Research-Stack/shared-data/pist_route_repair_v13b_benchmark.json

816 lines
No EOL
19 KiB
JSON

{
"n": 28,
"recovered": 15,
"results": [
{
"name": "rw_missing_dir_1",
"obstruction": "missing_rewrite_direction",
"initial_status": "failed",
"recovered": false,
"attempts": [
{
"attempt": 1,
"tag": "rw_reverse_simp",
"patch": "rw [\u2190 {'name': 'h', 'type': 'a = b', 'is_prop': False, 'is_n",
"ok": false
},
{
"attempt": 2,
"tag": "rw_forward_simp",
"patch": "rw [{'name': 'h', 'type': 'a = b', 'is_prop': False, 'is_nat",
"ok": false
},
{
"attempt": 3,
"tag": "rw_reverse_rfl",
"patch": "rw [\u2190 {'name': 'h', 'type': 'a = b', 'is_prop': False, 'is_n",
"ok": false
},
{
"attempt": 4,
"tag": "rw_forward_rfl",
"patch": "rw [{'name': 'h', 'type': 'a = b', 'is_prop': False, 'is_nat",
"ok": false
}
],
"best_attempt": {
"attempt": 1,
"tag": "rw_reverse_simp",
"patch": "rw [\u2190 {'name': 'h', 'type': 'a = b', 'is_prop': False, 'is_n",
"ok": false
},
"n_candidates": 4
},
{
"name": "rw_missing_dir_2",
"obstruction": "arithmetic_gap",
"initial_status": "failed",
"recovered": true,
"attempts": [
{
"attempt": 1,
"tag": "omega",
"patch": "omega",
"ok": true
}
],
"best_attempt": {
"attempt": 1,
"tag": "omega",
"patch": "omega",
"ok": true
},
"n_candidates": 5
},
{
"name": "rw_missing_dir_3",
"obstruction": "arithmetic_gap",
"initial_status": "failed",
"recovered": true,
"attempts": [
{
"attempt": 1,
"tag": "omega",
"patch": "omega",
"ok": true
}
],
"best_attempt": {
"attempt": 1,
"tag": "omega",
"patch": "omega",
"ok": true
},
"n_candidates": 5
},
{
"name": "rw_missing_dir_4",
"obstruction": "missing_rewrite_direction",
"initial_status": "failed",
"recovered": false,
"attempts": [
{
"attempt": 1,
"tag": "rw_reverse_simp",
"patch": "rw [\u2190 {'name': 'h', 'type': 'a = b', 'is_prop': False, 'is_n",
"ok": false
},
{
"attempt": 2,
"tag": "rw_forward_simp",
"patch": "rw [{'name': 'h', 'type': 'a = b', 'is_prop': False, 'is_nat",
"ok": false
},
{
"attempt": 3,
"tag": "rw_reverse_rfl",
"patch": "rw [\u2190 {'name': 'h', 'type': 'a = b', 'is_prop': False, 'is_n",
"ok": false
},
{
"attempt": 4,
"tag": "rw_forward_rfl",
"patch": "rw [{'name': 'h', 'type': 'a = b', 'is_prop': False, 'is_nat",
"ok": false
}
],
"best_attempt": {
"attempt": 1,
"tag": "rw_reverse_simp",
"patch": "rw [\u2190 {'name': 'h', 'type': 'a = b', 'is_prop': False, 'is_n",
"ok": false
},
"n_candidates": 4
},
{
"name": "rw_missing_dir_5",
"obstruction": "arithmetic_gap",
"initial_status": "failed",
"recovered": true,
"attempts": [
{
"attempt": 1,
"tag": "omega",
"patch": "omega",
"ok": true
}
],
"best_attempt": {
"attempt": 1,
"tag": "omega",
"patch": "omega",
"ok": true
},
"n_candidates": 5
},
{
"name": "rw_missing_dir_6",
"obstruction": "missing_rewrite_direction",
"initial_status": "failed",
"recovered": false,
"attempts": [
{
"attempt": 1,
"tag": "rw_reverse_simp",
"patch": "rw [\u2190 {'name': 'h', 'type': 'a = b', 'is_prop': False, 'is_n",
"ok": false
},
{
"attempt": 2,
"tag": "rw_forward_simp",
"patch": "rw [{'name': 'h', 'type': 'a = b', 'is_prop': False, 'is_nat",
"ok": false
},
{
"attempt": 3,
"tag": "rw_reverse_rfl",
"patch": "rw [\u2190 {'name': 'h', 'type': 'a = b', 'is_prop': False, 'is_n",
"ok": false
},
{
"attempt": 4,
"tag": "rw_forward_rfl",
"patch": "rw [{'name': 'h', 'type': 'a = b', 'is_prop': False, 'is_nat",
"ok": false
}
],
"best_attempt": {
"attempt": 1,
"tag": "rw_reverse_simp",
"patch": "rw [\u2190 {'name': 'h', 'type': 'a = b', 'is_prop': False, 'is_n",
"ok": false
},
"n_candidates": 4
},
{
"name": "rw_missing_dir_7",
"obstruction": "arithmetic_gap",
"initial_status": "failed",
"recovered": true,
"attempts": [
{
"attempt": 1,
"tag": "omega",
"patch": "omega",
"ok": true
}
],
"best_attempt": {
"attempt": 1,
"tag": "omega",
"patch": "omega",
"ok": true
},
"n_candidates": 5
},
{
"name": "missing_assume_1",
"obstruction": "missing_assumption_bridge",
"initial_status": "failed",
"recovered": true,
"attempts": [
{
"attempt": 1,
"tag": "apply_exact",
"patch": "apply hAB\n exact hA",
"ok": false
},
{
"attempt": 2,
"tag": "exact_apply",
"patch": "exact hAB hA",
"ok": true
}
],
"best_attempt": {
"attempt": 2,
"tag": "exact_apply",
"patch": "exact hAB hA",
"ok": true
},
"n_candidates": 5
},
{
"name": "missing_assume_2",
"obstruction": "missing_assumption_bridge",
"initial_status": "failed",
"recovered": false,
"attempts": [
{
"attempt": 1,
"tag": "apply_exact",
"patch": "apply hAB\n exact hA",
"ok": false
},
{
"attempt": 2,
"tag": "exact_apply",
"patch": "exact hAB hA",
"ok": false
},
{
"attempt": 3,
"tag": "apply_exact_gen",
"patch": "apply hAB\n exact A",
"ok": false
},
{
"attempt": 4,
"tag": "apply_exact_gen",
"patch": "apply hAB\n exact B",
"ok": false
},
{
"attempt": 5,
"tag": "apply_exact_gen",
"patch": "apply hAB\n exact C",
"ok": false
}
],
"best_attempt": {
"attempt": 1,
"tag": "apply_exact",
"patch": "apply hAB\n exact hA",
"ok": false
},
"n_candidates": 5
},
{
"name": "missing_assume_3",
"obstruction": "missing_destructuring",
"initial_status": "failed",
"recovered": true,
"attempts": [
{
"attempt": 1,
"tag": "dot_left",
"patch": "exact h.left",
"ok": true
}
],
"best_attempt": {
"attempt": 1,
"tag": "dot_left",
"patch": "exact h.left",
"ok": true
},
"n_candidates": 4
},
{
"name": "missing_assume_4",
"obstruction": "case_split_missing",
"initial_status": "failed",
"recovered": true,
"attempts": [
{
"attempt": 1,
"tag": "case_same",
"patch": "cases h with\n | inl h => left; exact h\n | inr h => right; ",
"ok": true
}
],
"best_attempt": {
"attempt": 1,
"tag": "case_same",
"patch": "cases h with\n | inl h => left; exact h\n | inr h => right; ",
"ok": true
},
"n_candidates": 2
},
{
"name": "missing_assume_5",
"obstruction": "missing_assumption_bridge",
"initial_status": "failed",
"recovered": true,
"attempts": [
{
"attempt": 1,
"tag": "apply_exact",
"patch": "apply h\n exact hA",
"ok": false
},
{
"attempt": 2,
"tag": "exact_apply",
"patch": "exact h hA",
"ok": true
}
],
"best_attempt": {
"attempt": 2,
"tag": "exact_apply",
"patch": "exact h hA",
"ok": true
},
"n_candidates": 5
},
{
"name": "missing_assume_6",
"obstruction": "intro_chain_missing",
"initial_status": "failed",
"recovered": false,
"attempts": [
{
"attempt": 1,
"tag": "intro_first",
"patch": "intro h0\n intro h1\n exact h0",
"ok": false
},
{
"attempt": 2,
"tag": "intro_last",
"patch": "intro h0\n intro h1\n exact h1",
"ok": false
},
{
"attempt": 3,
"tag": "intro_prop",
"patch": "intro h0\n intro h1\n exact A",
"ok": false
},
{
"attempt": 4,
"tag": "intro_single",
"patch": "intro h\n exact h",
"ok": false
}
],
"best_attempt": {
"attempt": 1,
"tag": "intro_first",
"patch": "intro h0\n intro h1\n exact h0",
"ok": false
},
"n_candidates": 4
},
{
"name": "missing_assume_7",
"obstruction": "intro_chain_missing",
"initial_status": "failed",
"recovered": false,
"attempts": [
{
"attempt": 1,
"tag": "intro_first",
"patch": "intro h0\n exact h0",
"ok": false
},
{
"attempt": 2,
"tag": "intro_prop",
"patch": "intro h0\n exact P",
"ok": false
},
{
"attempt": 3,
"tag": "intro_single",
"patch": "intro h\n exact h",
"ok": false
}
],
"best_attempt": {
"attempt": 1,
"tag": "intro_first",
"patch": "intro h0\n exact h0",
"ok": false
},
"n_candidates": 3
},
{
"name": "arith_gap_1",
"obstruction": "arithmetic_gap",
"initial_status": "failed",
"recovered": true,
"attempts": [
{
"attempt": 1,
"tag": "omega",
"patch": "omega",
"ok": true
}
],
"best_attempt": {
"attempt": 1,
"tag": "omega",
"patch": "omega",
"ok": true
},
"n_candidates": 4
},
{
"name": "arith_gap_2",
"obstruction": "arithmetic_gap",
"initial_status": "failed",
"recovered": true,
"attempts": [
{
"attempt": 1,
"tag": "omega",
"patch": "omega",
"ok": true
}
],
"best_attempt": {
"attempt": 1,
"tag": "omega",
"patch": "omega",
"ok": true
},
"n_candidates": 4
},
{
"name": "arith_gap_3",
"obstruction": "arithmetic_gap",
"initial_status": "failed",
"recovered": false,
"attempts": [
{
"attempt": 1,
"tag": "omega",
"patch": "omega",
"ok": false
},
{
"attempt": 2,
"tag": "norm_num",
"patch": "norm_num",
"ok": false
},
{
"attempt": 3,
"tag": "rw_omega",
"patch": "rw [{'name': 'h', 'type': 'a + b = b + a', 'is_prop': False,",
"ok": false
},
{
"attempt": 4,
"tag": "simp_omega",
"patch": "simp\n omega",
"ok": false
},
{
"attempt": 5,
"tag": "simp_norm_num",
"patch": "simp\n norm_num",
"ok": false
}
],
"best_attempt": {
"attempt": 1,
"tag": "omega",
"patch": "omega",
"ok": false
},
"n_candidates": 5
},
{
"name": "arith_gap_4",
"obstruction": "arithmetic_gap",
"initial_status": "failed",
"recovered": true,
"attempts": [
{
"attempt": 1,
"tag": "omega",
"patch": "omega",
"ok": true
}
],
"best_attempt": {
"attempt": 1,
"tag": "omega",
"patch": "omega",
"ok": true
},
"n_candidates": 4
},
{
"name": "arith_gap_5",
"obstruction": "arithmetic_gap",
"initial_status": "failed",
"recovered": false,
"attempts": [
{
"attempt": 1,
"tag": "omega",
"patch": "omega",
"ok": false
},
{
"attempt": 2,
"tag": "norm_num",
"patch": "norm_num",
"ok": false
},
{
"attempt": 3,
"tag": "simp_omega",
"patch": "simp\n omega",
"ok": false
},
{
"attempt": 4,
"tag": "simp_norm_num",
"patch": "simp\n norm_num",
"ok": false
}
],
"best_attempt": {
"attempt": 1,
"tag": "omega",
"patch": "omega",
"ok": false
},
"n_candidates": 4
},
{
"name": "arith_gap_6",
"obstruction": "arithmetic_gap",
"initial_status": "failed",
"recovered": true,
"attempts": [
{
"attempt": 1,
"tag": "omega",
"patch": "omega",
"ok": true
}
],
"best_attempt": {
"attempt": 1,
"tag": "omega",
"patch": "omega",
"ok": true
},
"n_candidates": 4
},
{
"name": "arith_gap_8",
"obstruction": "arithmetic_gap",
"initial_status": "failed",
"recovered": false,
"attempts": [
{
"attempt": 1,
"tag": "omega",
"patch": "omega",
"ok": false
},
{
"attempt": 2,
"tag": "norm_num",
"patch": "norm_num",
"ok": false
},
{
"attempt": 3,
"tag": "simp_omega",
"patch": "simp\n omega",
"ok": false
},
{
"attempt": 4,
"tag": "simp_norm_num",
"patch": "simp\n norm_num",
"ok": false
}
],
"best_attempt": {
"attempt": 1,
"tag": "omega",
"patch": "omega",
"ok": false
},
"n_candidates": 4
},
{
"name": "arith_gap_9",
"obstruction": "arithmetic_gap",
"initial_status": "failed",
"recovered": true,
"attempts": [
{
"attempt": 1,
"tag": "omega",
"patch": "omega",
"ok": true
}
],
"best_attempt": {
"attempt": 1,
"tag": "omega",
"patch": "omega",
"ok": true
},
"n_candidates": 4
},
{
"name": "case_split_1",
"obstruction": "case_split_missing",
"initial_status": "failed",
"recovered": true,
"attempts": [
{
"attempt": 1,
"tag": "case_swap",
"patch": "cases h with\n | inl h => right; exact h\n | inr h => left; ",
"ok": true
}
],
"best_attempt": {
"attempt": 1,
"tag": "case_swap",
"patch": "cases h with\n | inl h => right; exact h\n | inr h => left; ",
"ok": true
},
"n_candidates": 2
},
{
"name": "case_split_2",
"obstruction": "missing_destructuring",
"initial_status": "failed",
"recovered": false,
"attempts": [
{
"attempt": 1,
"tag": "dot_left",
"patch": "exact h.left",
"ok": false
},
{
"attempt": 2,
"tag": "dot_right",
"patch": "exact h.right",
"ok": false
},
{
"attempt": 3,
"tag": "cases_and_elim",
"patch": "cases h with\n | intro h1 h2 => exact h1",
"ok": false
},
{
"attempt": 4,
"tag": "rcases_elim",
"patch": "rcases h with \u27e8h1, h2\u27e9\n exact h1",
"ok": false
}
],
"best_attempt": {
"attempt": 1,
"tag": "dot_left",
"patch": "exact h.left",
"ok": false
},
"n_candidates": 4
},
{
"name": "case_split_3",
"obstruction": "constructor_missing",
"initial_status": "failed",
"recovered": false,
"attempts": [
{
"attempt": 1,
"tag": "constructor_exact",
"patch": "constructor\n \u00b7 exact hA",
"ok": false
},
{
"attempt": 2,
"tag": "constructor_exact",
"patch": "constructor\n \u00b7 exact hB",
"ok": false
},
{
"attempt": 3,
"tag": "constructor_assumption",
"patch": "constructor\n \u00b7 assumption\n \u00b7 assumption",
"ok": false
},
{
"attempt": 4,
"tag": "constructor_props",
"patch": "constructor\n \u00b7 exact A\n \u00b7 exact B",
"ok": false
}
],
"best_attempt": {
"attempt": 1,
"tag": "constructor_exact",
"patch": "constructor\n \u00b7 exact hA",
"ok": false
},
"n_candidates": 4
},
{
"name": "case_split_4",
"obstruction": "case_split_missing",
"initial_status": "failed",
"recovered": true,
"attempts": [
{
"attempt": 1,
"tag": "case_same",
"patch": "cases h with\n | inl h => left; exact h\n | inr h => right; ",
"ok": true
}
],
"best_attempt": {
"attempt": 1,
"tag": "case_same",
"patch": "cases h with\n | inl h => left; exact h\n | inr h => right; ",
"ok": true
},
"n_candidates": 2
},
{
"name": "case_split_5",
"obstruction": "case_split_missing",
"initial_status": "failed",
"recovered": false,
"attempts": [],
"best_attempt": null,
"n_candidates": 0
},
{
"name": "case_split_6",
"obstruction": "constructor_missing",
"initial_status": "failed",
"recovered": false,
"attempts": [
{
"attempt": 1,
"tag": "constructor_exact",
"patch": "constructor\n \u00b7 exact hA",
"ok": false
},
{
"attempt": 2,
"tag": "constructor_exact",
"patch": "constructor\n \u00b7 exact hB",
"ok": false
},
{
"attempt": 3,
"tag": "constructor_assumption",
"patch": "constructor\n \u00b7 assumption\n \u00b7 assumption",
"ok": false
},
{
"attempt": 4,
"tag": "constructor_props",
"patch": "constructor\n \u00b7 exact A\n \u00b7 exact B",
"ok": false
}
],
"best_attempt": {
"attempt": 1,
"tag": "constructor_exact",
"patch": "constructor\n \u00b7 exact hA",
"ok": false
},
"n_candidates": 4
}
]
}