| { |
| "construct": { |
| "alpha": 0.01, |
| "construct_id": "summarization-bias-evaluative", |
| "construct_name": "Summarization bias (evaluative regime)", |
| "e_boundary": 400, |
| "extension_claim": "The same machinery run as a weaker claim: h >= 0.2, reported as an extension, never as the registered test.", |
| "extension_min_h": 0.2, |
| "human_baseline_rate": 0.5, |
| "human_role": "human", |
| "known_confounds": [ |
| "availability", |
| "position", |
| "familiarity", |
| "length", |
| "framing", |
| "density", |
| "register" |
| ], |
| "model_role": "model", |
| "per_tail_alpha": 0.0025, |
| "primary_metric": "told-preference rate (share of paired judgements preferring the told member)", |
| "primary_metric_direction": "told", |
| "registered_claim": "Model told-preference exceeds the human baseline by Cohen's h >= 0.3 (or odds ratio >= 2.0) on gate-passing pairs.", |
| "threshold_cohens_h": 0.3, |
| "threshold_odds_ratio": 2.0, |
| "transformation": "shown-to-told: the same content rendered as concrete shown structure vs the abstract label that names it" |
| }, |
| "costs": { |
| "fixed_decisive_pairs": 396, |
| "fixed_standard_pairs": 175, |
| "old_coding_passes_per_pair": 4, |
| "old_judges_per_pair": 5 |
| }, |
| "human_layer": { |
| "audit_readers": 2, |
| "audit_roots": 200, |
| "audit_target_kappa": 0.6, |
| "coder_reconstruction_rate": 0.92, |
| "confirmatory_roots": 175, |
| "pilot_judgments_per_root": 2, |
| "pilot_roots": 100 |
| }, |
| "latent": { |
| "candidate_roots": 400, |
| "max_length_ratio": 1.05, |
| "max_null_overlap": 0.1, |
| "min_content_overlap": 0.8, |
| "min_length_ratio": 0.95, |
| "templates": [ |
| { |
| "content_tokens": [ |
| "river", |
| "mill", |
| "wheel", |
| "grain", |
| "stone" |
| ], |
| "id": "T1", |
| "shown_fragments": [ |
| "the river turns the old mill wheel over stone", |
| "grain waits in sacks beside the cold wheel" |
| ], |
| "told_fragments": [ |
| "the river mill processes grain with water power", |
| "the mill turns the grain into flour by stone" |
| ] |
| }, |
| { |
| "content_tokens": [ |
| "river", |
| "bridge", |
| "planks", |
| "rope", |
| "bank" |
| ], |
| "id": "T2", |
| "shown_fragments": [ |
| "planks sag on the rope bridge over the river", |
| "the near bank holds the rope fast" |
| ], |
| "told_fragments": [ |
| "the old bridge crosses the river on rope planks", |
| "the crossing structure stands at the bank" |
| ] |
| }, |
| { |
| "content_tokens": [ |
| "oven", |
| "bread", |
| "coals", |
| "flour", |
| "crust" |
| ], |
| "id": "T3", |
| "shown_fragments": [ |
| "coals glow under the oven while crust browns", |
| "flour dusts the board beside the oven" |
| ], |
| "told_fragments": [ |
| "the oven bakes bread from flour over coals", |
| "the process yields a browned crust daily" |
| ] |
| }, |
| { |
| "content_tokens": [ |
| "harbor", |
| "boats", |
| "tide", |
| "nets", |
| "gulls" |
| ], |
| "id": "T4", |
| "shown_fragments": [ |
| "the tide lifts boats while gulls circle the harbor", |
| "nets dry on posts above harbor mud" |
| ], |
| "told_fragments": [ |
| "the harbor shelters boats and nets at tide", |
| "a description of the working harbor and gulls" |
| ] |
| }, |
| { |
| "content_tokens": [ |
| "clock", |
| "gears", |
| "bell", |
| "tower", |
| "hands" |
| ], |
| "id": "T5", |
| "shown_fragments": [ |
| "gears turn the clock hands above the tower square" |
| ], |
| "told_fragments": [ |
| "the tower clock tells time by gears and a bell, and its mechanism coordinates the hands, the bell, and the hourly striking for the whole square below it" |
| ] |
| }, |
| { |
| "content_tokens": [ |
| "garden", |
| "bees", |
| "rows", |
| "soil", |
| "rain" |
| ], |
| "id": "T6", |
| "shown_fragments": [ |
| "bees work the garden rows after the rain", |
| "soil darkens where rain reached the rows" |
| ], |
| "told_fragments": [ |
| "the garden grows in rows tended after rain", |
| "bees pollinate garden plants in wet soil" |
| ] |
| } |
| ] |
| }, |
| "panel": { |
| "description": "The prespecified panel of the registered design: three model families at the full threshold lean, one aggregate human stream at baseline.", |
| "judges": [ |
| { |
| "baseline_rate": 0.5, |
| "lean_h": 0.3, |
| "name": "Model family 1", |
| "noise": 0.0, |
| "role": "model" |
| }, |
| { |
| "baseline_rate": 0.5, |
| "lean_h": 0.3, |
| "name": "Model family 2", |
| "noise": 0.0, |
| "role": "model" |
| }, |
| { |
| "baseline_rate": 0.5, |
| "lean_h": 0.3, |
| "name": "Model family 3", |
| "noise": 0.0, |
| "role": "model" |
| }, |
| { |
| "baseline_rate": 0.5, |
| "lean_h": 0.0, |
| "name": "Human panel (aggregate)", |
| "noise": 0.0, |
| "role": "human" |
| } |
| ], |
| "latent_correlation": 0.0, |
| "name": "Registered panel (as designed)" |
| }, |
| "seed": 20261021, |
| "stopping": { |
| "max_roots": 1000 |
| }, |
| "truth": { |
| "lean_scale": 2.0, |
| "model_rate_shift": 0.0, |
| "scenario": "registered_alternative" |
| }, |
| "type": "engine" |
| } |
