The film argues for where this goes. What follows is the smaller and more accurate account: the system has been built and run one reviewer at a time.
Three modes
Taste is a framework for understanding and evaluating creative sensibility. It runs in three modes, and each is useful on its own. Together they close a loop: one extracts the signal, one turns it into a profile, one holds new work up against it.
- Taste Builder
- Builds the profile, for a person or for a brand. Every claim about taste enters as a hypothesis and has to earn its evidence, with conservative priors so that a handful of comparisons can’t manufacture confidence. What comes out is a structured artifact rather than an opinion held firmly.
- Taste Check
- Reads a piece of work against an established profile. Does this fit the sensibility we defined, which preferences apply, and where is the work in tension with them?
- Taste Match
- Extracts the signal in the first place, through comparison. Two options, one pick, no explaining, because the explaining is the part people do badly. This is the one you can run yourself Opens in a new tab
What came out
The process has been run on working creative leaders, one at a time, and it is still being run. One vault holds fourteen active preferences, each traceable to specific evidence, each carrying a confidence value. They were more specific than I expected and considerably less flattering than a brand guideline: resisting category defaults, a discipline about where a piece of communication places its burden, a binary quality threshold with no partial credit. Several were things the reviewer had been enforcing for years without naming once.
A description of taste is a curiosity. A description you can review against is a tool.
What it taught me
- Forced choice beats self-report by a distance. People describe their taste inaccurately and demonstrate it consistently. Every attempt to shortcut the picking with a questionnaire produced worse output.
- Confidence has to travel with the claim. The moment a preference appears without its evidence count, someone treats it as policy.
- Proposing beats concluding. When I tested a version that stated conclusions outright, reviewers argued with it and disengaged.
On what’s shown here. The reviewers this has been run on aren’t named, and neither are the projects the evidence came from. The preferences above are paraphrased. The film uses an invented reviewer and a synthetic session. The method is the transferable part; a vault belongs to the person it describes.
Built in Python, FastAPI, React and SQLite.
Talk to the site
Ask it what this page left out.
A voice assistant runs on every page here. It reads a handful of files I wrote by hand, it won’t pretend to be me, and there’s a written list of things it refuses to discuss. Client specifics are on that list.
Needs a microphone.