What does it mean for a researcher to accept a feature before it enters a study? This note describes that gate, using a sentiment score on Tesla headlines as the public example, and explains why a computed column is still only a proposal.
A column is not yet a feature
A pipeline can compute a number for every row. The number may be a score, a count, a rank, or a flag. Computation is not acceptance. Acceptance is a person's decision that this version of the number, built under a stated rule, is allowed to sit beside the other inputs of a study. Until that decision is recorded, the column is an experiment. Experiments are supposed to be easy to throw away. Features are supposed to be hard to change without a note.
The decision has content. It names the source, the version, the population, and the known failures. It says what the number is allowed to mean and what it is not allowed to mean. A later reader should be able to reject the feature by reading the note, without rerunning the pipeline. If the note is only a file name, there is nothing to accept or refuse. The gate is the sentence, not the column.
Accepting a feature is a written decision that a named version of a computed column may enter a named study, with its source, population, and known failures stated.
A score on Tesla headlines
Take headlines about Tesla as the example. A language model, or a simpler dictionary, can assign each headline a sentiment score. The score is a proposal about the headline's tone under that tool's definition of tone. It is not a measurement of the company, and it is not a measurement of the stock. Headlines are written by editors. They cluster on event days. They repeat one another. A daily average of scores will move because the mix of headlines moved, which may or may not have anything to do with a filing or a price.
Acceptance starts by reading a sample, not by plotting the average against a price. The sample should include a dull day, a delivery announcement, a recall, and a headline that uses the name in passing. The researcher writes down which of those the score handled in a way the study is willing to live with. If the passing mention is scored like a major event, the feature may still be accepted for a narrow use, or it may be refused. Either outcome is a result. Silence is not.
Versioning is part of the same gate. A model that is retrained or swapped will change scores on old headlines. The old accepted series and the new proposal are different features. Replacing the column in place erases the decision. The honest file keeps the accepted version frozen and stores the new proposal beside it until someone accepts that one too. Tesla is a useful example because the headline stream is public and loud. The loudness makes leakage and double-counting easier to see, not harder.
What acceptance does not authorize
Accepting a feature authorizes its use as an input to further research under the limits in the note. It does not authorize a live order. It does not say the score is accurate in a philosophical sense. It says a named person looked at a named version and allowed that version into a named study. Another study that wants a different population, such as filings rather than headlines, needs its own acceptance. The Tesla headline score does not travel with a new label.
Refusal is a normal outcome. A feature that fails the sample, or that cannot be rebuilt on a point-in-time basis, stays a proposal. The proposal can remain on disk. It does not enter the study's input list. The habit is the point. A computed column will always be available. A feature exists only after someone has agreed, in writing, to be responsible for what it is allowed to mean.
This piece does not claim that sentiment is useful, that headlines about Tesla are a signal, or that acceptance is a form of approval for trading. It separates a computed proposal from the decision to let that proposal into a study.
