Decision log
← ARI3 overview
Decisions that shape what ARI3 collects, learns from and claims, newest first. Published decisions run through 2026-09-26. A decision is replaced by a new entry, never edited.
- D-092026-09-26
Keep ambiguous lexicon terms, but do not count them
- Decision
- Lexicon v1 keeps all 73 terms the editor reviewed. Seven words that often mean something else (western, brat, demure, lace, bow, glamour, sustainable) are kept in the lexicon and their matches are stored, but they are excluded from every count.
- Why
- Dropping them would hide real style uses. Counting them blindly would count the wrong meanings.
- Alternatives
- Drop the ambiguous terms, or count them without notes.
- Tradeoff
- Real style uses of those seven words go uncounted until a way to tell the meanings apart exists.
- D-082026-09-26
Keep hard cases out of calibration as well as testing
- Decision
- Active-learning labels were used only for training in EXP-002.
- Why
- Split conformal prediction assumes calibration items look like test items. Items chosen for being hard do not.
- Alternatives
- Use them in calibration as their hash placed them.
- Tradeoff
- 7 labels went unused. The choice was not in the pre-registration and is disclosed as such.
See EXP-002.
- D-072026-09-24
Pre-register every experiment that compares versions
- Decision
- EXP-002's question, hypotheses and pass criteria were committed publicly before any of its labels existed.
- Why
- Criteria chosen after seeing results can make any result look good.
- Alternatives
- Report results and explain them afterwards.
- Tradeoff
- Weak or failed results have to be published as they are.
See EXP-002.
- D-062026-09-23
Beauty is separate from style, nails count
- Decision
- Makeup, skincare, fragrance, hair and beauty packaging are labeled not style. Nails are style.
- Why
- Beauty trade coverage caused half of v0.0.1's test mistakes, so the boundary had to be written down before more labeling.
- Alternatives
- Count all beauty as style, or none of it.
- Tradeoff
- Beauty coverage is not measured until it gets a task of its own.
See EXP-002.
- D-052026-09-23
Fix each label's split when it is written
- Decision
- A SHA-256 hash of the item and task decides whether a label is for training, calibration or testing, and it never changes.
- Why
- A test item that ever reaches training makes every later score look better than it is.
- Alternatives
- Re-split randomly for each experiment.
- Tradeoff
- Split sizes are only approximately 65/15/20.
See EXP-001.
- D-042026-09-23
Collect every 4 hours
- Decision
- Feeds are read six times a day.
- Why
- Measured: 28 of 65 feeds held under 7 days of items, and the fastest held about 8 hours. Weekly collection would miss items.
- Alternatives
- Weekly, daily or every 12 hours.
- Tradeoff
- More requests to each outlet, kept polite by robots.txt checks and a one-second delay per host.
- D-032026-09-22
Every claim needs a dated article link
- Decision
- A report cannot be published if any claim lacks a link to a specific dated article, or links only to a homepage, or cites something published after the report date. Continuous integration enforces this.
- Why
- A reader has to be able to open the source behind each claim and check it.
- Alternatives
- Allow outlet-level citations.
- Tradeoff
- Fewer, slower reports.
- D-022026-09-22
Split what one person can build from what needs partners
- Decision
- Each capability is either Viable (buildable on public data by one person) or Frontier (needs private data, real-world experiments or a team). Moving one across the line needs its own recorded decision.
- Why
- The line stops the site from claiming capabilities the project cannot yet support.
- Alternatives
- One undivided roadmap.
- Tradeoff
- Frontier ideas wait, however interesting.
- D-01Standing rule
No scraping of social platforms
- Decision
- Social data may come only from official APIs or hand sampling. Runway images are linked, never stored.
- Why
- Scraping breaks platform terms and collects people's posts without their consent.
- Alternatives
- Scrape public posts.
- Tradeoff
- Social signals are missing until API access is granted.