Skip to content
中文

Explore pi-rsi airline experiments and their research Wiki

PocketPlay Journal2026-10-04Huanyi Xie2 min read

Open recorded score curves, search branches and knowledge revisions; follow supporting and opposing evidence without treating the highest score as independent confirmation.

Explore the recorded airline-satisfaction search in three interactive views: score progression and branches, the full search tree, and the research Wiki.

Read the scores

The experiment snapshot records a baseline ROC-AUC of 0.9572004066 and a highest valid completed score of 0.9579240131 at n016, a difference of 0.0007236065. Click a node to inspect its result and inherited branch. Failed or invalid measurements cannot improve the best-score curve.

The n016 Depthwise configuration is only 0.0001394425 above the separate SymmetricTree reference, below its prespecified 0.0002 practical screen. Numerical rank does not establish repeatability or statistical significance. The final holdout remains sealed; these are adaptively reused local search-validation scores, not Kaggle leaderboard results.

Follow a knowledge entry

Select an observation, hypothesis or experience entry. Follow the separate supporting and opposing records to inspect a recorded experiment score or documentation source. Expand scope, alternatives and the next test before deciding whether a claim applies to another setting.

The version selector shows published before/after statuses and the recorded reason for a change. “Supported” is a scoped Wiki interpretation, not independent confirmation; multiple records can describe one experiment or source.

Understand the snapshots

The experiment and Wiki each display their own saved time. These pages do not poll the running campaign. Original experimental statements remain in their source language, while navigation and controls have English and Chinese versions.

The snapshot summary, score ledger, Wiki provenance and frozen protocol document what is shown. The public pages contain aggregate records, not dataset rows, held-out labels, prediction tables or model-session transcripts.

Sources & checks

How we write and check articles: Editorial method

Have a follow-up question?

Reviews, recommendations and red flags from people who tried it.

Open the community