Survival of Fitness

MRI-002

Method / Data / Limitations

Category Capture

Back to MRI-002

2026 Current Snapshot

  • 30 high-attention New York Times Opinion discussions.
  • Dates: August 17-25, 2026.
  • Every article had at least 500 displayed comments.
  • 297 captured Reader Picks.
  • 29 articles contributed 10 Reader Picks; ATT-014 contributed 7.
  • Participation = displayed comment count.
  • Recognition ranking in the 2026 visual = sum of recommendation counts among the captured Reader Picks for each article.
  • This is not all recommendation activity across all comments, and it is not representative of all New York Times readers.

Historical Corpus

  • 2,118,617 normalized comments.
  • 9,298 unique article metadata records.
  • 9,424 discussion/article IDs in comments.
  • 1,550,491 top-level comments.
  • 568,126 replies.
  • 2,086,862 comments linked to article metadata.
  • 31,755 unmatched to metadata.
  • Normalized comment timestamp range: January 1, 2017 through May 2, 2018.
  • 944 discussions were flagged high-engagement in the full discussion table.
  • 933 high-engagement discussions met the full calibration/holdout eligibility rules and entered the frozen validation population.
  • The 933 eligible discussions split into 120 calibration discussions and 813 holdout discussions.

Woke Lexical Audit

  • Exact word-boundary search for woke and wokeness.
  • Case-insensitive matching.
  • Ordinary wake/sleep verb uses were excluded.
  • Candidate comments were manually reviewed using a transparent rule.
  • 1,437 exact candidate comments.
  • 1,494 exact matched tokens.
  • 391 retained categorical comment-level uses.
  • 418 categorical matched tokens.
  • Retained date span: January 4, 2017 to May 1, 2018.
  • Uses were mixed from the beginning.
  • No linear semantic progression is claimed.
  • Raw monthly counts were not normalized by all monthly comments, so no prevalence comparison is made.
  • Usage groups: political behavior / awareness; positive self / ally; cultural / ambiguous; meta / term discussion; negative / pejorative; identity / group label.
  • Public aggregate monthly counts CSV

The reviewed row-level Woke categorical-use table and Markdown audit summary remain in the local research folder unless explicitly approved for publication.

Calibration

  • 120 high-engagement discussions were selected reproducibly.
  • The first 25 top-level comments represented the early field.
  • Future top-10 recognition = the highest-recommendation top-level comments after comment 25.
  • Earlier work revealed a major early-position recognition confound.
  • The future-winner definition excluded the first 25 comments to prevent leakage.

Matching

  • For each future winner, three ordinary controls were selected from the same discussion.
  • Controls were matched as closely as possible on chronological order percentile.
  • Control selection used order position, time, and comment ID rather than recommendations or text.
  • 8,130 future winners.
  • 24,390 controls.
  • Median absolute order-percentile difference: 0.001757.
  • 90th percentile order-percentile difference: 0.004175.
  • 97.79% within +/-0.01.
  • 99.70% within +/-0.025.
  • 100% within +/-0.05.
  • No reused controls within article.
  • No controls from the first 25 comments.
  • No controls that were also future winners.

Text Representation

  • HTML and URLs removed.
  • Text lowercased.
  • TF-IDF with English stop words.
  • Word unigrams and bigrams.
  • min_df=2.
  • sublinear_tf=True.
  • max_features=50,000.
  • TruncatedSVD with 100 components.
  • random_state=42.
  • Vectors normalized.
  • Cosine similarity was calculated against the early-field representation.
  • This measures semantic/language alignment, not category capture directly.

Frozen Holdout

  • The method was frozen before holdout evaluation.
  • 813 holdout discussions.
  • 0 calibration article IDs in holdout.
  • Pairwise early-alignment win rate: 61.3571%.
  • Winner greater than mean of controls: 65.1415%.
  • Top-of-four accuracy: 36.2485%.
  • Chance baseline: 25%.
  • Discussions with positive mean winner-minus-control difference: 699 / 813.
  • Positive discussion direction: 85.9779%.

Length Robustness

Conditional logistic model: winner_status ~ standardized early alignment + standardized log1p word count.

  • Early alignmentcoef 0.135104
  • OR 1.144656.
  • 95% CI 1.103629-1.187208.
  • p 4.0253e-13.
  • Word countcoef 0.607256
  • OR 1.835388.
  • 95% CI 1.767831-1.905526.
  • p 4.84793e-221.
  • Word count is a substantially stronger recognition predictor in this model.

Limitations

  • New York Times commenters are self-selected.
  • Reader Picks are not a census of readers.
  • Captured Reader Picks do not represent every recommendation action.
  • Semantic alignment is broader than category alignment.
  • This is observational; it does not support causal inference.
  • Examples are interpretive illustrations.
  • The class-portability hypothesis is prospective and unconfirmed.
  • Woke historical uses were mixed from the beginning.
  • The Woke audit does not establish a linear change in meaning.
  • Monthly Woke counts are raw counts, not normalized prevalence rates.
  • The current sample is small and restricted to high-attention Opinion discussions.
  • Editorial-selection effects are not tested here.
  • MRI-002 does not establish that New York Times comments cause wider cultural change.

Data / Code

Source corpus, derived results, audits and reproducibility files are available here.

Data / code →