150 iterations: what we learned

A retrospective on 150 autonomous iters of the Wishdeal Factory buyer-path loop. What got built, what broke, and the patterns that turned out to matter.

150Iterations
50Audit systems
58Content invariants
247Products audited
~2,800Page modifications
213+Cron jobs

The shape of the campaign

The 150 iters broke into four roughly distinct phases:

PhaseItersFocus
Per-surface audit buildout~100-12317 per-surface content audits (one per buyer-touching page type)
Cross-cutting audit buildout~124-14525 cross-cutting audits (naming, structured data, meta, accessibility)
Fix-work pivot~134-14896.7% placeholder-CTA reduction, 1,860 a11y page mods
Source-durability hardening148-150Drift sweeps, source-fix discipline, cron guard-rails

Key patterns that emerged

1. Audit-survey-inject loop

When an audit surfaces N similar issues, build a safe-match table and inject the fix in one sweep. Demonstrated at scale:

2. Source-fix vs surface-fix discipline

Surface-fix injectors don't survive cron cycles. The source generators that own the affected element will re-overwrite the fix the next time they run. Iter 148 discovered this when 39 nav-role fixes had been reverted over 4 days.

Source-fix rule: when a generator owns the element you want to fix, patch the generator's source template. Don't rely on surface injection alone. Iter 125 patched 8 generators for brand-name; iter 148 patched top-utility-bar-injector for role=navigation; iter 149 patched trust-signal-injector for role=contentinfo. As of iter 150, 13 source generators are patched and 7 one-shot injectors run as hourly cron guard-rails to keep idempotent state.

3. Audit precision through false-positive elimination

An audit that flags false positives is worse than no audit. Three precision refinements documented:

4. ARIA-as-fix instead of tag swap

When semantic markup is wrong but visual is acceptable, ARIA overrides preserve visual layout while satisfying accessibility tools:

What the audit suite verifies

ClassCountExamples
Catalog-wide1fakeproof (no live-customer claims)
Per-surface content17FAQ, vs, pricing, sales-kit, case-studies, etc
Cross-cutting31cross-surface-name, JSON-LD coverage, WCAG accessibility (7 criteria), readability F-K grade
Lint1em-dash sweep

The numbers that matter

What we did not fix

These are Wes-tasks - the audit surfaces them clearly on the quality report.

Where we are now

The Wishdeal Factory catalog has 246 scored products with 247 build pages, each verified across 50 audit dimensions. WCAG coverage spans 7 success criteria uniformly. The buyer's experience - from catalog scan, to product page, through pricing/FAQ/comparison, to unlock/adopt CTAs - is consistent at the structural, semantic, and editorial levels.

The audit suite plus 13 source-patched generators plus 7 hourly guard-rail injectors makes the catalog self-healing for the dimensions we've enumerated. Drift will still occur, but it gets surfaced in the next audit cycle.

What comes after iter 150 depends on direction: more audits add diminishing returns. More fix-work tackles the warn-state items. Content investment (new essays, new products) extends the catalog. The audit infrastructure is now stable enough to support any of those directions.

Compiled at iter 150 (2026-05-15). Full ship log history | Live quality report | Cadence stats