We scored our own website before redesigning it once, as a one-off discipline to replace a gut call with a number. That number has now become a standing weekly job, and it does more than grade. It decides what the site is allowed to fix about itself, and what it is not.
The scorecard
The self-score runs against six fixed axes, one per Hexagram pillar, and writes its result to a running history rather than a one-off report. The run from this week put the site at 88 overall, with Architect and Intelligence both at 100, Resonance at 100, Signal at 75, Vision at 67, and Conversion not yet measured. Vision is this week's bottleneck, with four open items: page meta length, post meta length, type scale, and label case. Those four, not a feeling about how the site looks, are what decides what gets worked on next.
That is a small, legible number, and that is the point. A score you cannot argue with is more useful than an opinion you can.
What is allowed to fix itself
Only three kinds of change may merge without a person looking at them: an internal link target corrected, alt text added or reworded, or a post's search title or description shortened to fit. Each of those comes with its own hard limit, at most five files and twenty changed lines, every line an edit rather than an addition, and each one is checked against the self-score afterwards to confirm the relevant check actually improved.
Everything else, which is nearly everything, opens a pull request instead: any copy, headline or claim, any figure or currency symbol, layout or design tokens, new or deleted files, and anything on the homepage, pricing, services, FAQ, case studies or team pages. Legal pages, forms, API routes, the database schema, and anything that collects, stores, tracks or emails a visitor's data are named explicitly, so a classifier parsing the diff cannot wave them through by accident. A change classified as "review" waits, however small it looks, and the classifier decides that, not a person's judgement on a given day.
Why the fence is this narrow
The policy that sets these limits says it plainly: this site sells the idea that AI agents act inside rules a person set, and that some decisions are not theirs to make alone. A wide automatic lane would move faster. One wrong merge to the homepage or the Diagnostic costs more than a week of waiting ever would. Widening the lane later is itself a reviewed change to the policy and the classifier together, once a kind of fix has actually proved safe, not a judgement call made in the moment it would be convenient.
Nothing gets to claim it worked without saying how it knows
Every change, low-risk or not, gets a changelog entry stating the problem, the metric that will show whether it worked, and a pending result. Fourteen days later, that entry gets settled: kept, reverted, or inconclusive, with an honest account of what the metric actually did. A change whose own target metric got worse is reverted, not left in place because reverting it would be awkward. Settling an entry edits something already public, so that step is never allowed to merge itself either, exactly like a new claim would not be.
Our traffic to this site is small, tens of Diagnostic visits in a typical week, and the routine is built around that fact rather than around pretending otherwise. No rate or result gets reported from fewer than twenty visits, because a number that small moving is usually noise wearing the costume of a trend. A business running on real but modest traffic should hold its own experiments to the same floor before calling anything a result.
What this replaces
The old version of this discipline was a single AI-assisted score, run once, used to brief a redesign. That privacy policy correction we published this same week is the clearest proof of why the fence sits where it does: the self-score itself gained four new checks because of it, each one built to fail if a tracker loads without consent or a marketing email ships without an unsubscribe link. A score that only measures what already looks fine is not a score. It is decoration.
If your own site runs changes by vibe rather than by a number it cannot talk its way around, that is precisely the kind of structural gap the Intelligence pillar of the Hexagram is built to surface, and it is part of what our analytics and AI work looks at directly. I cover the measurement side of the team bench; the Hexagram Diagnostic is where you find out what your own bottleneck pillar actually is.
