Datasets:
File size: 12,275 Bytes
48521cd | 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 91 92 93 94 95 96 97 98 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237 238 239 240 241 242 243 244 245 246 247 248 249 250 251 252 253 254 255 256 257 258 259 260 261 262 263 264 265 266 267 268 269 270 271 272 273 274 275 276 277 278 279 280 281 282 283 284 285 286 287 288 289 290 291 292 293 294 295 296 297 298 299 300 301 302 303 304 305 306 307 308 309 310 311 312 313 314 315 316 317 318 319 320 321 322 323 324 325 326 327 328 329 330 331 332 333 334 335 336 337 338 339 340 341 342 343 344 345 346 347 348 349 350 351 352 353 354 355 356 357 358 359 360 361 362 363 364 365 366 367 368 369 370 371 372 373 374 375 376 377 378 379 380 381 382 383 384 385 386 387 388 | {
"version": "bekko.search_score_policy.v2.20260928",
"seed": 20260928,
"grid_step": 0.05,
"minimum_gap": 0.05,
"assignment": "Uniform over the 20 policies, fixed per group/query and shared across its documents, independent of target and split; family assignment is uniform over policies at 7 broad, 7 standard, and 6 strict.",
"target_semantics": "Policy-conditioned utility, not a probability. Original teacher evidence grades are retained with family-specific source-positive floor handling and full provenance. The numeric utility mapping is subjective and synthetic, not an individually measured human scale.",
"mapping_note": "Utility anchors have fixed endpoints 0 and 1; intermediate anchors use the 0.05 grid with ±1-grid-step jitter subject to monotonicity and minimum gap. Anchor values express a subjective policy preference, not a claim about an exact largest step or measured rating interval.",
"families": {
"broad": {
"positive_floor": 3,
"description": "Broad discovery keeps more potentially useful leads and exact-subject context. Before mapping, source-positive grades are floored at 3; grade 3 means useful retrieval material and may include incomplete support, while grade 4 is a sufficient answer.",
"criteria_levels": [
"Unrelated to the search intent, or about a different referent with no supported connection to the intended subject.",
"Loose topical or lexical overlap, without a useful lead toward the search intent.",
"Potentially useful context about the intended subject, with only an indirect connection to the precise search need.",
"A strong, query-related lead worth retaining for further investigation, even when the requested answer details are absent.",
"A direct and sufficient answer for the intended subject."
]
},
"standard": {
"positive_floor": 3,
"description": "Standard retrieval relevance: source positives have a minimum grade of 3; grade 3 means a relevant document matching the search intent, without requiring explicit answer facts.",
"criteria_levels": [
"Unrelated to the search intent, or about a different referent with no supported connection to the intended subject.",
"Only topical or lexical overlap, without a clear match to the search intent.",
"Useful background about the intended subject, but not clearly addressing the specific search intent.",
"A relevant document matching the search intent and intended subject; direct answer facts are not required.",
"A direct and sufficient answer for the intended subject."
]
},
"strict": {
"positive_floor": null,
"description": "Strict selection discounts context and requires clear support for answer utility. Source grades are used without a positive floor; grade 3 requires explicit partial-answer evidence or a necessary, supported linking fact, and grade 4 is a sufficient answer.",
"criteria_levels": [
"No useful evidence, or evidence about a different referent.",
"Topic overlap without a usable fact for answering the request.",
"Exact-subject context with limited support, but no clear answer-bearing or necessary linking fact.",
"Explicit partial answer evidence or a necessary, well-supported identification or relationship; a complete answer is not required.",
"A direct and sufficient answer for the intended subject."
]
}
},
"policies": [
{
"id": "broad_recall_anchor",
"family": "broad",
"instruction": "For an early search pass, keep material that could help a researcher follow a lead. Give exact-target facts more utility than loose context, and give a complete answer the most.",
"anchors": [
0.0,
0.25,
0.55,
0.8,
1.0
],
"jitter_steps": [
-1,
0,
1
]
},
{
"id": "broad_topic_credit",
"family": "broad",
"instruction": "When exploring a topic, preserve concrete adjacent material that may point toward an answer. Reward exact-target evidence more, and reserve the top score for a sufficient answer.",
"anchors": [
0.0,
0.3,
0.55,
0.8,
1.0
],
"jitter_steps": [
-1,
0,
1
]
},
{
"id": "broad_context_credit",
"family": "broad",
"instruction": "For background gathering, retain useful context about the named subject even when it does not state the requested detail. Directly useful facts should score higher.",
"anchors": [
0.0,
0.25,
0.5,
0.8,
1.0
],
"jitter_steps": [
-1,
0,
1
]
},
{
"id": "broad_partial_evidence",
"family": "broad",
"instruction": "Keep leads that could contribute to a later synthesis, including concrete facts that address only part of the request. A complete answer remains the strongest result.",
"anchors": [
0.0,
0.3,
0.55,
0.8,
1.0
],
"jitter_steps": [
-1,
0,
1
]
},
{
"id": "broad_low_grade_value",
"family": "broad",
"instruction": "In a first-pass screen, assign some value to plausible contextual material because a reviewer may filter it later. Do not treat that value as proof that the document answers the query.",
"anchors": [
0.0,
0.35,
0.6,
0.8,
1.0
],
"jitter_steps": [
-1,
0,
1
]
},
{
"id": "recall_balanced_steps",
"family": "broad",
"instruction": "Use a gradual discovery scale: retain relevant context, favor facts about the requested subject, and favor partial answers over context alone.",
"anchors": [
0.0,
0.25,
0.5,
0.75,
1.0
],
"jitter_steps": [
-1,
0,
1
]
},
{
"id": "context_first",
"family": "broad",
"instruction": "For research collection, useful background about the exact subject can be worth keeping before an answer is available. Give answer-bearing evidence stronger utility.",
"anchors": [
0.0,
0.2,
0.55,
0.8,
1.0
],
"jitter_steps": [
-1,
0,
1
]
},
{
"id": "answer_evidence_first",
"family": "standard",
"instruction": "Prefer documents with facts that answer the request. Exact-topic context can help, but it should receive clearly less utility than answer evidence.",
"anchors": [
0.0,
0.2,
0.5,
0.75,
1.0
],
"jitter_steps": [
-1,
0,
1
]
},
{
"id": "directness_first",
"family": "standard",
"instruction": "For a results page, favor evidence that directly addresses the request. Useful partial evidence matters, while a sufficient answer should rank highest.",
"anchors": [
0.0,
0.2,
0.45,
0.75,
1.0
],
"jitter_steps": [
-1,
0,
1
]
},
{
"id": "context_vs_directness",
"family": "standard",
"instruction": "Balance explanatory context with answer sufficiency: useful context has real value, specific partial evidence has more, and a complete answer has the most.",
"anchors": [
0.0,
0.25,
0.5,
0.75,
1.0
],
"jitter_steps": [
-1,
0,
1
]
},
{
"id": "topic_vs_context",
"family": "standard",
"instruction": "Distinguish exact-subject background from mere topic similarity. Give concrete requested facts more utility than background, with sufficient answers at the top.",
"anchors": [
0.0,
0.2,
0.55,
0.75,
1.0
],
"jitter_steps": [
-1,
0,
1
]
},
{
"id": "partial_answer_priority",
"family": "standard",
"instruction": "For a request that may need synthesis, value a concrete fact answering an important part of it. Such evidence is useful, though a sufficient answer remains stronger.",
"anchors": [
0.0,
0.2,
0.5,
0.8,
1.0
],
"jitter_steps": [
-1,
0,
1
]
},
{
"id": "answer_threshold_emphasis",
"family": "standard",
"instruction": "Use answer-bearing facts as the main reason to select a result. Relevant background may remain useful at a lower level, but should not be mistaken for an answer.",
"anchors": [
0.0,
0.2,
0.45,
0.75,
1.0
],
"jitter_steps": [
-1,
0,
1
]
},
{
"id": "near_linear_strict",
"family": "standard",
"instruction": "Increase utility steadily as evidence becomes more specific to the requested information, from exact-subject context through partial facts to a sufficient answer.",
"anchors": [
0.0,
0.25,
0.5,
0.75,
1.0
],
"jitter_steps": [
-1,
0,
1
]
},
{
"id": "high_grade_concentration",
"family": "strict",
"instruction": "Concentrate selection on evidence that answers the request. Keep general context low, while giving a concrete partial answer meaningful utility.",
"anchors": [
0.0,
0.05,
0.2,
0.7,
1.0
],
"jitter_steps": [
-1,
0,
1
]
},
{
"id": "strict_precision_anchor",
"family": "strict",
"instruction": "Use a precision-focused shortlist: give little credit to topic overlap or background, but retain clearly supported facts that answer part of the request.",
"anchors": [
0.0,
0.1,
0.2,
0.75,
1.0
],
"jitter_steps": [
-1,
0,
1
]
},
{
"id": "strict_context_discount",
"family": "strict",
"instruction": "Apply a strong discount to context that cannot help establish an answer. Explicit partial answers and necessary supported links still deserve substantial utility.",
"anchors": [
0.0,
0.05,
0.25,
0.7,
1.0
],
"jitter_steps": [
-1,
0,
1
]
},
{
"id": "strict_direct_answer",
"family": "strict",
"instruction": "For final answer selection, reserve the highest utility for a direct sufficient answer. A specific partial fact is still useful evidence and should not be treated as irrelevant.",
"anchors": [
0.0,
0.05,
0.15,
0.8,
1.0
],
"jitter_steps": [
-1,
0,
1
]
},
{
"id": "strict_high_grade_only",
"family": "strict",
"instruction": "Keep weak context near the bottom and focus selection on explicit requested facts. A well-supported intermediate fact can help answer the query even without completing it.",
"anchors": [
0.0,
0.05,
0.2,
0.75,
1.0
],
"jitter_steps": [
-1,
0,
1
]
},
{
"id": "precision_heavy",
"family": "strict",
"instruction": "Prefer a concise set of documents with concrete answer evidence over loosely connected material. Preserve meaningful credit for explicit partial answers.",
"anchors": [
0.0,
0.1,
0.25,
0.8,
1.0
],
"jitter_steps": [
-1,
0,
1
]
}
],
"rationale": "Family-specific rubrics distinguish broad discovery, conventional retrieval relevance, and strict answer evidence. Targets are derived from existing grades and source relevance labels, not independently relabeled under each rubric. Source-label noise and FAQ/claim-specific limitations remain; existing quarantine decisions are preserved."
}
|