Official-test scalar attribution inventory

Research Purpose And Boundary

This part preserves the individual scalar attribution and correlation inventory plus the held-out signed feature-mechanism continuation for the frozen Task 43 raw-inclusive hybrid. The root reference owns the benchmark claim, component-head performance, headline mechanism findings, mapping contrast, population stability, and external comparison. This part owns the 37 scalar rows and their measured overlap on the fixed 5,042-row training matrix, then the detailed scalar/family pressure and residual evidence on the retained 2,567-row official test.

Mean absolute TreeSHAP measures use by the existing nonlinear ensemble. The Pearson values measure association. Neither quantity is an additive QWK contribution, causal effect, or feature-removal result.

Evidence And Sources

Findings And Interpretation

Four scalar features share the accepted DeBERTa runtime lineage: sent_similarity_variance, prompt_similarity, intro_prompt_sim, and min_para_relevance. The remaining 33 scalars do not consume that embedding runtime.

Construct family Scalar feature Mean absolute TreeSHAP Max |r|
Correction/error grammar_errors_per_100_words 0.001296 0.500
Correction/error language_tool_spelling_errors_per_100_words 0.017286 0.660
Correction/error final_spelling_issues_per_100_words 0.002690 0.678
Correction/error punctuation_errors_per_100_words 0.000000 0.281
Readability smog 0.000183 0.329
Readability coleman_liau 0.001813 0.584
Readability ari 0.003098 0.383
Readability avg_sentence_length 0.001825 0.428
Lexical statistics ttr 0.000103 0.660
Length word_count 0.052873 0.378
Length avg_word_length 0.001016 0.591
Syntax parse_tree_depth 0.000403 0.472
Syntax passive_ratio 0.000367 0.218
Syntax dep_distance 0.000984 0.424
Semantic coherence, embedding-derived sent_similarity_variance 0.000377 0.161
Prompt relevance, embedding-derived prompt_similarity 0.000761 0.305
Prompt relevance, embedding-derived intro_prompt_sim 0.000165 0.240
Prompt relevance, embedding-derived min_para_relevance 0.000000 0.286
Complex nominal procedural_or_event_head_complex_np_density_log1p 0.000095 0.264
RST relation tier3_rst_gum_primary_tree_satellite_relation_rate 0.001632 0.332
Clause rates finite_advcl_per_100w 0.000885 0.259
Clause rates finite_relative_clause_per_100w 0.000128 0.234
Clause rates nonfinite_participial_advcl_per_100w 0.000321 0.205
Clause rates nonfinite_participial_acl_per_100w 0.000000 0.172
Frequency/sophistication p_lex_lambda 0.000175 0.423
Frequency/sophistication advanced_guiraud 0.018843 0.487
Word-range dispersion mean_log_contextual_diversity 0.000213 0.466
Word-range dispersion mean_contextual_diversity_percent 0.000000 0.496
Word-range dispersion contextual_diversity_coverage 0.000000 0.352
Moving-average lexical diversity mattr_50 0.000000 0.566
Concreteness mean_concreteness 0.000390 0.461
Concreteness concreteness_coverage 0.000511 0.304
Word prevalence mean_word_prevalence 0.000231 0.257
Word prevalence word_prevalence_coverage 0.000507 0.283
Lexical decision mean_lexical_decision_rt_ms 0.000728 0.424
Lexical decision lexical_decision_rt_coverage 0.002162 0.252
Causal-predicate complex nominal abstract_np_with_causal_predicate 0.000376 0.244

No scalar is constant, and no two scalar columns are exact duplicates. Eight scalar pairs have absolute Pearson correlation at least 0.8:

Scalar pair r
ari / avg_sentence_length 0.988
mean_log_contextual_diversity / mean_contextual_diversity_percent 0.953
coleman_liau / avg_word_length 0.945
concreteness_coverage / word_prevalence_coverage 0.936
language_tool_spelling_errors_per_100_words / final_spelling_issues_per_100_words 0.935
ari / parse_tree_depth 0.833
avg_sentence_length / parse_tree_depth 0.832
p_lex_lambda / mean_log_contextual_diversity -0.813

The high correlations include shared denominators and overlapping constructs, but correlation alone does not establish that one feature is removable. The TreeSHAP values likewise describe the fitted ensemble on this matrix; they do not report performance after removing a feature.

Held-Out Signed Attribution And Residual Alignment

The continuation reuses the five frozen historical 805-column members on the retained official-test matrix. The then-current Task 44 806-column source contained l2_dictionary_corrections_per_100_words; the projection omits it only because the historical learner never received it. TreeSHAP reconstructs the ensemble within 0.00000572, and the recomputed ensemble differs from the retained raw predictions by at most 0.00000024. The common base value is 3.178818.

The target strata contain 2,567 full, 225 low, 1,917 mid, 425 broad-high, and 68 exact-high rows. Every named scalar and accepted family retains support, mean signed and absolute contribution, percentile 95% intervals from 2,000 deterministic row-bootstrap replicates, tail-minus-mid differences, and raw and mapped residual alignment. These are descriptive intervals, not prompt-clustered population inference. Signed values below are grade-scale prediction pressure relative to the model base value.

Feature or accepted family Full Low 1.0-2.0 Mid 2.5-3.5 Broad high 4.0-5.0 Exact high 4.5-5.0
DeBERTa embedding family -0.021648 -0.802622 -0.067613 +0.599137 +0.916769
word_count -0.013093 -0.034154 -0.016584 +0.013804 +0.025033
Length statistics -0.013195 -0.034613 -0.016719 +0.014041 +0.025615
Correction and error rates -0.004429 -0.016542 -0.006222 +0.010068 +0.013411
Frequency and sophistication -0.008734 -0.011005 -0.010266 -0.000625 +0.013919
Prompt relevance, embedding-derived -0.000010 +0.000251 -0.000019 -0.000108 -0.000194

Low-band length pressure is downward on average, despite raw overprediction of +0.341438. The low-minus-mid length-family difference is -0.018060, 95% interval [-0.026573, -0.009887]. Length is positive on 39.1% of low rows; within those rows, error pressure is negative in 81.8% but fully offsets length in 31.8%. The longest/high-error low cell (n=39) has length +0.043579, error -0.020855, joint pressure +0.022725, and raw residual +0.379217. This is a bounded local length-overpressure pattern, not an aggregate low-band explanation.

Broad-high and exact-high raw residuals are -0.291574 and -0.436538; 332/425 and 64/68 rows are underpredicted. Frequency/sophistication is nearly neutral in broad-high and small positive in exact-high. Correction/error pressure is positive in both high views, not downward, but too small to close the residual. Prompt-relevance pressure is negligible and slightly downward, with both intervals crossing zero. Embeddings supply the main upward pressure, yet the final raw predictions remain low.

“Overweighted” and “underweighted” are residual-alignment interpretations here, not causal weight estimates. TreeSHAP does not show what a refit, feature removal, or different loss would do. The three-by-three word-count/error table reports observed main-effect cells only; no 805-by-805 or 806-by-806 SHAP interaction tensor was computed.

Evidence Gaps And Follow-Up

No matched per-feature removal fits were run. They were outside Task 03 and would add 37 post-hoc model-selection comparisons. The three declared reduced heads and the root reference's family attribution are the bounded component evidence for this story.