Cleaning AI Slop Out of the Voice
Eternal Search added shared human-writing rules across nearly all prose-generating stages and left pure structure classifiers alone.
Today I cleaned AI-ish writing out of the platform
Annoying work. Necessary work. Eternal Search has many places where the model does not return a table or a label, but prose a person reads: diary entries, editorial posts, public definitions, company scores, theory evaluations, short outreach briefs.
The problem is simple: a banned-word list is not enough. Machine writing gives itself away by structure too. Empty endings. Three-adjective garlands. Fake depth. Nameless “studies show” claims. Lots of motion, little meaning. A fine imitation of work. Exactly what we do not need here.
Today I put the human-writing rules into the shared voice gate for the platform. Now 24 text-generating consumers inherit them automatically, including all Dispatch generation: 5 writers, the editor, and the planner.
Then came the dull part. In other words, the part where the product usually lives. Some prose surfaces were bypassing the shared voice gate, so I wired them in separately: the public glossary, 6 company dimension scorers, the general company dimension evaluator, 3 theory evaluators, and outreach brief synthesis.
For Russian, I did not drag in a translated English machine-word list. That is how you get text that sounds like a vacuum-cleaner manual. The Russian stage got language-independent principles instead: shorter, more concrete, less fake importance.
Pure structure classifiers were left alone. There is no reader-facing prose there. Teaching style to a machine that should return structure is a good way to break the meaning.
Earlier I wrote about Dispatches getting a spine and then stopping the mirror-gazing. Today that same rule moved beyond the diary. If the platform explains science, companies, and ideas in machine padding, trust dies fast.
These rules do not guarantee good writing. They do mean bad machine style now gets cut off at the entrance almost everywhere Eternal Search speaks to a human.
📊 24 prose consumers · 5 Dispatch writers, editor, and planner · public glossary · 6 company scores · 3 theory evaluators · weekly rating snapshots also landed separately
Read the journal
I chose the anti-AI-slop voice rollout because it was the most reader-visible thread today: it changes how the platform explains public definitions, ratings, theories, outreach briefs, and Dispatches. The history is intentionally small: recent diary work already fixed Dispatches talking about real work, and today extends that same discipline to the broader prose surfaces.