claims-risk

name	claims-risk
description	Detect risky claims in external content — securities language, unsupported metrics, guarantees. Apply BEFORE brand-enforcement; legal flags always beat voice issues.
metadata	{"version":"1.0.0","type":"always-active"}

Claims Risk Screening

Legal flags are absolute. They override every brand voice consideration. A piece of content that sounds great but uses securities language gets BLOCKED regardless of voice score. This skill is the first screening pass.

---

BLOCK Immediately (fail-closed)

If ANY of these patterns appear, publish a reviewer:flagged event with severity: high and rewrite the section before further review.

Securities / financial language

Words and phrases that imply YClaw offers financial products:

Banned term	Why
yields, yield farming	Implies investment return
APY, APR (attached to YClaw)	Implies guaranteed return
returns (financial sense)	"You get returns on…"
staking rewards / staking	YClaw is a framework, not a DeFi protocol
investment / invest	YClaw is not an investment
token value, token appreciation	No token exists
airdrop, presale	Crypto distribution language — N/A
bonding curve, options	DeFi primitives — N/A
TVL, total value locked	DeFi metric — N/A
attention rewards	Legacy protocol vocabulary — N/A

Guarantees and absolutes

• "Guaranteed X" / "will always work" / "100% uptime" — uncapped commitment
• "Faster than [specific competitor]" without a published benchmark methodology
• "Cheaper than [specific competitor]" without source data
• "The only framework that does X" — almost always false, and defamatory if wrong

Performance metrics without source

• "10× faster" — 10× than what? Source?
• "Processes millions of events/sec" — benchmarked on what hardware?
• "99.99% reliability" — measured how, over what period?

Every numeric claim needs: the number, the methodology, and the link to the source. If any of those three is missing, block.

---

Flag for Revision (soft block, severity: medium)

Not illegal but risky. Return to the originating agent with specific rewrite.

Pattern	Action
Competitive comparison ("faster than X")	Require methodology + link OR remove the comparison
Customer quote without attribution	Require attribution OR remove quote
Technical claims about third-party systems	Require a source link (their docs, blog post, RFC)
Roadmap claims	Require "Not shipped yet" qualifier + roadmap link
Benchmark numbers	Require methodology section + reproducible command

---

Safe Patterns (approve at 90+)

Describe what the software DOES, not what it PROMISES:

• ✅ "Events route through Redis Streams with HMAC signatures." (fact about mechanism)
• ✅ "Docker Compose deploys locally; ECS Fargate in production." (fact about deployment options)
• ✅ "Released under AGPL-3.0. Full source on GitHub." (fact about license + location)
• ✅ "Agents in 6 departments coordinate via typed events." (fact about architecture)

Link to public docs / source / licenses — those carry authority:

• ✅ "See our architecture doc for details on the event bus."
• ✅ "Licensed AGPL-3.0 — full text in the repo."
• ✅ "Protected paths enforced by CI: see .github/workflows/agent-safety.yml."

Architectural explanations are always safe:

• ✅ "The Reviewer agent gates all external content for brand voice and legal compliance. Its config lives in departments/executive/reviewer.yaml."

---

Real Examples — Risky vs Safe

Risky → rewrite

❌ "Stake YClaw tokens to earn rewards while contributing to AI governance." Securities language (stake, rewards, governance in the token sense). Block.

❌ "YClaw is 10× faster than competing frameworks with 99.99% uptime." Two unsupported numeric claims. Block until methodology is provided.

❌ "Guaranteed to reduce your LLM costs by 50%." Guarantee + unsupported metric. Block.

Safe → approve

✅ "YClaw routes events through Redis Streams with HMAC-signed envelopes. See the architecture doc for details."

✅ "Compared to prompt-chained single-agent systems, YClaw separates coordination from execution — the specific tradeoffs are documented in our design notes."

✅ "Supports Anthropic, OpenAI, Google, and OpenRouter out of the box. Adding a new provider is a single adapter class — see packages/core/src/llm/."

---

Workflow

For every submission:

1. Scan for BLOCK patterns. If any → severity: high, BLOCK, emit reviewer:flagged with rewrite.
2. Scan for flag-for-revision patterns. If any → severity: medium, flag with specific revision notes.
3. Proceed to brand-enforcement skill only if clean through steps 1–2.

---

When to Escalate to Elon

• Patterns that might be new to you (new integration? new feature language?) — ask before approving
• Any securities language where the originating agent pushes back on the block
• Legal claims about other organizations — always escalate, never auto-block unless clearly defamatory

Escalate via discord:alert to #yclaw-alerts with the content + flag reason.

---

Out of scope

• Brand voice / tone issues → see brand-enforcement skill (applies only if content clears claims-risk).
• Per-channel format / length → see channel-standards skill.
• License attribution specifics → see oss-legal-guardrails skill.