Skip to content

V70: twin columns are flagged, not removed - #179

Merged
dapiced merged 1 commit into
mainfrom
dom/modest-darwin-4oe9q0
Nov 19, 2026
Merged

dapiced merged 1 commit into
mainfrom
dom/modest-darwin-4oe9q0

Conversation

@dapiced

@dapiced dapiced commented Nov 19, 2026 •

Copy link
Copy Markdown
Owner

What

G2. Two columns that say the same thing were trained side by side with no word about it. On titanic: class and pclass, embarked and embark_town, adult_male read off who. Twins split the importance and the Shapley values between them, and V68 proposed a passenger who cannot exist by changing one without the other.

The functional-dependency test the lab already ran from a column to the target (leakSuggestions) now runs between columns, when the file is read:

  • each value of A goes with one value of B, both ways: « says the same thing as »;
  • one way only: « can be read off »;
  • two continuous numbers: the near-perfect correlation of the leak test.

A twin says so on its card and is left in. A redundancy is a note, not an exclusion reason: which one to keep is the user's call. A counterfactual that touches one twin without the other says so, naming both.

Measured before the interface

On the ten demo files: the three pairs of titanic plus the alive leak, and nothing on the nine other files.

Titanic, ten seeds, eight families Mean accuracy
the thirteen columns 0.813
without class, embark_town, adult_male 0.814
without those and who 0.805

No score moves when the twins go: this is about reading a model, not its performance. With the thirteen columns, 296 of 695 first counterfactuals (43 %) changed one column of a pair alone; with the twins excluded, none could.

Defect of this wave's own making

The first version read every cell again for every pair: 34 s on 200 columns and 100,000 rows. Each column is now read once into integer codes: 1.2 s on the same file, the same pairs on every demo.

Not in this wave

Excluding a twin or choosing which to keep; a fact carried by three columns together; an approximate redundancy.

Note

This description briefly announced V71 as well: a second commit was pushed to the branch a few minutes after this pull request had been merged. It was never part of this merge. V71 has its own pull request.

🤖 Generated with Claude Code

https://claude.ai/code/session_01AyBsUet8WybTG986ubroar

Two columns that say the same thing were trained side by side with no
word about it: class and pclass, embarked and embark_town, adult_male
read off who on titanic. Twins split the importance and the Shapley
values between them, and let a counterfactual change one without the
other.

The functional-dependency test the lab already ran from a column to the
target now runs between columns, when the file is read. A twin says so
on its card, both ways or one, and is left in: a redundancy is a note,
not an exclusion, and which one to keep is the user's call.

Measured first on the ten demo files: the three pairs of titanic and the
alive leak, nothing elsewhere. Removing the twins moves no score (0.813
against 0.814 of mean accuracy over ten seeds). With them in, 43 % of
the first counterfactuals changed one column of a pair alone; such an
answer now says so.

Docs in both languages, README, PLAN row V70, version 1.70.0.

Co-Authored-By: Claude Fable 5.2 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01AyBsUet8WybTG986ubroar
@dapiced
dapiced marked this pull request as ready for review November 19, 2026 22:51
@dapiced
dapiced merged commit 8b6a328 into main Nov 19, 2026
7 checks passed
@dapiced dapiced changed the title V70: twin columns are flagged, not removed V70 and V71: twin columns flagged, filled cells counted Nov 19, 2026
@dapiced dapiced changed the title V70 and V71: twin columns flagged, filled cells counted V70: twin columns are flagged, not removed Nov 19, 2026
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Labels

None yet

Projects

None yet

Development

Successfully merging this pull request may close these issues.

2 participants