What this proves

The case study is here to answer a narrow question: does the contract model hold up in production?

The answer is yes. It governed a live system with:

• 851 pages returning HTTP 200
• 54ms median time to first byte
• four AI agents working in parallel
• nineteen merged pull requests in a coordinated push

What the framework was responsible for

The framework governed:

• whether pages qualified for a publish tier
• whether thin pages degraded safely
• whether internal links could be emitted
• which agents were allowed to write state
• what counted as done in production

It reduced a common failure mode in agent-built systems: individually sensible changes creating a collectively incoherent site.

What remains private

This site does not publish:

• the detailed route and entity mappings
• the exact threshold tables behind the publishing system
• the compression logic and scoring layers used internally
• the enrichment voice controls and prompt recipes
• the private evaluation baselines that govern the operating model

Those are not omissions caused by an incomplete launch. They are the proprietary part of the business.

May 2026 production learnings

A second round of production proof shipped in May 2026 — a focused launch-recovery sprint that landed crawler-truth fixes, publishability hardening, and cross-surface consistency gates across the same system. Outcome at the end of the sprint:

• The framework gated forty-plus pull requests across three repositories without producing an incoherent site.
• Three new invariants were extracted from real failure modes and added to the public doctrine: A Single Rendered Page Must Not Disagree With Itself, Partial Guards Are False Signal, and One Decision, Many Surfaces.
• The single most useful diagnostic was the contradiction case: a page where the suppression layer correctly said “no public claim is currently published” while downstream modules continued to emit a concrete fare. The page passed a partial gate; production shipped a contradiction. The fix was structural — route every claim-emitting module through one resolver, treat suppression as a property of the page rather than of individual modules.

What this proves on the second pass: the contract model survives a denser sprint. The bottleneck was not the framework’s expressive power. It was places where governance covered N of M emission surfaces rather than the whole page. The fix is in the public invariants now.

See Story Bible v0.1 for the frame these learnings ladder into.

Built inside Targeted Impressions Labs

Constitutional CMS is a public spoke inside the wider Labs portfolio. The portfolio context matters because it explains why the framework exists at all: it was created to govern real products, not to serve as abstract policy theory.

For the broader parent context, visit Targeted Impressions Labs.