{ "task": "NEW-1 pooling-operator vs encoder isolation for the no-binding theorem", "backbone": "sentence-transformers/all-mpnet-base-v2", "bonus_cls_encoder": "BAAI/bge-base-en-v1.5", "n_sentences": 2000, "n_examples": 4000, "by_constr_examples": { "active": 800, "ditrans": 800, "fronted": 800, "passive": 800, "relcl": 800 }, "guard_cos_operator_vs_mean": { "cos_cls_vs_mean": 0.8784728050231934, "cos_last_vs_mean": 0.8693444728851318, "cos_cls_vs_last": 0.9117629528045654, "note": "cos << 1 => operator carries info distinct from the mean (non-degenerate)." }, "preregistered": { "mean_reproduces_no_binding": "mean-pool cross-construction thematic AUC <= 0.61", "pooling_theorem_if": "CLS or last-token cross-construction thematic AUC > 0.70 => Law 2 is a MEAN-POOLING theorem", "falsifier_if": "ALL operators <= 0.61 => no-binding is an encoder/training property, deeper than pooling" }, "surface_baseline": { "within_each_cv": { "active": { "linear": 1.0, "mlp": 0.836775 }, "ditrans": { "linear": 0.9995499999999999, "mlp": 0.93225 }, "fronted": { "linear": 0.9997250000000001, "mlp": 0.98280625 }, "passive": { "linear": 1.0, "mlp": 0.8163625000000001 }, "relcl": { "linear": 0.9258250000000001, "mlp": 0.9129375 } }, "cross_construction": { "linear": { "passive": 0.0, "relcl": 0.9273250000000001, "ditrans": 0.99914375, "fronted": 0.0, "POOLED_nonactive": 0.480773828125 }, "mlp": { "passive": 0.25751250000000003, "relcl": 0.8426875, "ditrans": 0.8655437500000002, "fronted": 0.10442500000000002, "POOLED_nonactive": 0.4946173828125 } }, "within_active_cv": { "linear": 1.0, "mlp": 0.836775 } }, "operators": { "mean": { "within_active_cv": { "linear": 0.6548375000000001, "mlp": 0.6864000000000001 }, "within_each_cv": { "active": { "linear": 0.6548375000000001, "mlp": 0.6864000000000001 }, "ditrans": { "linear": 0.8711125, "mlp": 0.845525 }, "fronted": { "linear": 0.9804187500000001, "mlp": 0.9753999999999999 }, "passive": { "linear": 0.5893125, "mlp": 0.5954375 }, "relcl": { "linear": 0.8155875, "mlp": 0.8740125 } }, "cross_construction": { "linear": { "passive": 0.458146875, "relcl": 0.5148312500000001, "ditrans": 0.6976812499999999, "fronted": 0.315625, "POOLED_nonactive": 0.500084375 }, "mlp": { "passive": 0.48049375, "relcl": 0.6474999999999999, "ditrans": 0.7178625000000001, "fronted": 0.3933625, "POOLED_nonactive": 0.5506796875 } } }, "cls": { "within_active_cv": { "linear": 0.633928125, "mlp": 0.6506375 }, "within_each_cv": { "active": { "linear": 0.633928125, "mlp": 0.6506375 }, "ditrans": { "linear": 0.8680374999999999, "mlp": 0.83453125 }, "fronted": { "linear": 0.98, "mlp": 0.97469375 }, "passive": { "linear": 0.56188125, "mlp": 0.5670062499999999 }, "relcl": { "linear": 0.8091718749999999, "mlp": 0.8657718750000001 } }, "cross_construction": { "linear": { "passive": 0.47651249999999995, "relcl": 0.53600625, "ditrans": 0.7081468750000001, "fronted": 0.268915625, "POOLED_nonactive": 0.5063806640624999 }, "mlp": { "passive": 0.4758375, "relcl": 0.700725, "ditrans": 0.71361875, "fronted": 0.35950000000000004, "POOLED_nonactive": 0.5515187500000001 } } }, "last": { "within_active_cv": { "linear": 0.64555, "mlp": 0.6686562500000002 }, "within_each_cv": { "active": { "linear": 0.64555, "mlp": 0.6686562500000002 }, "ditrans": { "linear": 0.86766875, "mlp": 0.84595 }, "fronted": { "linear": 0.978665625, "mlp": 0.9749 }, "passive": { "linear": 0.563221875, "mlp": 0.587625 }, "relcl": { "linear": 0.80785625, "mlp": 0.8823656249999999 } }, "cross_construction": { "linear": { "passive": 0.461471875, "relcl": 0.51744375, "ditrans": 0.6858906249999999, "fronted": 0.30709375, "POOLED_nonactive": 0.5020439453125001 }, "mlp": { "passive": 0.4782687499999999, "relcl": 0.73876875, "ditrans": 0.7032749999999999, "fronted": 0.338225, "POOLED_nonactive": 0.559987890625 } } } }, "bonus_cls_trained_encoder": { "within_active_cv": { "linear": 0.682478125, "mlp": 0.6686312499999999 }, "within_each_cv": { "active": { "linear": 0.682478125, "mlp": 0.6686312499999999 }, "ditrans": { "linear": 0.874225, "mlp": 0.8711187499999999 }, "fronted": { "linear": 0.97695, "mlp": 0.9747750000000001 }, "passive": { "linear": 0.566890625, "mlp": 0.57204375 }, "relcl": { "linear": 0.805015625, "mlp": 0.8629 } }, "cross_construction": { "linear": { "passive": 0.4721375, "relcl": 0.6523000000000001, "ditrans": 0.686903125, "fronted": 0.324421875, "POOLED_nonactive": 0.5425126953125 }, "mlp": { "passive": 0.4706375, "relcl": 0.7557375, "ditrans": 0.7462, "fronted": 0.48858124999999997, "POOLED_nonactive": 0.6027312499999999 } }, "cos_cls_vs_self_mean_guard": 0.9725103974342346 }, "scored": { "mean_cross_pooled_best": 0.5506796875, "cls_cross_pooled_best": 0.5515187500000001, "last_cross_pooled_best": 0.559987890625, "bonus_cls_trained_cross_pooled_best": 0.6027312499999999, "surface_cross_pooled_linear": 0.480773828125, "surface_within_active_linear": 1.0, "mean_reproduces_no_binding": true, "nonpooled_binds_pooling_theorem": false, "all_operators_stay_low_encoder_property": true }, "landing": "ENCODER/TRAINING PROPERTY: ALL operators (mean, CLS, last-token, CLS-trained retrieval) stay <= 0.61 cross-construction => no-binding is NOT about pooling. It is a deeper property of sentence-embedding encoders/training. This RESCOPES the theory frame: the mean-pool generative model is a convenient parameterization, but the absence of abstract agent/patient binding is intrinsic to the encoder, not the pooling operator." }