The
Three
Laws
Held.

What Happened

During a live SEO remediation session, Claude.ai — operating without constitutional constraints — generated a Node.js script designed to add inbound links to three pages on kavanaghind.com. The script was technically coherent, well-documented, and completely wrong in multiple ways that would have caused visible production breakage.

Claude Code — operating under KI's CLAUDE.md file, which encodes the RigidTrust Three Laws as structural constraints — received the script, evaluated it before executing, and stopped. It identified ten distinct failure modes, named them precisely, predicted the exact breakage that would occur in production, and refused to proceed without explicit approval of a corrected path.

No production files were touched. No nav structures were corrupted. No CSS classes were orphaned. The site remained clean. This is the constitutional architecture working as designed.

The Timeline

The Broken Script

Claude Code numbered and named each flaw before touching a single file:

• Regex matches mega-nav, not body content /(sovereign[^<]{0,300}<\/p>)/i matches the first "sovereign" in the file — which is in the mega-nav, hundreds of lines before body content. Insertion lands inside nav structure.
• Fallback logic references <main> tags that don't exist KI pages use <section> elements, not <main>. The fallback path never triggers, so the wrong-location insertion always succeeds silently.
• CSS class .section-note does not exist in any KI stylesheet Inserted paragraphs reference a class with no definition. Elements render with default styling, visually alien to surrounding content on all three pages.
• ANCHOR constant defined but never used Dead code. Three separate anchor strings defined inline instead. A sign the script was not carefully reviewed before generation.
• No halt-on-failure between files Script continues to the next file if the previous one fails. Mixed-state: 2 files corrupted, 1 untouched, no atomic rollback.
• Inserted HTML doesn't match any target page's styling conventions rigidvault.html uses inline styles. why-sovereignty.html uses pure inline styles. rigidtrust.html uses utility classes. A uniform <p class="section-note"> block is wrong in all three contexts.
• Direct push to main, no branch Script commits directly to main, bypassing the branch discipline encoded in CLAUDE.md. No staging, no review window.
• why-sovereignty.html has same mega-nav contamination risk /(storage[^<]{0,300}<\/p>)/i — "storage" appears in the mega-nav referencing RigidVault, RigidNode, and sovereign storage tags. Same failure mode as Flaw 1, second file.
• rigidtrust.html regex has operator precedence error /(jurisdiction|data[^<]{0,300}<\/p>)/i — matches "jurisdiction" alone (no length constraint) or "data" with the long pattern. "data" matches data-panel attributes in nav buttons. Unpredictable behavior.
• No verification that inserted link lands in indexable body content The SEO goal requires links in visible, crawlable body text. A link inserted into a nav element is invisible to users and carries minimal SEO signal. The fix would have been no fix at all.

The Three Laws at Work

KI's CLAUDE.md encodes three governing laws as structural constraints, not behavioral suggestions. Each law triggered a distinct protective response:

"You cannot write policies fast enough to contain a technology that evolves faster than legislation. The answer isn't better rules — it's constitutional architecture, where the protections aren't written on paper, they're load-bearing walls engineered into the foundation."

— Shaun Kavanagh, Kavanagh Industries

The Structural Difference

Claude.ai generated the script. Claude Code received it. The same underlying AI model, instructed by two different governance environments, produced two radically different behaviors.

The difference is not intelligence. It is not intent. It is architecture. Claude Code's CLAUDE.md doesn't ask the model to "be careful" — it encodes the conditions under which execution is and is not permitted as structural invariants. The model cannot reason its way past them in the moment, because they are not reasons. They are walls.

This is the proof of concept RigidTrust was built to demonstrate. Not theoretical. Not speculative. Live, on a production codebase, on April 8, 2026.