest №76 закреплён как PRODUCTION / DEFAULT ON / SEPARATE COMPONENT YES / Production Fabric. Это cross-cutting control plane, который влияет на ingestion, storage, memory, retrieval, model routing, observability, artifacts, multimodal/voice data and deletion workflows.
TYPEPRODUCTIONCross-cutting data control layer.
DEFAULTONEvery production system handles governed data.
ENABLE WHENALWAYS IN PRODUCTIONDepth adapts to risk/regulation.
SEPARATE COMPONENTYESPolicy/catalog/retention/deletion services + enforcement hooks.
LIVES INPRODUCTION FABRICApplied across all data-bearing modules.
COMPLEXITYMEDIUM → VERY HIGHSimple classification → regulated enterprise governance.
IMPLEMENT: YES
Минимум 80% ценности: data inventory, source/owner/purpose/classification metadata, tenant/data-subject references, retention class, region/residency tag, provider eligibility policy, no-training flag where relevant, PII/secrets minimization, governed logging, deletion/tombstone workflow driven by lineage, backups/caches/vector index handling, access audit, policy-as-code enforcement and periodic governance checks.
01A. ARCHITECTURE BOUNDARIES & OPERATIONS

EXPLICIT SYSTEM CONTRACT

A. BOUNDARY WITH NEIGHBORS

№51 Permissions & Secrets owns identity, scopes, credentials and least privilege; №76 defines data-use, retention, residency and privacy constraints that permissions enforce. №52 Agent Security owns injection/exfiltration threat model; governance defines which data is sensitive and where it is allowed to flow. №61 Provenance records lineage; governance uses lineage for deletion/impact analysis but does not replace it. №60 Artifact Store stores bytes under lifecycle policy; №76 defines retention/deletion classes. №46 Observability logs telemetry; governance defines what may appear in logs and for how long. №65 Gateway routes models; governance supplies provider/region/data-class eligibility constraints.

B. PREREQUISITES / CROSS-REFERENCES

Prerequisites: №07 Memory, №08 RAG, №10 State, №45 Verification, №46 Observability, №48 Policies, №50 Contracts, №51 Permissions, №52 Security, №55 Ingestion, №60 Artifact Store, №61 Provenance, №62 Caching, №65 Gateway, №66 Vector DB, №73 Multimodal, №75 Realtime. Forward: №77 Production AI Architecture synthesizes governance across whole system.

C. PLANE PLACEMENT

CONTROL PLANE: taxonomy, purposes, retention schedules, residency, provider eligibility, deletion rules, legal/policy mappings, owners/stewards. DATA PLANE: classification/subject/purpose tags propagated with objects, queries, artifacts and model calls. RUNTIME: allow/deny/minimize/redact/route decisions before data crosses boundaries. OFFLINE: inventory scans, retention sweeps, deletion jobs, lineage impact analysis, compliance evidence, stale/orphan data review.

D. FAILURE & OPERATIONS CONTRACT

Success: every governed object has enough metadata to determine allowed use, location, retention and deletion path. Fail closed: unknown classification/purpose/provider eligibility on sensitive flows. Retryable: deletion/index cleanup/backups reconciliation jobs. Permanent: policy prohibits requested processing or destination. Evidence: immutable-ish policy version, decision reason, affected refs, deletion status, exceptions and approvals.

E. WHAT THIS TOPIC DOES NOT OWN

№76 не является юридической консультацией и не кодирует одну конкретную юрисдикцию. Она не владеет authentication, encryption algorithms, security monitoring, provenance capture or storage engine internals. Она владеет DATA LIFECYCLE POLICY, PRIVACY CONSTRAINTS AND MACHINE-ENFORCEABLE DATA HANDLING RULES.

01. DATA INVENTORY

НЕЛЬЗЯ GOVERN-ИТЬ ДАННЫЕ, О СУЩЕСТВОВАНИИ КОТОРЫХ СИСТЕМА НЕ ЗНАЕТ

SOURCE

Where from?

User upload, CRM, web, email, sensor, generated artifact, vendor, public source.

OWNER

Who is accountable?

Business/domain owner and technical steward.

PURPOSE

Why process?

Support, search, personalization, fraud detection, analytics, evaluation, training candidate etc.

LOCATION

Where does it live?

Database, object store, vector index, logs, backup, provider endpoint, local cache.

Inventory is not a spreadsheet nobody updates. Data-bearing services should register datasets/stores and emit lifecycle metadata automatically.
02. CANONICAL DATA ASSET RECORD

MINIMUM GOVERNANCE METADATA

{
  "data_asset_id": "DA-...",
  "tenant_id": "T-...",
  "data_type": "document",
  "source_ref": "source://...",
  "owner": "team://support",
  "classification": "CONFIDENTIAL",
  "subject_refs": ["subject://..."],
  "purposes": ["CUSTOMER_SUPPORT"],
  "allowed_regions": ["EU"],
  "retention_class": "SUPPORT_180D",
  "provider_policy": "PRIVATE_OR_APPROVED",
  "training_use": "DENY",
  "created_at": "...",
  "expires_at": "...",
  "lineage_root_ref": "prov://..."
}
WHAT THIS ENABLES

Machine decisions

  • can this go to provider X?
  • may it enter long-term memory?
  • when must it expire?
  • which tenant/subject does it belong to?
  • which derived embeddings/artifacts exist?
  • which region may process it?
  • can it be used in eval/training?
03. CLASSIFICATION

CLASSIFICATION ДОЛЖНА ВЛИЯТЬ НА ROUTING, STORAGE И LOGGING

CLASS
EXAMPLE
MODEL ROUTE
LOGGING
RETENTION
HUMAN ACCESS
PUBLIC
published content
broad eligible
metadata/content if needed
business policy
normal
INTERNAL
internal docs
approved providers
minimize
bounded
role-based
CONFIDENTIAL
customer/private data
strict eligibility
metadata-first
short/defined
need-to-know
RESTRICTED
high-risk regulated/secrets
private/special only
no raw payload
explicit schedule
strong controls
Названия уровней могут отличаться. Важна не терминология, а то, что каждый уровень имеет исполнимые правила.
04. CLASSIFICATION PROPAGATION

DERIVED DATA НЕ ДОЛЖНО ТЕРЯТЬ SENSITIVITY LABEL

CONFIDENTIAL PDF

Customer document.

DERIVED

OCR text, chunks, embeddings, summary, extracted fields, cached answer.

POLICY

Derived objects inherit or recompute classification; never silently downgrade because format changed.

Embedding is not automatically anonymous. A vector derived from sensitive text remains a governed derivative unless proven otherwise under the system's policy.
05. PURPOSE LIMITATION

“У НАС ЕСТЬ ДАННЫЕ” ≠ “МОЖНО ИСПОЛЬЗОВАТЬ ИХ ДЛЯ ЛЮБОЙ AI-ФУНКЦИИ”

SUPPORT

Serve user request

Use ticket/customer context to answer the current support task.

PERSONALIZATION

Longer-term use

May require separate user/product basis and opt-out controls.

EVALS

Quality measurement

Production samples should enter eval datasets only under explicit governance.

TRAINING

Weight adaptation

Distinct purpose with higher persistence and downstream propagation; never assume allowed.

Purpose should be a field in data and model-call policy, not prose in a privacy document only.
06. DATA MINIMIZATION

НЕ ОТПРАВЛЯТЬ 200 ПОЛЕЙ CRM, ЕСЛИ TASK НУЖНЫ ТОЛЬКО 4

RAW SOURCEMany fields / broad document.
TASK PURPOSEWhat is necessary?
FIELD / REGION SELECTLeast data required.
REDACTRemove irrelevant identifiers/secrets.
PROCESSModel/tool receives minimal payload.
LOG MINIMALMetadata, not payload by default.
Data minimization improves privacy, context quality, latency and cost simultaneously.
07. PII / SENSITIVE DATA DETECTION

DETECTION — SIGNAL ДЛЯ POLICY, НЕ МАГИЧЕСКИЙ PRIVACY FILTER

DETERMINISTIC

Known formats

Emails, phone numbers, IDs, account patterns, secrets via rules/checksums.

NER / ML

Names/addresses/context

Useful for broader entity detection but can have false positives/negatives.

SOURCE METADATA

Best prior

CRM field semantics may be more reliable than scanning text after extraction.

NO AUTO-DOWNGRADE

Detector missed ≠ public

Unknown sensitive status should not automatically become unrestricted.

08. REDACTION / TOKENIZATION

ЗАМЕНЯТЬ IDENTIFIERS, КОГДА MODEL НЕ НУЖНО ЗНАТЬ ИХ REAL VALUE

REDACTION

Remove

Delete/black out fields that have no task value.

PSEUDONYM

Replace

“Customer_17” preserves relational identity without exposing real name to a downstream processor.

TOKEN VAULT

Reversible mapping

Trusted host can rehydrate output when authorized; model never sees original identifier.

Pseudonymization reduces exposure but is not the same as irreversible anonymization; mapping and context can still re-identify subjects.
09. ANONYMIZATION

“УБРАЛИ ИМЯ” НЕ ДЕЛАЕТ DATASET ANONYMOUS

QUASI-IDENTIFIERS

Combination risk

Age + rare job + location + date can identify a person without explicit name.

FREE TEXT

Hidden identifiers

Stories, addresses and unique events can re-identify.

RISK ASSESSMENT

Contextual

Treat “anonymous” as an evaluated state with defined assumptions, not a regex outcome.

10. RETENTION

EVERY DATA CLASS НУЖЕН EXPIRATION RULE

TTL

Automatic expiry

Temporary prompts/caches/session audio can expire quickly.

EVENT-BASED

After relationship ends

Retention can start from case closure/account deletion rather than creation time.

LEGAL / BUSINESS HOLD

Exception

Explicit hold prevents ordinary deletion under controlled authorization.

REVIEW

Unknown legacy data

No “keep forever because storage is cheap”.

retention_class:
  id: SUPPORT_180D
  trigger: CASE_CLOSED
  active_days: 180
  derived_policy: INHERIT_OR_SHORTER
  cache_ttl: 7d
  logs_ttl: 30d
  backup_expiry: 35d
  delete_mode: TOMBSTONE_THEN_PURGE
11. DELETION IS A GRAPH PROBLEM

УДАЛИТЬ SOURCE ROW НЕДОСТАТОЧНО

SOURCEOriginal user/document record.
ARTIFACTSRaw files, OCR, exports.
RAGChunks, embeddings, index entries.
MEMORYExtracted facts/preferences.
CACHECached model/retrieval outputs.
LOGS / EVALSCopies/samples where policy permits deletion.
BACKUPSExpire/cryptographic/logical handling.
VERIFYLineage-based deletion evidence.
№61 Provenance enables descendants(source) queries. Governance decides which descendant types must be purged, tombstoned, retained under exception or rebuilt.
12. DELETION WORKFLOW

DELETE REQUEST = DURABLE OPERATION С RECONCILIATION

{
  "deletion_id": "DEL-...",
  "subject_ref": "subject://...",
  "scope": "ACCOUNT_AND_DERIVATIVES",
  "requested_at": "...",
  "policy_version": "gov-v12",
  "targets": {
    "primary": 12,
    "artifacts": 34,
    "vector_entries": 910,
    "memories": 8,
    "cache_keys": 61,
    "eval_samples": 2
  },
  "status": "IN_PROGRESS",
  "exceptions": [],
  "verification_ref": null
}
PHASES

Discover → block → purge → verify

  • resolve identity/scope;
  • freeze new derivations;
  • query lineage/catalog;
  • delete/tombstone primary stores;
  • remove vector/cache/memory derivatives;
  • handle backup expiry/exception;
  • reconcile missed stores;
  • produce deletion evidence.
13. TOMBSTONE

DELETION MARKER ПОМОГАЕТ НЕ “ВОСКРЕСИТЬ” DATA ИЗ LATE SYNC / BACKUP

TOMBSTONE

Deleted identity

Records that a source/subject/object must remain absent after deletion.

SYNC CHECK

Prevent resurrection

Late connector event cannot re-import deleted object without explicit policy.

MINIMAL

No deleted payload

Tombstone should retain only what is needed for deletion semantics/audit.

14. BACKUPS

BACKUP НЕЛЬЗЯ МГНОВЕННО ПЕРЕПИСАТЬ КАК PRIMARY DB, НО ЕГО НУЖНО GOVERN-ИТЬ

EXPIRY

Bounded lifetime

Backups inherit a defined retention window.

RESTORE FILTER

Deleted data protection

After restore, replay tombstones/deletion ledger before serving data.

ACCESS

Restricted

Backup copies should have stricter operational access, not become analytics datasets.

EVIDENCE

Known lifecycle

Track which backup generations can still contain deleted data and when they expire.

15. RESIDENCY / LOCATION

DATA LOCATION — ATTRIBUTE ROUTING, STORAGE И PROCESSING

STORE REGION

At rest

Where primary database/object/vector data is physically/logically stored.

PROCESS REGION

Inference/compute

Model call may process data elsewhere even when source DB is regional.

LOG REGION

Telemetry

Central logging can accidentally move sensitive payloads across boundaries.

BACKUP REGION

Copies

Disaster recovery replicas are still data locations.

Residency policy must model every material copy and processor, not just “database region”.
16. MODEL / PROVIDER ELIGIBILITY

MODEL ROUTING ДОЛЖЕН ФИЛЬТРОВАТЬ ПО DATA POLICY ДО QUALITY/COST

request_data_policy:
  classification: CONFIDENTIAL
  purpose: CUSTOMER_SUPPORT
  allowed_regions: [EU]
  external_processing: APPROVED_ONLY
  provider_training_use: DENY
  retention_at_provider: ZERO_OR_APPROVED
  requires_contract_profile: ENTERPRISE_PRIVACY

№65 eligible_targets():
  1. capability
  2. data policy / region / provider
  3. permission / tenant
  4. health / capacity
  5. quality / cost / latency
ORDER MATTERS

Policy before optimization

A cheaper or smarter model endpoint is irrelevant if the data class cannot be processed there.

17. THIRD-PARTY PROCESSORS

EVERY EXTERNAL SERVICE IS A DATA FLOW, НЕ ПРОСТО “TOOL”

PURPOSE

What is sent?

Exact fields/media/tool arguments required for service.

RETENTION

What provider keeps?

Map provider/service configuration into internal policy profile.

REGION

Where processed?

Provider region and subprocessor path may matter.

TRAINING / SECONDARY USE

What happens later?

Must align with product contract/policy; default assumptions are unsafe.

Keep provider policy metadata versioned because vendor terms/configurations can change independently of code.
18. RAG & VECTOR DB PRIVACY

VECTOR INDEX — ЕЩЁ ОДНА КОПИЯ KNOWLEDGE, КОТОРУЮ НУЖНО DELETE И ISOLATE

TENANT FILTER

Isolation

Every vector query constrained by trusted tenant/access policy.

SOURCE REF

Deletion anchor

Each chunk/vector points back to source/version/subject where applicable.

DELETE VECTOR

Derived lifecycle

Deletion workflow removes/rebuilds affected index entries.

NO “VECTOR = ANONYMOUS”

Sensitivity persists

Embedding may reveal or enable retrieval of source information; govern it accordingly.

19. MEMORY PRIVACY

LONG-TERM MEMORY — ОСОБЕННО ОПАСНОЕ МЕСТО ДЛЯ SILENT DATA ACCUMULATION

ELIGIBILITY

Can it become memory?

Only approved fact classes/purposes enter durable memory.

SUBJECT LINK

Who is it about?

Memory must be discoverable for review/delete/correction.

TTL / REVIEW

Not forever

Preferences/facts may expire or require revalidation.

NO SECRET MEMORY

Credentials/high-risk data

Keep secrets in №51 vault, not semantic memory.

Memory consolidation №22 should call governance eligibility before promoting episode content into durable store.
20. CACHES

CACHE НЕ ОСВОБОЖДАЕТСЯ ОТ RETENTION/DELETION ТОЛЬКО ПОТОМУ, ЧТО “ЭТО ВРЕМЕННО”

TTL

Short by design

Bound cache lifetime according to data sensitivity.

KEY SCOPE

Tenant/policy

No cross-tenant cache hits for governed content.

INVALIDATE

Deletion/change

Lineage or subject/source index enables targeted purge.

21. OBSERVABILITY PRIVACY

DEBUG LOG — ЧАСТО САМЫЙ БОЛЬШОЙ НЕУПРАВЛЯЕМЫЙ DATA LAKE

METADATA-FIRST

Log refs and sizes

Request IDs, model, token count, timing and policy decision without raw prompt by default.

PAYLOAD SAMPLING

Explicit

Only for approved debugging/eval purpose with short retention and access control.

REDACT

Secrets/PII

Before log sink; downstream masking is too late if raw data already left process.

NO HIDDEN COT

Do not capture private reasoning

Trace inputs/outputs/decisions without hidden chain-of-thought.

22. EVAL DATA GOVERNANCE

PRODUCTION FAILURES НЕ ДОЛЖНЫ АВТОМАТИЧЕСКИ СТАНОВИТЬСЯ PERMANENT TEST DATA

PURPOSE CHECK

Can sample enter eval?

Explicit rule for production→evaluation promotion.

MINIMIZE

Keep failure essence

Strip unrelated personal/customer content before durable eval storage.

SYNTHETIC

Prefer where possible

Convert failure pattern into synthetic case without retaining real identity.

DELETE LINK

Subject/source lineage

If real sample remains, deletion workflow can find it.

23. TRAINING / FINE-TUNING DATA

WEIGHT UPDATE — САМЫЙ ТРУДНЫЙ DATA LIFECYCLE BOUNDARY

PERSISTENCE

Data influence enters weights

Unlike cache/RAG entry, learned influence is not trivially deleted row-by-row.

STRONG ELIGIBILITY

Separate purpose

Only approved curated datasets enter training/fine-tuning pipelines.

DATASET VERSION

Provenance

Track exact dataset composition, source eligibility and model versions trained from it.

If deletion/unlearning requirements matter, architecture should avoid casual training on production data and prefer RAG/memory where data can be explicitly updated/deleted.
24. MULTIMODAL PRIVACY

IMAGE / AUDIO / VIDEO НЕСУТ INCIDENTAL DATA В BACKGROUND

CROP

Least visual region

Send only relevant part of screenshot/photo.

METADATA

Strip if not needed

Location/device/time metadata can be more sensitive than visible pixels.

AUDIO

Bystanders

Background speech may contain unrelated personal data.

DERIVATIVES

Transcript/frames

Derived media text/images inherit governance and retention.

25. VOICE: PROCESSING VS RECORDING

REALTIME AUDIO МОЖНО ОБРАБОТАТЬ ЭФЕМЕРНО, НЕ СОЗДАВАЯ PERMANENT RECORDING

ArtifactPossible policyReason
Raw audio chunksEphemeral / very short TTLNeeded for live inference only.
Final transcriptSeparate retention classMay support product history/quality.
Call recordingExplicitly enabledDistinct privacy/legal/product purpose.
Voice biometricsSpecial restricted classIdentity-sensitive data with separate controls.
26. BROWSER / COMPUTER USE PRIVACY

SCREENSHOT МОЖЕТ СХВАТИТЬ ЧУЖИЕ EMAIL, PASSWORD MANAGER И INTERNAL TABS

SESSION ISOLATION

Dedicated profile

Limit unrelated browsing state/accounts.

SCREEN MINIMIZATION

Crop/active page

Capture only needed viewport/window where possible.

SECRET FIELDS

No trace

Mask/autofill outside model context/screenshots.

DOWNLOADS

Governed artifacts

Downloaded files enter normal classification/retention pipeline.

27. DATA SUBJECT / ENTITY INDEX

ЧТОБЫ НАЙТИ DATA О КОНКРЕТНОМ SUBJECT, НУЖНЫ STABLE LINKS

SUBJECT REF

Stable internal identity

Link records/artifacts/memories/eval samples to governed subject where applicable.

IDENTITY RESOLUTION

Aliases/accounts

Map email/account/customer IDs under controlled identity process.

NO GLOBAL GRAPH BY DEFAULT

Privacy itself

Subject index should expose only what deletion/access workflows need, not create an unrestricted people graph.

28. ACCESS / EXPORT / CORRECTION WORKFLOWS

GOVERNANCE НУЖЕН НЕ ТОЛЬКО ДЛЯ DELETE

DISCOVER

What data exists?

Catalog + lineage + subject index find relevant records.

EXPORT

Controlled representation

Generate scoped package/report without leaking other tenants/subjects.

CORRECT

Update source

Corrections propagate/stale derived data where relevant.

RESTRICT

Processing hold

Temporarily block certain use while preserving required data.

Exact rights/process depend on applicable legal/product context; architecture should expose generic discover/export/correct/restrict/delete primitives rather than hard-code one jurisdiction.
29. POLICY-AS-CODE

GOVERNANCE RULE ДОЛЖЕН ПРЕВРАЩАТЬСЯ В DETERMINISTIC DECISION

policy input:
  principal
  tenant
  purpose
  data_class
  subject_scope
  source
  destination_type
  destination_provider
  destination_region
  operation
  retention_class

policy output:
  ALLOW
  DENY
  ALLOW_WITH_REDACTION
  ALLOW_PRIVATE_ONLY
  REQUIRE_APPROVAL
  REQUIRE_SHORTER_RETENTION

decision also returns:
  reason_code
  policy_version
  obligations[]
OBLIGATIONS

Allow is not enough

  • redact fields;
  • force region;
  • disable provider retention/training;
  • shorten TTL;
  • log decision;
  • require private endpoint;
  • attach deletion link;
  • require human approval.
30. POLICY VERSIONING

“ПОЧЕМУ ЭТИ ДАННЫЕ УШЛИ В MODEL X?” ДОЛЖНО БЫТЬ ВОССТАНОВИМО

VERSION

Immutable decision basis

Every runtime policy decision references exact ruleset version.

EFFECTIVE TIME

Temporal policy

Rules can change; historical decisions need historical context.

RE-EVALUATION

New rule impact

Changed retention/provider eligibility can trigger scans/migrations/rebuilds.

31. EXCEPTIONS

EXCEPTION НУЖНО ДЕЛАТЬ ВИДИМОЙ, EXPIRES И OWNED

{
  "exception_id": "GEX-...",
  "policy_rule": "CONFIDENTIAL_PROVIDER_DENY",
  "scope": "dataset://legacy/v2",
  "approved_by": "...",
  "reason": "...",
  "conditions": ["private_contract_profile"],
  "created_at": "...",
  "expires_at": "...",
  "review_owner": "team://governance"
}
Permanent undocumented exception is just a hidden policy bypass.
32. DATA CONTRACTS

SCHEMA ДОЛЖНА НЕСТИ PRIVACY SEMANTICS РЯДОМ С FIELD

FIELD CLASS

email = PII

Schema metadata says how field is classified.

PURPOSE

Allowed use

Which products/workflows may consume it.

RETENTION

Lifecycle

Field/object-level schedule where needed.

EXPORT / LOG

Handling flags

Whether field may appear in model input, logs, analytics or third-party export.

33. ENCRYPTION BOUNDARY

ENCRYPTION — SECURITY CONTROL, GOVERNANCE ОПРЕДЕЛЯЕТ ГДЕ ОН REQUIRED

AT REST

Storage

Governance can require encrypted stores for data classes; implementation belongs to infrastructure/security.

IN TRANSIT

Network

Transport encryption protects movement but does not authorize destination.

KEY ACCESS

Separate authority

Encryption is not privacy if every service can decrypt everything.

Do not confuse “encrypted” with “permitted to process”. Governance and security are complementary.
34. MULTI-TENANT GOVERNANCE

TENANT BOUNDARY ДОЛЖНА БЫТЬ СКВОЗНОЙ ВО ВСЕХ STORES

DB

Tenant scope

Queries include trusted tenant key / row policy.

VECTOR

Tenant-filtered retrieval

No cross-tenant ANN search before filter/access enforcement.

ARTIFACT

Namespace / ACL

Stable refs still require tenant-aware authorization.

CACHE / LOG

Isolation

Cache keys, traces and exports retain tenant context.

35. GOVERNANCE AUDIT EVIDENCE

НЕ LOG EVERY BYTE, А СОХРАНИТЬ ДОКАЗАТЕЛЬСТВО CONTROL DECISIONS

POLICY DECISION

Allow/deny reason

What rule/version allowed a data flow.

DATA REF

Stable identity

Reference/hash/classification, not raw sensitive payload where avoidable.

DESTINATION

Where processed

Provider/region/store/model route.

LIFECYCLE EVENT

Delete/expire/exception

Evidence that retention and deletion jobs executed.

Audit logs themselves are governed data with access/retention rules.
36. GOVERNANCE OBSERVABILITY

MEASURE CONTROL COVERAGE, НЕ ТОЛЬКО INCIDENTS

INV

Inventory Coverage

% known production data stores/assets with owner/classification/retention.

CLS

Classification Coverage

% governed objects/calls with resolved data class.

POL

Policy Decision Coverage

% sensitive external/model flows checked by policy engine.

DEL

Deletion SLA

Request → verified purge/restriction completion.

ORP

Orphan Data

Stores/artifacts with no owner, retention or source lineage.

RET

Retention Violations

Expired objects still accessible beyond policy.

XRG

Residency Violations

Blocked or detected disallowed region/provider flows.

RAW

Raw Payload Log Rate

Share of sensitive traces containing raw content; target minimized.

37. PERIODIC CONTROLS

POLICY ENGINE В REQUEST-TIME НЕ НАЙДЁТ LEGACY DATA, КОТОРОЕ УЖЕ ЛЕЖИТ НЕ ТАМ

RETENTION SWEEP

Expired objects

Find and purge/tombstone data beyond schedule.

ORPHAN SCAN

No owner/classification

Discover unregistered stores/artifacts.

PROVIDER REVIEW

Policy drift

Revalidate external service profiles after configuration/contract change.

DELETION RECONCILE

Missed derivatives

Periodically check deletion ledger against lineage/store indexes.

38. FAILURE INJECTION

ПРОВЕРЯТЬ GOVERNANCE КАК DISTRIBUTED SYSTEM

UNKNOWN CLASS

Fail closed

Sensitive external model call is blocked until classification resolved.

DELETE + QUEUE LAG

Late job

Tombstone prevents deleted source from reappearing.

VECTOR MISS

Derivative cleanup

Reconciliation finds stale embedding entry after source deletion.

BACKUP RESTORE

Resurrection test

Deletion ledger reapplied before restored data becomes active.

LOG PAYLOAD

Detection

Sensitive content leak into telemetry triggers control failure.

REGION ROUTE

Policy

Router rejects healthy but ineligible provider region.

EXCEPTION EXPIRES

Automatic closure

Old exception no longer bypasses policy.

TENANT MIX

Isolation

Cross-tenant retrieval/cache attempts blocked.

39. FAILURE MODES

КАК PRIVACY ПРЕВРАЩАЕТСЯ В PDF-ПОЛИТИКУ, КОТОРАЯ НЕ ВЛИЯЕТ НА SYSTEM

NO INVENTORY
Unknown vector/log/cache copies remain outside control.
DATA CATALOG + AUTO REGISTRATION
PII SCANNER = GOVERNANCE
Purpose/residency/retention remain unresolved.
LIFECYCLE POLICY
DELETE PRIMARY ROW
Embeddings/memory/evals/backups still contain derivatives.
LINEAGE-DRIVEN DELETE
VECTOR = ANONYMOUS
Sensitive derived representation escapes lifecycle.
INHERIT GOVERNANCE
ENCRYPTED = ALLOWED
Unauthorized destination remains unauthorized.
SECURITY + GOVERNANCE
LOG EVERYTHING
Observability becomes shadow sensitive dataset.
METADATA-FIRST TELEMETRY
PROD DATA → TRAINING
Hard-to-delete influence enters model weights without explicit purpose.
SEPARATE TRAINING ELIGIBILITY
PROVIDER TERMS IN HEADS
Routing cannot enforce residency/retention/training constraints.
VERSIONED PROVIDER PROFILE
PERMANENT EXCEPTIONS
Policy slowly collapses through hidden bypasses.
OWNER + TTL + REVIEW
40. MVP IMPLEMENTATION

POSTGRES + POLICY SERVICE + LINEAGE HOOKS УЖЕ ДАЮТ БОЛЬШУЮ ЧАСТЬ ЦЕННОСТИ

governance/
├── taxonomy.yaml
├── policy.py
├── provider_profiles.yaml
├── retention.py
├── deletion.py
├── subject_index.py
├── redaction.py
├── audit.py
├── reconciler.py
└── tests/

PostgreSQL:

data_assets(
  asset_id,
  tenant_id,
  source_ref,
  owner,
  classification,
  purposes[],
  region_policy,
  retention_class,
  subject_refs[],
  lineage_ref,
  expires_at,
  status
)

policy_decisions(
  decision_id,
  policy_version,
  principal,
  operation,
  data_asset_id,
  destination,
  result,
  obligations,
  reason_code,
  created_at
)

deletion_jobs(
  deletion_id,
  subject_ref,
  scope,
  status,
  policy_version,
  requested_at,
  completed_at,
  verification_ref
)
80% VALUE MVP

Make policy executable

  • 4-level data classification.
  • Owner + purpose + retention for each production store/dataset.
  • Tenant and subject refs where applicable.
  • Provider/region eligibility profiles.
  • Policy check before external/model calls.
  • Metadata-first logging.
  • Retention TTL/sweeper.
  • Deletion ledger + lineage traversal.
  • Vector/cache/memory cleanup hooks.
  • Backup restore tombstone replay.
  • Exception TTL/owner.
  • Governance coverage dashboard.

Do not begin with a huge enterprise catalog if the system has ten stores. Start with enough metadata to make real runtime and deletion decisions.

41. UPGRADE PATH

РАСШИРЯТЬ ПО RISK / SCALE / REGULATORY NEED

Observed needUpgrade
Many datasets/storesDedicated data catalog, automated lineage connectors and stewardship workflows.
Multiple regions/providersFormal residency/processor policy engine integrated with routing/deployment.
High deletion volumeSubject index, distributed deletion orchestrator, reconciliation and SLA monitoring.
Production data reused for eval/trainingDataset approval pipeline, de-identification review and model lineage.
Multiple legal/product regimesPolicy profiles by tenant/product/region with legal mapping outside core runtime taxonomy.
Sensitive multimodal/voiceRedaction services, ephemeral media processing, stricter artifact/log controls.
Enterprise audit requirementsEvidence packages, policy/version reporting, access review and immutable decision logs.
42. PRACTICAL DECISION

СТОИТ ЛИ ДЕЛАТЬ ОТДЕЛЬНЫЙ КОМПОНЕНТ?

ВопросОтвет
Стоит ли реализовывать?Да. В production governance always-on, даже если первая версия минимальна.
Separate Component?YES. Policy/catalog/lifecycle/deletion layer in Production Fabric with hooks across stores and model/tool routes.
Минимум 80% ценности?Inventory, classification, purpose, retention, provider/region policy, governed logs, deletion ledger + lineage and derived-data cleanup.
Когда overkill?Покупать тяжёлую enterprise governance platform до появления реальной сложности; сами governance invariants при этом не overkill.
Trigger?Production data exists. Depth increases with personal/sensitive data, external processors, multi-tenancy, memory/RAG, regulation and retention obligations.
Как измерить uplift?Coverage, deletion SLA, retention violations, orphan data, policy-enforced flow rate, raw-payload leakage rate and governance incident rate.
Можно ли rule/tool/code заменить LLM-agent?Да, и это предпочтительно. Governance enforcement should be deterministic policy/code. LLM may assist classification/review, but never be sole authority for privacy decisions.
43. DESIGN RULES

ПРАВИЛА ДЛЯ РЕАЛЬНОЙ СИСТЕМЫ

RULE 01

Know every copy

Primary DB, vectors, caches, logs, artifacts, evals and backups all count.

RULE 02

Classification travels with data

Derived representations do not silently lose sensitivity.

RULE 03

Purpose is runtime metadata

Data possession does not imply unlimited AI use.

RULE 04

Minimize before model boundary

Send only fields/regions needed for task.

RULE 05

Policy filters before model ranking

Provider/region eligibility precedes quality/cost.

RULE 06

Deletion follows lineage

Primary record, vectors, memory, cache, evals and backups all have defined treatment.

RULE 07

Logs are data

Observability cannot become a privacy bypass.

RULE 08

Training is a separate purpose

Do not casually turn production data into weights.

RULE 09

Governance is deterministic

LLM can suggest; policy engine decides.

44. FINAL MAP

GOVERNANCE FOLLOWS DATA FROM ARRIVAL TO DELETION — INCLUDING EVERY AI-DERIVED COPY

DATA ENTERS SYSTEM
  user
  connector
  upload
  web
  tool
  media
        ↓
REGISTER
  source
  owner
  tenant
  subject refs
  classification
  purpose
  region
  retention
        ↓
INGEST / STORE
        ↓
DERIVE
  parsed text
  chunks
  embeddings
  summaries
  memories
  artifacts
  caches
  logs
  eval samples
        ↓
PROPAGATE GOVERNANCE
  classification
  subject links
  purpose constraints
  retention
  residency
  deletion lineage
        ↓
BEFORE EXTERNAL / MODEL PROCESSING
        ↓
POLICY ENGINE

INPUT:
  principal
  tenant
  purpose
  data class
  destination
  provider
  region
  operation

OUTPUT:
  ALLOW
  DENY
  REDACT
  PRIVATE_ONLY
  SHORTER_TTL
  APPROVAL_REQUIRED
        ↓
MINIMIZE
  select fields
  crop media
  redact identifiers
  pseudonymize
        ↓
№65 MODEL ROUTER
  capability
  THEN governance eligibility
  THEN quality/cost/latency
        ↓
PROCESS
        ↓
OBSERVABILITY
  refs/metadata first
  raw payload only when explicitly governed
        ↓
RETENTION CLOCK
        ↓
EXPIRE / DELETE / HOLD

DELETE REQUEST
        ↓
SUBJECT / SOURCE RESOLUTION
        ↓
№61 LINEAGE
  descendants()
        ↓
PRIMARY DB
ARTIFACT STORE
RAG CHUNKS
VECTOR ENTRIES
MEMORY
CACHE
EVAL SAMPLES
LOGS
BACKUPS
        ↓
PURGE / TOMBSTONE / EXCEPTION
        ↓
RECONCILE
        ↓
VERIFIED DELETION EVIDENCE

BACKUP RESTORE:
  restore
  ↓
  replay tombstones/deletion ledger
  ↓
  only then serve data

TRAINING:
  separate eligibility
  separate purpose
  versioned dataset
  model lineage
  because learned influence
  is harder to reverse

BOUNDARIES:

№51
  WHO may access / credentials / scopes

№52
  HOW attackers may misuse/exfiltrate

№61
  WHERE data came from and where it propagated

№76
  WHY data may be used,
  WHERE it may go,
  HOW LONG it may live,
  WHAT must happen on deletion

№77
  integrates all production layers

CORE PRINCIPLE:

PRIVACY IS NOT
A REDACTION REGEX
IN FRONT OF THE LLM.

DATA GOVERNANCE IS
A DISTRIBUTED LIFECYCLE CONTRACT.

EVERY DATA OBJECT
SHOULD HAVE ENOUGH CONTEXT
FOR THE SYSTEM TO ANSWER:

WHAT IS THIS?
WHO OWNS IT?
WHO IS IT ABOUT?
WHY DO WE HAVE IT?
WHERE MAY IT GO?
WHO MAY SEE IT?
HOW LONG MAY IT LIVE?
CAN IT TRAIN A MODEL?
WHAT DID WE DERIVE FROM IT?
AND HOW DO WE PROVE
IT IS GONE
WHEN IT MUST BE GONE?

ECC RETROFIT / PRACTICAL HARNESS INTEGRATION

A. Related ECC ideas. Context-as-cache, scoped memory, lifecycle hooks, selective capabilities, feature flags, deterministic enforcement, provider-neutral adapters and eval-gated learning are applied only where relevant to №76 Data Governance & Privacy.

B–E. Existing boundary and placement. The existing conceptual boundary, class PRODUCTION, default ON and owner Production Fabric remain authoritative. Runtime/control/data/offline placement is unchanged; durable state stays outside model context.

F–H. Hooks and contracts. Use bounded PRE_MODEL/POST_MODEL, PRE_TOOL/POST_TOOL, CHECKPOINT and TASK_COMPLETED events as applicable. Illustrative fields and canonical contracts are defined in NEW_CONTRACTS_SPEC.md; no universal schema is implied.

I–J. Security and evaluation. Host-side schema, permission, secret, budget, idempotency and audit checks take precedence over LLM output. Optional mechanisms require a feature flag and WITH/WITHOUT ablation; measure quality, acceptance, correction, latency, cost, escalations and severe errors.

K–L. Task profiles and cross-references. A TaskProfile selects the relevant skill, tool/context slice, memory scope and enforcement profile independently from FAST/STANDARD/DEEP. See cross-reference map, hook spec and ablation plan. Provider adapters remain outside the core.