Experiment protocol: validating TACS and IPR

Status: draft. The protocol has not yet been registered on OSF Registries and has not been reviewed externally. This page will be updated with the pre-registration link once available. No data have been collected.

Design

A controlled experiment with 300 original images (60 per condition), each placed on its own page with an identical template and identical surrounding text. Conditions are assigned at random, stratified by image category. All pages are published on the same day.

Condition	File name	Alt text	Caption
C0 Control	Short descriptive	One short phrase	None
C1	TACS format	As C0	None
C2	As C0	TACS format	None
C3	As C0	As C0	TACS format
C4 Full TACS	TACS format	TACS format	TACS format

Hypotheses

H1 (confirmatory): C4 appears in the top 10 of image search results more often than C0. Minimum effect considered meaningful: 15 percentage points.
H2 (secondary): each single anchor raises the chance of appearing relative to C0, with Holm correction.
H3 (secondary): the reformulated IPR score predicts the chance of appearing, tested on held-out data not used to choose the formula's form.

Measurement

Primary outcome: whether the image appears in the top 10 image search results on day 28 after publication, for queries written by annotators who see only the image. Each query is repeated five times at different times, and outcomes are coded by raters who do not know the image's condition. Additional measurement points: days 7, 14, and 56. Searches are run in English, with interface language, region, and location recorded.

Bias control

The embedding model used to compute IPR differs from the evaluator model, to avoid circularity.
This research domain is used for all conditions and is separate from the author's business site.
One confirmatory test; all other analyses are labelled exploratory.
A 30-image pilot precedes the main experiment to check the baseline.

Transparency

After pre-registration, code, the per-image condition list, queries, and raw data will be released in an open repository with a DOI. All results are reported, including null or negative ones, and deviations from the protocol are recorded.