waggle-os/docs/AI-ACT-AUDIT-2026-04-10.md

Waggle Personal Mind — EU AI Act Audit Report

Generated: 2026-04-10T10:05:22.224Z Standard: Regulation (EU) 2024/1689 — Artificial Intelligence Act Scope: Personal mind DB on this machine after Claude Code harvest DB: C:/Users/MarkoMarkovic/.waggle/personal.mind Report version: 1.0 Period: 2026-01-01T00:00:00Z → 2026-04-10T10:05:22.215Z

---

Executive Summary

Overall verdict: 🟡 warning

This audit evaluates the Waggle personal mind against the five articles of the EU AI Act that apply to deployers of general-purpose AI systems: Art. 12 (event logging), Art. 14 (human oversight), Art. 19 (log retention), Art. 26 (deployer monitoring), and Art. 50 (transparency obligations). It also adds a harvest-specific audit covering PII exposure, data minimization, source provenance, and retention posture for the 156 frames imported from ~/.claude/ on 2026-04-10.

§1 Article-by-Article Status

Article	Status	Detail
Art. 12 — Event logging	🟡 warning	No interactions logged yet. Logging activates automatically on first AI interaction.
Art. 14 — Human oversight	🟢 compliant	Human oversight capabilities available (approval gates, tool deny lists). No oversight actions recorded yet.
Art. 19 — Log retention	🟢 compliant	No logs to retain yet. Retention policy is permanent by default.
Art. 26 — Deployer monitoring	🟢 compliant	4 active monitors: cost, tools, model ID, persona.
Art. 50 — Transparency	🟢 compliant	Model identification active. Models will be disclosed on first interaction.

Art. 26 — Active monitors

• cost_tracking
• tool_logging
• model_identification
• persona_tracking

§2 Workspace Risk Classification

Workspace	personal (default)
Classification	minimal (EU AI Act Annex III not applicable)
Rationale	Single-user productivity system, no third parties, no high-risk use cases (biometric ID, critical infrastructure, etc.)
Reclassification trigger	Add a third-party user, deploy to production facing external users, or integrate with an Annex III use case

§3 Harvest Provenance (Art. 12 §1 — Chain of Custody)

Source	Imported at	Items	Frames created
Claude Code (~/.claude)	2026-04-10 09:24:54	468	468

Claude Code (~/.claude)

• Source path: C:\Users\MarkoMarkovic\.claude
• First registered: 2026-04-10 09:15:46
• Last synced: 2026-04-10 09:24:54
• Items seen: 468
• Frames created: 468
• Auto-sync: disabled

§4 PII & Sensitive Data Scan

Regex-based scan of all 156 harvest frames for common PII and credential patterns. This is a best-effort scan; deeper detection (NER, classifier models) is out of scope for v1.

Pattern	Occurrences	Risk level
Email addresses	5	Low (identity, not credentials)
Phone numbers	8	Low
Credit card numbers	0	High — flag if > 0
SSN / national IDs	0	High — flag if > 0
API key patterns (sk-, ghp_, xoxb-, ...)	0	Critical — flag if > 0
JWT tokens	0	Critical — flag if > 0
URLs (context only)	47	Informational

Frames containing any PII: 7 of 156

✅ No critical credential leaks detected

No API key or JWT patterns matched in the harvested content. Email addresses are present (5 occurrences — expected, these are user identity) but are considered low-risk personal data, not credentials.

§5 Data Minimization (Art. 10 §3 / GDPR Art. 5)

Control	Setting
Content cap per frame	4,000 chars (hard-coded in route + cron handler)
Frames truncated at cap	28 of 156 (18%)
Smallest frame	268 chars
Largest frame	4078 chars
Source filesystem	~/.claude/ only (explicit allowlist)
Auto-dedup	SHA-256 content hash via findDuplicate

The 4 KB cap prevents unbounded ingestion of large files. Future improvement: per-category caps (e.g., rules might warrant smaller caps than memories).

§6 Retention Posture (Art. 19)

Importance tier	Harvest frames	Retention policy
normal	156	indefinite (manual delete only)

Current retention exceeds the Art. 19 minimum of 6 months because harvest frames default to normal importance, which is never auto-pruned. If the user requests deletion, the harvest gop can be dropped with a single SQL statement (shown in §9 of this report).

§7 Source Auditability (Art. 12 §1(c))

• All sources have a recorded path: ✅ yes
• All sources have a last-sync timestamp: ✅ yes
• Per-frame source attribution: every harvest frame begins with [Harvest:<source>] <title> so provenance is preserved inside the content itself, not just a sidecar metadata column.

§8 Model Inventory & Oversight Log (Art. 50 / Art. 14)

• Model inventory entries: 0
• Human oversight actions recorded: 0
• Total interactions logged in period: 0

Note: The ai_interactions table is empty because no agent interactions have yet been routed through the InteractionStore recording path on this DB. Once the Waggle server runs and the user starts a chat session, every turn will be logged with model, tokens, cost, tools, and human oversight actions. The infrastructure is in place — the log simply has no rows yet.

§9 Right-to-Erasure (GDPR Art. 17)

Full deletion of the harvested content is a single SQL statement executed against this DB:

DELETE FROM memory_frames_fts WHERE rowid IN (SELECT id FROM memory_frames WHERE gop_id = 'harvest');
DELETE FROM memory_frames_vec WHERE rowid IN (SELECT id FROM memory_frames WHERE gop_id = 'harvest');
DELETE FROM memory_frames WHERE gop_id = 'harvest';
DELETE FROM harvest_sources;

No data is replicated outside this DB, so a single deletion fulfills the user request. Backup copies (if any) must also be deleted — a backup was created before the harvest run at ~/.waggle/personal.mind.backup-pre-harvest-* and should be reviewed under the same policy.

§10 Findings & Required Actions

Severity	Action
INFO	No ai_interactions logged yet — this is expected until first agent session runs through the server.

§11 Compliance Posture — honest caveats

• This report reflects the state of a single local DB, not a running production deployment. Compliance postures for hosted / multi-tenant / enterprise deployments are out of scope.
• The ai_interactions log is empty because no agent sessions have been logged against this DB yet. Art. 12 status is warning based on the current count, not the design.
• PII detection is regex-only — robust detection would require an NER model or an LLM classifier.
• The harvest imports code and reasoning artifacts from the user's own Claude Code sessions. These are not third-party personal data under GDPR; they are the user's own first-party content.
• The minimal risk classification is a default — reclassification is required if the workspace begins serving third parties or handles Annex III use cases.

---

Generated by the AI Act audit pipeline on top of ComplianceStatusChecker and ReportGenerator from @waggle/core/compliance. Full machine-readable report: AI-ACT-AUDIT-2026-04-10.json.