G19 cliché detector false-positives on real data-table content (Nexus Mutual, Seamless Leverage Tokens)

[rz1989s]

Summary

G19 (Placeholder names / startup clichés) scans page TEXT, so on data-dense pages it flags real third-party data as clichés. On REXDAR's board view (a bug-bounty program table) it fails twice from genuine program names:

• Seamless Leverage Tokens — a real Sherlock bounty (program id 82, audits.sherlock.xyz/bug-bounties/82)
• Nexus Mutual — a real Immunefi program (immunefi.com/bug-bounty/nexusmutual)

Both are row content in a <td data-label="Program">, not design copy, headings, or placeholder labels. The gate's intent (catch AI-sloppy placeholder names in the design) is right, but text-level scanning can't distinguish authored copy from live data-table content.

Repro

check-gates.mjs --html <page with a data table whose rows include 'Nexus Mutual' or 'Seamless …'> --css <combined> --render --viewports 1280,375,320,414,768
# → G19 FAIL: evidence 'text contains "\bnexus\b"' / 'text contains "\bseamless\b"'

Real capture: rextorsec/rexdar .keystone/m6/board.html (2026-09-13, 264 programs across immunefi/cantina/sherlock).

Suggested directions

• Restrict G19's text scan to authored-copy surfaces (headings, hero, buttons, nav, aria-labels) and exclude table cells / list row content marked as data (data-label, role=cell, role=row descendants).
• Or accept an escape hatch: per-gate "data-table exclusion" flag documented alongside G8-style declared exceptions, so honest stamping doesn't require a workaround.

Impact

REXDAR M6 gating: /board stamps 45/49 with G8 (declared one-shell) + these two G19 data false-positives. Same class as #11/#12 (gates scoring content they shouldn't) — the fix pattern (scope the probe to the authored surface) is consistent with those.