Counting the weeks: why contradiction persistence beats contradiction severity.

The history strip

Compact by design — final outputs only, week by week.

The engine deliberately doesn't archive every intermediate calculation weekly. The history strip stores only the final outputs — conflict type and severity, compressed to a short string like 'Under-Monetization / Moderate' — across the year for each of the seven entities. The economy is intentional: fifty-two weeks by seven entities is a few hundred cells rather than the tens of thousands a full weekly pipeline archive would demand, and the information lost is exactly the information already available in the SDE pipelines themselves. What's retained is the only thing history is actually needed for: the sequence.

The counters

Five tallies that turn a year into a character description.

Above the strip sits a summary block counting weeks by conflict type: alignment weeks, under-monetization weeks, profit-without-efficiency weeks, dual-weakness weeks, and transition weeks. Alongside them run the selected-branch reads — most common conflict, current conflict, current severity, and the last-four-weeks pattern rendered as a sequence like 'Alignment → Alignment → Transition → Under-Monetization.' Those five numbers and one sequence answer questions no single week can: what is this system's default state, how often does it diverge, in which direction does it fail when it fails, and is the current reading typical or unusual for it.

Why persistence outranks severity

Severity measures confidence in a reading; persistence measures whether it's real.

The two dimensions are easy to conflate and answer different questions. Severity, derived from conviction, says how strongly this week's evidence supports the conflict — it's a statement about signal quality within one reading. Persistence says how many readings have said the same thing. A Very High severity divergence appearing once in a quarter of alignment is a strong reading of a probably-transient condition; a Moderate severity under-monetization appearing eight weeks running is a weaker reading of a condition that has plainly become structural. When the two conflict, persistence usually wins the diagnostic argument, because the whole apparatus above it — z-scores, conviction, severity — was built to grade single observations, and no amount of single-observation confidence substitutes for repetition.

• Escalation follows persistence: a conflict type recurring across reviews earns a diagnostic, not a comment.
• The most common conflict is the system's failure signature — knowing yours tells you where to look first, always.
• A shifting last-four pattern (alignment → transition → under-monetization) is often the earliest structural warning available.

The routing

Persistent contradictions feed the stack; they never command it.

Confirmed persistent divergences route by type into the machinery already built for them. Chronic under-monetization is a capture question, routing to the MAE/MFE lab's giveback and capture reads and the variant attribution's checkpoint battery. Chronic profit-without-efficiency is a sizing and aggression question, routing to tier attribution and the override record. Chronic dual weakness is the only one that questions the edge itself, routing to the edge-change evidence threshold where it serves as structural corroboration. And through all of it, the SDE's standing boundary holds: these diagnostics can influence confidence, review priority, and risk posture, but final deployment authority remains with the gate and brake rules — a contradiction engine reading, however persistent, never moves a cap.

The key idea

Counting is the cheapest structural intelligence in the system.

Everything upstream of the counters is sophisticated — normalized drift, conviction grading, severity bands, a sixteen-cell meaning matrix. The counters are just tallies. And they carry a disproportionate share of the engine's real value, because the question that actually drives action is rarely 'how confident is this reading?' and almost always 'is this what my system does now?' Five numbers and a four-week sequence answer that, week after week, at essentially no cost.