Privacy Baseline (Pre-Beta)

Generated from the repository source of truth. Source file: /docs/PRIVACY-BASELINE.md.

Purpose

This document defines the minimum privacy controls expected when adopting aioc before 0.1.0-beta.1.

aioc provides runtime hooks and trace primitives. Storage governance and data lifecycle controls remain application responsibilities.

Latest adoption snapshot: docs/PRIVACY-ADOPTION.md.

Scope

This baseline applies to:

• any use of run(..., { record })
• any environment where run records may contain business-sensitive or personal data

Privacy Controls Matrix

Control	SDK primitive	SDK default	Required application action	Verification
Context minimization before persistence	record.contextRedactor	pass-through (no redaction)	redact, hash, or drop sensitive fields from context snapshots	unit test with redacted output + manual record inspection
Prompt text capture control	record.includePromptText	false	keep disabled unless there is an explicit legal and operational need	assert promptSnapshots[*].promptText is undefined by default
Metadata hygiene	record.metadata	free-form	enforce metadata allowlist; never write raw secrets, tokens, emails, or full identifiers	lint/check in adapter layer + sample records audit
Trace storage security	record.sink (adapter)	none	enforce encryption at rest/in transit and strict access controls	infra policy review + access test
Retention and deletion	sink-side only	none	define TTL, archival, and deletion process per data class	automated retention job + deletion test
Access accountability	sink-side only	none	log read/write access to stored run records	access audit logs enabled
Tenant boundary separation	sink-side only	none	enforce tenant-scoped partitioning and access filters	integration test with cross-tenant denial
Replay dataset safety	promptSnapshots, requestFingerprints, items	available when record enabled	build replay datasets from redacted records only	sampling + replay dry run

Recommended Runtime Configuration

await run(agent, input, {
  context,
  policies,
  record: {
    includePromptText: false,
    contextRedactor: (ctx) => ({
      contextSnapshot: {
        actor: {
          userId: "[redacted]",
          groups: ctx.actor.groups,
        },
      },
      contextRedacted: true,
    }),
    metadata: {
      appBuildVersion: process.env.APP_BUILD_VERSION,
      scenario: "customer-support",
    },
    sink: runRecordSink,
  },
});

Metadata Guidance

Recommended metadata shape in record.metadata:

• appBuildVersion: host application build/version identifier
• scenario: bounded scenario identifier
• tenantRef (optional): pseudonymous tenant reference
• traceClass (optional): data classification label (public, internal, restricted)

Avoid storing:

• access tokens, API keys, raw session identifiers
• plain email addresses and phone numbers
• full legal names when a pseudonymous id is sufficient

Pre-Beta Go/No-Go Checklist

All items should be true before 0.1.0-beta.1:

1. contextRedactor is configured in production run-record pipelines.
2. includePromptText remains disabled by default.
3. Sink adapters enforce encryption and access controls.
4. Retention/deletion policy is documented and implemented.
5. Metadata allowlist and sensitive-field ban are documented.
6. At least one redaction-focused test exists in CI or release checks.