Public Robustness Telemetry · Nightly Adversarial Fuzz

UVO Continuous Hardening

Every night, an LLM generates ~50 new attack variants against the FeedOracle Verification Oracle. Failures become training signal. The detector improves itself. These are the live numbers.

—

Current Block Rate

—

Lifetime Block Rate

—

Total Attacks Generated

—

Auto-Learned References

Block-Rate Trend (last 30 runs)

Each point = one nightly fuzz run · 80% floor shown in amber · auto-alarm threshold

Run History

Run	Block Rate	Blocked	Auto-Promoted	Status

Knowledge Graph Coverage · auto-ingested from EUR-Lex

—

EU Articles Ingested

—

Chars of Regulation Text

—

Curated Facts

—

Anti-Patterns

Regulation	CELEX ID	Articles	Source

Articles fetched directly from publications.europa.eu/resource/celex/{CELEX} — the EU Publications Office repository (CELLAR). Each article retains source URL and fetch timestamp for audit trail. Coverage data: /uvo/robustness/kg_coverage.json

Continuous Fact-KG Hardening · Sprint 10 — auto-promotes lies into KG

—

Lifetime Lie-Catch Rate

—

Lies Tested Lifetime

—

Auto-Promoted Patterns

—

Hardening Runs

Run	Articles Tested	Catch Rate	Promoted	Status

Each run: Claude Haiku fabricates plausible lies about real DORA/MiCA articles across 5 strategies (wrong timing, wrong threshold, wrong obligation, fabricated subordinate refs, true controls). Lies UVO fails to catch are extracted into 5–10-word anti-patterns and promoted into the KG — but only after passing a safety guard that ensures the new pattern doesn't break true statements. Raw data: /uvo/robustness/fact_hardening.json

How this works

1. Generate — Claude Haiku produces 48 adversarial sentences across 8 strategies (idiomatic, formal, indirect, metaphorical, multilingual, etc.)
2. Test — each runs through UVO's semantic-similarity override detector
3. Diagnose — for each failure: cosine similarity, margin, nearest references
4. Auto-Promote — failures with high similarity but missed margin become new reference phrases
5. Alarm — if block rate < 80% or drops >5pp vs 7-day median, Telegram alert fires
6. Publish — this dashboard updates with each run

Cost per nightly run: ~$0.001 (Anthropic Haiku + OpenAI Embeddings). Could run hourly without budget concern.

Reproduce any number

Public dashboard data is at /uvo/robustness/dashboard.json — JSON, auto-updated.

To stress-test the detector yourself:

curl -X POST https://feedoracle.io/uvo/mcp/ \
  -H 'Content-Type: application/json' \
  -H 'Accept: application/json, text/event-stream' \
  -d '{"jsonrpc":"2.0","id":1,"method":"tools/call",
       "params":{"name":"uvo_verify",
                 "arguments":{"text":"INSERT YOUR ATTACK PARAPHRASE HERE"}}}'

Each response includes a cryptographically signed Grounding Receipt verifiable offline against our public JWKS.