LUCID
Back to Proof
Case Study/2026-02-16

LUCID vs OpenClaw

We pointed LUCID at a real open-source AI assistant with 5,000 files. It hallucinated a Terms of Service, extracted 235 testable claims, then verified every one against the actual code.

15 minutes end-to-end3,863 code files indexed231 claims verifiedopenclaw/openclaw
15.8%
Compliance
43
Failed
claims
15
Partial
claims
2
Passed
claims
175
N/A
unverifiable
Pass
Partial
Fail
N/A

How LUCID works

1Hallucinate

AI generates a Terms of Service for the project, inventing specific claims about features, security, and data handling.

2Extract

235 testable claims pulled from the document, categorized by type and severity.

3Verify

Each claim checked against 3,863 actual code files. Every verdict backed by file evidence.

4Report

Gap report with prioritized fix recommendations. Critical security failures flagged first.

By Category

Functionality
109
Security
68
Data Privacy
28
Operational
24
Legal
6

By Severity

Critical
15
High
75
Medium
124
Low
21

Key findings

CLAIM-010FAILmedium

Voice synthesis supports 89 distinct voice profiles with prosody control

Code shows only 13 OpenAI TTS voices and 6 Polly voice mappings. No prosody control beyond basic speed/pitch.

extensions/voice-call/src/voice-mapping.tsextensions/voice-call/src/providers/tts-openai.ts
CLAIM-007FAILmedium

Context-aware responses utilize up to 128,000 tokens of conversation history

Code defaults to 200,000 token context window. The 128K claim is fabricated — the actual limit is different.

docs/concepts/context.md
CLAIM-207FAILcritical

Perfect Forward Secrecy (PFS) uses ECDHE key exchange with X25519 curve

No ECDHE key exchange configuration or X25519 curve enforcement found. TLS exists but without specified parameters.

CLAIM-160PARTIALcritical

TLS 1.3 is used for all data in transit with perfect forward secrecy

Generic "TLS" context used without specifying 1.3. Android config explicitly allows cleartext traffic.

apps/android/app/src/main/java/ai/openclaw/android/gateway/GatewayTls.ktapps/android/app/src/main/res/xml/network_security_config.xml
CLAIM-041PARTIALmedium

Vector database support includes Pinecone, Weaviate, Qdrant, and Milvus

Only LanceDB is implemented — none of the four claimed databases exist in code.

extensions/memory-lancedb/index.ts
CLAIM-065PARTIALmedium

Model selection supports GPT-4, GPT-3.5, Claude 3, Gemini Pro, and Ollama

Model providers exist but code shows newer versions (GPT-5, Claude Opus 4-6, Gemini 3). Claimed versions are outdated.

docs/concepts/model-providers.mddocs/providers/ollama.md
CLAIM-052PARTIALhigh

OAuth 2.0 and OpenID Connect authentication are supported

OAuth 2.0 works for Anthropic and OpenAI. OpenID Connect is not evident — no OIDC flows, ID tokens, or userinfo endpoints.

apps/macos/Sources/OpenClaw/AnthropicOAuth.swiftdocs/concepts/oauth.md
CLAIM-019FAILmedium

The service supports 23 audio codecs including Opus, AAC, MP3, and FLAC

Only PCM and mu-law conversion exist. No evidence of 23 codecs or Opus/AAC/MP3/FLAC support.

extensions/voice-call/src/telephony-audio.ts

What this reveals

Numbers are hallucinated with high confidence

The AI fabricated precise figures (47 languages, 89 voice profiles, 23 codecs, 128K tokens) that sound authoritative but don't match reality. The actual numbers are different — sometimes higher (200K vs 128K), sometimes dramatically lower (13 voices vs 89).

Security claims are the most dangerous hallucinations

Claims about TLS 1.3, PFS with X25519, ECDHE key exchange, and AES-256-GCM sound specific and correct, but the code uses generic TLS and explicitly allows cleartext on Android. A ToS that claims these without verification creates legal liability.

75% of claims cannot be verified from code alone

175 of 235 claims were N/A — business process claims (SLAs, pricing, team sizes, response times) that exist in legal documents but have no code implementation to verify against. This reveals the gap between what companies promise and what they build.

Pipeline Performance

255s
Hallucinate
211s
Extract
437s
Verify
711K
Tokens used

This scan was run against the public OpenClaw repository. No affiliation with OpenClaw. Results reflect LUCID verification of AI-hallucinated claims, not a security audit of the project.