manifest for l2nav1_w0_v1_full

eval/l2nav1_w0_v1_full.manifest.json

@@ -0,0 +1,274 @@
+{
+  "dataset": "l2nav1_w0_v1_full",
+  "prompt_rev": "l2v1",
+  "git": "f1279b6",
+  "task": "l2nav1",
+  "qtype": "nav_dist_rel",
+  "config": {
+    "world": "w0",
+    "dataset_name": "l2nav1_w0_v1_full",
+    "seed": 881001,
+    "room_size": 16,
+    "n_objects": 6,
+    "n_movers": 2,
+    "fov_deg": 75.0,
+    "render_res": 512,
+    "context_levels": [
+      5,
+      6,
+      7,
+      8
+    ],
+    "plan_lens": [
+      4,
+      5,
+      6
+    ],
+    "ranks": [
+      "nearest",
+      "2nd-nearest",
+      "3rd-nearest",
+      "3rd-furthest",
+      "2nd-furthest",
+      "furthest"
+    ],
+    "sectors": [
+      "front",
+      "behind",
+      "left",
+      "right"
+    ],
+    "sector_rule": "dominance",
+    "sector_margin_deg": 8.0,
+    "rank_gap_min": 1.5,
+    "wall_clearance_min": 3.0,
+    "pixel_confirm_min": 0.5,
+    "coverage_fov_deg": 75.0,
+    "rank_gap_near_max": 2.0,
+    "sector_margin_near_max_deg": 12.5,
+    "per_cell": 13,
+    "prompt_rev": "l2v1"
+  },
+  "quota": {
+    "per_cell_target": 13,
+    "total": 156,
+    "cells": {
+      "L4:behind": 13,
+      "L4:front": 13,
+      "L4:left": 13,
+      "L4:right": 13,
+      "L5:behind": 13,
+      "L5:front": 13,
+      "L5:left": 13,
+      "L5:right": 13,
+      "L6:behind": 13,
+      "L6:front": 13,
+      "L6:left": 13,
+      "L6:right": 13
+    },
+    "ranks": {
+      "nearest": 16,
+      "2nd-nearest": 31,
+      "3rd-nearest": 31,
+      "3rd-furthest": 31,
+      "2nd-furthest": 31,
+      "furthest": 16
+    },
+    "answer_letters": {
+      "A": 39,
+      "B": 39,
+      "C": 39,
+      "D": 39
+    },
+    "hard_margin": 104,
+    "stale_flips": 0,
+    "mover_targets": 0,
+    "turn_imbalance_mean": 0.0427,
+    "surface_policy_pct": {
+      "most_forward": 25.3,
+      "fewest_turns": 25.3,
+      "most_turns": 26.6,
+      "most_waits": 25.0,
+      "max_net_rotation": 27.2,
+      "net_is_0": 28.7,
+      "net_is_90": 26.5,
+      "net_is_180": 25.0,
+      "net_is_270": 28.8,
+      "first_is_move_forward": 24.4,
+      "first_is_turn_left": 24.8,
+      "first_is_turn_right": 25.9,
+      "first_is_noop": 25.0
+    }
+  },
+  "generator_stats": {
+    "attempts": 1656,
+    "rejected": {
+      "world": 0,
+      "walk": 0,
+      "coverage": 0,
+      "wall_clearance": 0,
+      "no_feasible_cell": 671,
+      "cell_full": 791,
+      "ratio_budget": 12,
+      "pixel_coverage": 25
+    },
+    "walk_candidates_mean": 2.3,
+    "wall_clearance_rejects_mean": 0.21
+  },
+  "audits": {
+    "n_items": 156,
+    "plan_len": {
+      "4": 52,
+      "5": 52,
+      "6": 52
+    },
+    "sector": {
+      "behind": 39,
+      "front": 39,
+      "left": 39,
+      "right": 39
+    },
+    "rank": {
+      "2nd-furthest": 31,
+      "2nd-nearest": 31,
+      "3rd-furthest": 31,
+      "3rd-nearest": 31,
+      "furthest": 16,
+      "nearest": 16
+    },
+    "context": {
+      "5": 34,
+      "6": 50,
+      "7": 40,
+      "8": 32
+    },
+    "context_vs_l1": {
+      "l2": {
+        "5": 0.218,
+        "6": 0.321,
+        "7": 0.256,
+        "8": 0.205
+      },
+      "l1_set": "easy_eval_v1_full",
+      "l1": {
+        "5": 0.25,
+        "6": 0.25,
+        "7": 0.25,
+        "8": 0.25
+      },
+      "max_abs_share_gap": 0.071,
+      "caveat": "Context is soft-balanced, and W1's all-boxes-twice sighting rule bends it toward longer walks (W0, needing one sighting per box, stays near uniform). Recorded rather than corrected, on two grounds: context was measured on the v1 chain to be a redundancy axis with no monotone effect on accuracy, and the bend runs CONSERVATIVE for the W0-vs-W1 contrast — W1 carries the longer observations, which should help it, so any W1 deficit measured against W0 is if anything understated. Do not slice W1 by context: the shortest level is nearly empty."
+    },
+    "margin_bin": {
+      "near": 104,
+      "wide": 52
+    },
+    "query_target": {
+      "static": 156
+    },
+    "answer_letter": {
+      "A": 39,
+      "B": 39,
+      "C": 39,
+      "D": 39
+    },
+    "distinct_endpoint_positions": {
+      "2": 1,
+      "3": 50,
+      "4": 105
+    },
+    "distinct_target_ids": {
+      "1": 77,
+      "2": 72,
+      "3": 7
+    },
+    "last_frame_visible_share_mean": 0.227,
+    "answer_changed": {
+      "null": 156
+    },
+    "roles_per_item": {
+      "correct": 156,
+      "free": 135,
+      "rank_near_miss": 165,
+      "sector_only": 168
+    },
+    "reference_lines": {
+      "stale_world_equals_correct": 156,
+      "alt_reading_equals_correct": 94
+    },
+    "surface_correct": {
+      "n_turns_mean": 2.378,
+      "n_forward_mean": 2.622,
+      "n_waits_mean": 0.0,
+      "first_action": {
+        "move_forward": 63,
+        "turn_left": 43,
+        "turn_right": 50
+      },
+      "net_rotation": {
+        "0": 32,
+        "90": 37,
+        "180": 47,
+        "270": 40
+      }
+    },
+    "surface_distractors": {
+      "n_turns_mean": 2.335,
+      "n_forward_mean": 2.665,
+      "n_waits_mean": 0.0,
+      "first_action": {
+        "move_forward": 198,
+        "turn_left": 129,
+        "turn_right": 141
+      },
+      "net_rotation": {
+        "0": 81,
+        "90": 119,
+        "180": 158,
+        "270": 110
+      }
+    },
+    "surface_shortcut_audit": {
+      "n_items": 156,
+      "chance_pct": 25.0,
+      "accuracy_pct": {
+        "most_forward": 25.3,
+        "fewest_turns": 25.3,
+        "most_turns": 26.6,
+        "most_waits": 25.0,
+        "max_net_rotation": 27.2,
+        "net_is_0": 28.7,
+        "net_is_90": 26.5,
+        "net_is_180": 25.0,
+        "net_is_270": 28.8,
+        "first_is_move_forward": 24.4,
+        "first_is_turn_left": 24.8,
+        "first_is_turn_right": 25.9,
+        "first_is_noop": 25.0
+      },
+      "z": {
+        "most_forward": 0.09,
+        "fewest_turns": 0.09,
+        "most_turns": 0.46,
+        "most_waits": 0.0,
+        "max_net_rotation": 0.65,
+        "net_is_0": 1.06,
+        "net_is_90": 0.43,
+        "net_is_180": 0.0,
+        "net_is_270": 1.09,
+        "first_is_move_forward": -0.17,
+        "first_is_turn_left": -0.06,
+        "first_is_turn_right": 0.25,
+        "first_is_noop": 0.0
+      },
+      "worst_policy": "net_is_270",
+      "worst_z": 1.09,
+      "null_max_z_p95": 2.07,
+      "family_p_value": 0.6973,
+      "permutations": 3000,
+      "verdict": "clean"
+    },
+    "turn_imbalance_mean": 0.0427,
+    "plan_measured_mismatch_total": 22
+  }
+}