analysis/L5_ANOMALY.md

L4 to L5: baseline-adjusted reanalysis

Raw level means include all domains. Adjusted means use only domains nonconstant at BOTH levels. All differences are proportions (multiply by 100 for percentage points). Bootstrap uses 10,000 domain-stratified IMAGE-cluster resamples; an image receives the same multiplicity at both levels. Empirical ground-truth baselines are held fixed, so CIs are conditional on the evaluated truth distribution. The constant-exclusion difference is descriptive, not a causal attribution.

model	n	baseline	raw_accuracy	adjusted_score	excluded_constant	small_n	paired_domains	raw_L4	raw_L5	adjusted_L4	adjusted_L5	n_L4_all	n_L5_all	n_L4_paired	n_L5_paired	baseline_L4_paired	baseline_L5_paired	raw_L4_paired	raw_L5_paired	adjusted_L5_minus_L4	ci_low	ci_high	verdict	raw_L5_without_constants	raw_rise_before	raw_rise_after	raw_rise_removed_by_constant_exclusion
claude_opus_5	3100	0.321290	0.454839	0.183569	False	False	31	0.402941	0.551765	0.149082	0.218056	1700	1700	1550	1550	0.245161	0.397419	0.379355	0.530323	0.068974	0.024170	0.113034	survives	0.530323	0.148824	0.127381	0.021442
claude_sonnet_5	3100	0.321290	0.327097	-0.043400	False	False	31	0.297647	0.383529	0.008937	-0.095736	1700	1700	1550	1550	0.245161	0.397419	0.285806	0.368387	-0.104672	-0.157881	-0.051941	reverses	0.368387	0.085882	0.070740	0.015142
deepseek_v4_1_flash	3100	0.321290	0.269032	-0.191828	False	False	31	0.238824	0.348235	-0.077344	-0.306313	1700	1700	1550	1550	0.245161	0.397419	0.227097	0.310968	-0.228968	-0.283913	-0.173435	reverses	0.310968	0.109412	0.072144	0.037268
gemini_3_8_flash	3100	0.321290	0.484194	0.219608	False	False	31	0.454118	0.550000	0.220583	0.218634	1700	1700	1550	1550	0.245161	0.397419	0.430323	0.538065	-0.001949	-0.049164	0.044648	indistinguishable from zero	0.538065	0.095882	0.083947	0.011935
gpt_5_6_luna	3100	0.321290	0.263226	-0.169157	False	False	31	0.238824	0.340000	-0.074246	-0.264069	1700	1700	1550	1550	0.245161	0.397419	0.226452	0.300000	-0.189823	-0.247751	-0.131252	reverses	0.300000	0.101176	0.061176	0.040000
gpt_5_6_sol	3100	0.321290	0.329677	-0.033398	False	False	31	0.305882	0.404706	0.030280	-0.097077	1700	1700	1550	1550	0.245161	0.397419	0.291613	0.367742	-0.127357	-0.182329	-0.075919	reverses	0.367742	0.098824	0.061860	0.036964
grok_4_6	3100	0.321290	0.372903	0.057892	False	False	31	0.326471	0.463529	0.080726	0.035059	1700	1700	1550	1550	0.245161	0.397419	0.311613	0.434194	-0.045667	-0.091132	0.000745	indistinguishable from zero	0.434194	0.137059	0.107723	0.029336
inking	3100	0.321290	0.247097	-0.222736	False	False	31	0.232353	0.321176	-0.079506	-0.365966	1700	1700	1550	1550	0.245161	0.397419	0.210323	0.283871	-0.286461	-0.343340	-0.229998	reverses	0.283871	0.088824	0.051518	0.037306
muse_glimmer_30b	3100	0.321290	0.279677	-0.194990	False	False	31	0.258235	0.334706	-0.049386	-0.340594	1700	1700	1550	1550	0.245161	0.397419	0.246452	0.312903	-0.291208	-0.343732	-0.239474	reverses	0.312903	0.076471	0.054668	0.021803
perplexity_sonar_pro	3100	0.321290	0.232258	-0.215963	False	False	31	0.207059	0.313529	-0.136017	-0.295909	1700	1700	1550	1550	0.245161	0.397419	0.181935	0.282581	-0.159891	-0.215411	-0.106239	reverses	0.282581	0.106471	0.075522	0.030949

• Claude Opus 5: the adjusted L4-to-L5 rise survives (difference +6.90 pp; 95% CI +2.42 to +11.30 pp).
• Claude Sonnet 5: the adjusted L4-to-L5 rise reverses (difference -10.47 pp; 95% CI -15.79 to -5.19 pp).
• DeepSeek V4.1 Flash: the adjusted L4-to-L5 rise reverses (difference -22.90 pp; 95% CI -28.39 to -17.34 pp).
• Gemini 3.8 Flash: the adjusted L4-to-L5 rise indistinguishable from zero (difference -0.19 pp; 95% CI -4.92 to +4.46 pp).
• GPT-5.6 Luna: the adjusted L4-to-L5 rise reverses (difference -18.98 pp; 95% CI -24.78 to -13.13 pp).
• GPT-5.6 Sol: the adjusted L4-to-L5 rise reverses (difference -12.74 pp; 95% CI -18.23 to -7.59 pp).
• Grok 4.6: the adjusted L4-to-L5 rise indistinguishable from zero (difference -4.57 pp; 95% CI -9.11 to +0.07 pp).
• Inkling: the adjusted L4-to-L5 rise reverses (difference -28.65 pp; 95% CI -34.33 to -23.00 pp).
• Muse Glimmer 30B: the adjusted L4-to-L5 rise reverses (difference -29.12 pp; 95% CI -34.37 to -23.95 pp).
• Perplexity Sonar Pro: the adjusted L4-to-L5 rise reverses (difference -15.99 pp; 95% CI -21.54 to -10.62 pp).