Spaces:
Sleeping
Sleeping
File size: 12,097 Bytes
feb1b1c dd3dc8e fc87e83 dd3dc8e fc87e83 dd3dc8e fc87e83 | 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 | from __future__ import annotations
import math
import re
from collections import Counter
from typing import Final, final
from pydantic import BaseModel, ConfigDict
from redstack.domain.enums import ReasoningPolarity, Severity, ValidationCode
from redstack.domain.ranking import RankedCandidate, Ranking
from redstack.domain.validation import ValidationFinding, ValidationReport
_ID_PATTERN: Final[re.Pattern[str]] = re.compile(r"^CAND_[0-9]{7}$")
# Normalization for the templated-reasoning skeleton: collapse ids and numbers.
_NUM_TOKEN: Final[re.Pattern[str]] = re.compile(r"\d+(?:\.\d+)?")
_ID_TOKEN: Final[re.Pattern[str]] = re.compile(r"CAND_[0-9]{7}")
_TEMPLATED_FRACTION: Final[float] = 0.70
_ALL_CODES: Final[frozenset[ValidationCode]] = frozenset(ValidationCode)
@final
class ValidationEngine(BaseModel):
"""Stateless, pure validator over a (reasoned) ``Ranking``."""
model_config = ConfigDict(
frozen=True, extra="forbid", arbitrary_types_allowed=False
)
expected_size: int = 100
# ------------------------------------------------------------------ public
def validate_ranking(self, ranking: Ranking) -> ValidationReport:
"""Run all structural + Stage-4 checks; assemble the ``ValidationReport``.
Named ``validate_ranking`` rather than ``validate`` so this doesn't
shadow ``pydantic.BaseModel.validate`` (a deprecated classmethod with
an unrelated signature).
"""
findings: list[ValidationFinding] = []
ordered = ranking.ordered
findings.extend(self._row_count(ordered))
findings.extend(self._ranks(ordered))
findings.extend(self._ids(ordered))
findings.extend(self._scores(ordered))
findings.extend(self._reasoning(ordered))
is_valid = not any(f.severity is Severity.HARD for f in findings)
return ValidationReport(
findings=tuple(findings),
is_valid=is_valid,
checks_run=_ALL_CODES,
)
# ----------------------------------------------------------- structural
def _row_count(
self, ordered: tuple[RankedCandidate, ...]
) -> tuple[ValidationFinding, ...]:
if len(ordered) == self.expected_size:
return ()
return (
ValidationFinding(
code=ValidationCode.WRONG_ROW_COUNT,
severity=Severity.HARD,
message=f"expected {self.expected_size} rows, got {len(ordered)}",
location=None,
),
)
def _ranks(
self, ordered: tuple[RankedCandidate, ...]
) -> tuple[ValidationFinding, ...]:
out: list[ValidationFinding] = []
ranks = [c.rank for c in ordered]
for candidate in ordered:
if candidate.rank < 1 or candidate.rank > self.expected_size:
out.append(
ValidationFinding(
code=ValidationCode.RANK_OUT_OF_RANGE,
severity=Severity.HARD,
message=f"rank {candidate.rank} out of [1, {self.expected_size}]",
location=candidate.candidate_id,
)
)
counts = Counter(ranks)
for rank, freq in sorted(counts.items()):
if freq > 1:
out.append(
ValidationFinding(
code=ValidationCode.DUPLICATE_RANK,
severity=Severity.HARD,
message=f"rank {rank} appears {freq} times",
location=None,
)
)
expected = set(range(1, self.expected_size + 1))
missing = expected - set(ranks)
if missing:
out.append(
ValidationFinding(
code=ValidationCode.MISSING_RANK,
severity=Severity.HARD,
message=f"missing ranks: {sorted(missing)[:10]}",
location=None,
)
)
return tuple(out)
def _ids(
self, ordered: tuple[RankedCandidate, ...]
) -> tuple[ValidationFinding, ...]:
out: list[ValidationFinding] = []
seen: Counter[str] = Counter(c.candidate_id for c in ordered)
for candidate in ordered:
if _ID_PATTERN.match(candidate.candidate_id) is None:
out.append(
ValidationFinding(
code=ValidationCode.BAD_ID_FORMAT,
severity=Severity.HARD,
message=f"id {candidate.candidate_id!r} fails ^CAND_[0-9]{{7}}$",
location=candidate.candidate_id,
)
)
for cid, freq in sorted(seen.items()):
if freq > 1:
out.append(
ValidationFinding(
code=ValidationCode.DUPLICATE_ID,
severity=Severity.HARD,
message=f"id {cid} appears {freq} times",
location=cid,
)
)
return tuple(out)
def _scores(
self, ordered: tuple[RankedCandidate, ...]
) -> tuple[ValidationFinding, ...]:
out: list[ValidationFinding] = []
for candidate in ordered:
value = candidate.score
if not isinstance(value, float) or not math.isfinite(value):
out.append(
ValidationFinding(
code=ValidationCode.SCORE_NOT_FLOAT,
severity=Severity.HARD,
message=f"score {value!r} is not a finite float",
location=candidate.candidate_id,
)
)
for earlier, later in zip(ordered, ordered[1:], strict=False):
if later.score > earlier.score:
out.append(
ValidationFinding(
code=ValidationCode.SCORE_INCREASING,
severity=Severity.HARD,
message=(
f"score rises at rank {later.rank} "
f"({earlier.score} -> {later.score})"
),
location=later.candidate_id,
)
)
elif (
later.score == earlier.score
and earlier.candidate_id >= later.candidate_id
):
out.append(
ValidationFinding(
code=ValidationCode.TIEBREAK_VIOLATION,
severity=Severity.HARD,
message=(
f"tie at rank {later.rank}: "
f"{earlier.candidate_id} !< {later.candidate_id}"
),
location=later.candidate_id,
)
)
return tuple(out)
# ------------------------------------------------------------- Stage-4
def _reasoning(
self, ordered: tuple[RankedCandidate, ...]
) -> tuple[ValidationFinding, ...]:
out: list[ValidationFinding] = []
rendered: list[str] = []
skeletons: list[str] = []
for candidate in ordered:
reasoning = candidate.reasoning
if reasoning is None or not reasoning.rendered.strip():
out.append(
ValidationFinding(
code=ValidationCode.EMPTY_REASONING,
severity=Severity.HARD,
message="reasoning missing or blank",
location=candidate.candidate_id,
)
)
continue
if any(len(clause.evidence) == 0 for clause in reasoning.clauses):
out.append(
ValidationFinding(
code=ValidationCode.HALLUCINATION,
severity=Severity.HARD,
message="a reasoning clause carries no evidence",
location=candidate.candidate_id,
)
)
out.extend(self._rank_band_consistency(candidate))
rendered.append(reasoning.rendered)
skeletons.append(self._skeleton(reasoning.rendered))
out.extend(self._identical(ordered, rendered))
out.extend(self._templated(skeletons))
return tuple(out)
def _rank_band_consistency(
self, candidate: RankedCandidate
) -> tuple[ValidationFinding, ...]:
reasoning = candidate.reasoning
if reasoning is None:
return ()
expected = self._expected_band(candidate.rank)
out: list[ValidationFinding] = []
if reasoning.rank_band != expected:
out.append(
ValidationFinding(
code=ValidationCode.RANK_REASONING_MISMATCH,
severity=Severity.SOFT,
message=(
f"rank {candidate.rank} band {reasoning.rank_band!r} "
f"!= expected {expected!r}"
),
location=candidate.candidate_id,
)
)
has_strength = any(
c.polarity is ReasoningPolarity.STRENGTH for c in reasoning.clauses
)
if expected in ("top", "mid") and not has_strength:
out.append(
ValidationFinding(
code=ValidationCode.RANK_REASONING_MISMATCH,
severity=Severity.SOFT,
message=f"{expected}-band candidate has no strength clause",
location=candidate.candidate_id,
)
)
return tuple(out)
def _expected_band(self, rank: int) -> str:
top_cut = max(1, round(self.expected_size * 0.10))
mid_cut = max(top_cut + 1, round(self.expected_size * 0.50))
if rank <= top_cut:
return "top"
if rank <= mid_cut:
return "mid"
return "tail"
@staticmethod
def _identical(
ordered: tuple[RankedCandidate, ...], rendered: list[str]
) -> tuple[ValidationFinding, ...]:
if len(rendered) < 2:
return ()
counts = Counter(rendered)
out: list[ValidationFinding] = []
for text, freq in sorted(counts.items()):
if freq > 1:
out.append(
ValidationFinding(
code=ValidationCode.IDENTICAL_REASONING,
severity=Severity.HARD,
message=f"{freq} candidates share identical reasoning text",
location=None,
)
)
return tuple(out)
@staticmethod
def _templated(skeletons: list[str]) -> tuple[ValidationFinding, ...]:
if len(skeletons) < 2:
return ()
counts = Counter(skeletons)
_, top_freq = counts.most_common(1)[0]
fraction = top_freq / len(skeletons)
if fraction >= _TEMPLATED_FRACTION:
return (
ValidationFinding(
code=ValidationCode.TEMPLATED_REASONING,
severity=Severity.HARD,
message=(
f"{top_freq}/{len(skeletons)} renderings collapse to one "
f"templated skeleton ({fraction:.0%})"
),
location=None,
),
)
return ()
@staticmethod
def _skeleton(text: str) -> str:
"""Normalize ids and numbers out so genuine variation survives but a fixed
template (varying only by number/id) collapses to a single skeleton."""
without_ids = _ID_TOKEN.sub("<id>", text)
without_nums = _NUM_TOKEN.sub("<n>", without_ids)
return " ".join(without_nums.lower().split())
__all__: tuple[str, ...] = ("ValidationEngine",)
|