Spaces:
Sleeping
Sleeping
File size: 22,580 Bytes
feb1b1c 8b10520 feb1b1c fc87e83 feb1b1c 8b10520 feb1b1c 830d137 feb1b1c 830d137 feb1b1c 830d137 feb1b1c fc87e83 feb1b1c 8b10520 f0894e2 8b10520 9f858f4 f0894e2 9f858f4 8b10520 | 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 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237 238 239 240 241 242 243 244 245 246 247 248 249 250 251 252 253 254 255 256 257 258 259 260 261 262 263 264 265 266 267 268 269 270 271 272 273 274 275 276 277 278 279 280 281 282 283 284 285 286 287 288 289 290 291 292 293 294 295 296 297 298 299 300 301 302 303 304 305 306 307 308 309 310 311 312 313 314 315 316 317 318 319 320 321 322 323 324 325 326 327 328 329 330 331 332 333 334 335 336 337 338 339 340 341 342 343 344 345 346 347 348 349 350 351 352 353 354 355 356 357 358 359 360 361 362 363 364 365 366 367 368 369 370 371 372 373 374 375 376 377 378 379 380 381 382 383 384 385 386 387 388 389 390 391 392 393 394 395 396 397 398 399 400 401 402 403 404 405 406 407 408 409 410 411 412 413 414 415 416 417 418 419 420 421 422 423 424 425 426 427 428 429 430 431 432 433 434 435 436 437 438 439 440 441 442 443 444 445 446 447 448 449 450 451 452 453 454 455 456 457 458 459 460 461 462 463 464 465 466 467 468 469 470 471 472 473 474 475 476 477 478 479 480 481 482 483 484 485 486 487 488 489 490 491 492 493 494 495 496 497 498 499 500 501 502 503 504 505 506 507 508 509 510 511 512 513 514 515 516 517 518 519 520 521 522 523 524 525 526 527 528 529 530 531 532 533 534 535 536 537 538 539 540 541 542 543 544 545 546 547 548 549 550 551 552 553 554 555 556 557 558 559 560 561 562 563 564 565 566 567 568 569 570 571 572 573 574 575 576 577 578 579 580 | from __future__ import annotations
import enum
from datetime import datetime
from pydantic import (
BaseModel,
ConfigDict,
Field,
field_validator,
model_validator,
)
from redstack.domain.enums import IntegrityFlag, LocationFit, NoticeFit, Severity
__all__ = [
"RunMode",
"Profile",
"LogLevel",
"MalformedRecordPolicy",
"ScoreTransformKind",
"AnchorPolarity",
"HoneypotSeverity",
"DeterminismConfig",
"BudgetConfig",
"PathsConfig",
"LoggingConfig",
"ScorePresentationConfig",
"OnlineRuntimeConfig",
"OfflineRuntimeConfig",
"RedstackConfig",
"ScoringWeightsConfig",
"ConceptSeed",
"LexiconSeedConfig",
"CompiledLexiconConcept",
"CompiledLexicon",
"AnchorIntent",
"JdAnchorsConfig",
"EligibilityRule",
"EligibilityRulesConfig",
"HoneypotRule",
"HoneypotRulesConfig",
"IntegrityThresholds",
"EligibilityRuleSet",
"ScoringPolicy",
"BehavioralPolicy",
"LogisticsPolicy",
]
# --------------------------------------------------------------------------- #
# Closed config-layer vocabularies (IO/authoring concepts, not domain VOs). #
# --------------------------------------------------------------------------- #
class RunMode(enum.StrEnum):
"""Which runtime layer the loader composes (``runtime/<mode>.yaml``)."""
ONLINE = "online"
OFFLINE = "offline"
class Profile(enum.StrEnum):
"""The optional final override layer (``profiles/<profile>.yaml``)."""
CI = "ci"
LOCAL = "local"
class LogLevel(enum.StrEnum):
"""Structured-logging verbosity (see :mod:`redstack.observability.logging`)."""
DEBUG = "DEBUG"
INFO = "INFO"
WARNING = "WARNING"
ERROR = "ERROR"
class MalformedRecordPolicy(enum.StrEnum):
"""Ingestion policy for a ``Malformed`` source record (online R1 / offline O0).
``SKIP`` drops the record as data and continues; ``ABORT`` fails the run.
"""
SKIP = "skip"
ABORT = "abort"
class ScoreTransformKind(enum.StrEnum):
"""How an internal base score is presented in the submission CSV column."""
IDENTITY = "identity"
class AnchorPolarity(enum.StrEnum):
"""Authoring polarity of a ``jd.*`` anchor intent (Architecture §13, O6)."""
POSITIVE = "positive"
NEGATIVE = "negative"
class HoneypotSeverity(enum.StrEnum):
"""Authoring severity of a honeypot rule shape.
``HARD`` rules contribute to the ``>= 2 HARD`` honeypot decision; ``SOFT``
rules feed the composite. O3 calibrates the numeric thresholds against the
census; this seed only declares the rule *shape*.
"""
HARD = "hard"
SOFT = "soft"
# --------------------------------------------------------------------------- #
# Base model: every config VO is frozen and rejects unknown keys. #
# --------------------------------------------------------------------------- #
class _FrozenConfig(BaseModel):
"""Shared base enforcing the frozen / ``extra="forbid"`` discipline."""
model_config = ConfigDict(frozen=True, extra="forbid")
# --------------------------------------------------------------------------- #
# Runtime / IO config sections. #
# --------------------------------------------------------------------------- #
class DeterminismConfig(_FrozenConfig):
"""Seed and thread pins owned by :mod:`redstack.config.determinism`.
``omp_num_threads`` and ``mkl_num_threads`` default to ``1`` to guarantee
thread-count-invariant, byte-identical CSV outputs. They are configurable
only so the determinism test suite can prove 1-thread vs N-thread ranking
identity; production runs keep them pinned to ``1``.
"""
seed: int = Field(default=20240601, ge=0)
omp_num_threads: int = Field(default=1, ge=1)
mkl_num_threads: int = Field(default=1, ge=1)
onnx_intra_op_threads: int = Field(default=1, ge=1)
onnx_inter_op_threads: int = Field(default=1, ge=1)
class BudgetConfig(_FrozenConfig):
"""The hard online runtime budget asserted by the timing guard.
``max_wall_seconds`` is the disqualifying ceiling (300 s); the internal
``target_wall_seconds`` (130 s) is the regression alarm. ``max_rss_mb`` is
the 16 GB memory ceiling.
"""
max_wall_seconds: float = Field(default=300.0, gt=0.0)
target_wall_seconds: float = Field(default=130.0, gt=0.0)
max_rss_mb: int = Field(default=16384, gt=0)
@model_validator(mode="after")
def _target_within_ceiling(self) -> BudgetConfig:
"""The internal target must not exceed the disqualifying ceiling."""
if self.target_wall_seconds > self.max_wall_seconds:
msg = (
"target_wall_seconds "
f"({self.target_wall_seconds}) must be <= max_wall_seconds "
f"({self.max_wall_seconds})"
)
raise ValueError(msg)
return self
class PathsConfig(_FrozenConfig):
"""Filesystem locations. Paths are kept as ``str`` for byte-stable hashing.
``configs/`` is intent and ``data/`` is raw fact (read-only); only the
offline pipeline writes ``artifacts/``. No code writes ``data/`` or
``configs/``.
"""
artifacts_root: str = Field(default="artifacts", min_length=1)
data_root: str = Field(default="data", min_length=1)
candidates_path: str = Field(default="data/raw/candidates.jsonl", min_length=1)
golden_labels_path: str = Field(
default="data/golden/golden_labels.csv", min_length=1
)
manifest_path: str = Field(default="artifacts/MANIFEST.json", min_length=1)
submission_path: str = Field(default="submission.csv", min_length=1)
run_report_path: str = Field(default="run_report.json", min_length=1)
class LoggingConfig(_FrozenConfig):
"""Logging policy. ``deterministic`` strips timestamps from repro-relevant lines."""
level: LogLevel = LogLevel.INFO
deterministic: bool = True
class ScorePresentationConfig(_FrozenConfig):
"""How scores reach the CSV: transform, decimal precision, FLOOR sentinel.
``floor_sentinel`` is the value assigned to integrity/eligibility-gated
candidates so a gated record can never outrank a genuine one.
"""
transform: ScoreTransformKind = ScoreTransformKind.IDENTITY
decimals: int = Field(default=6, ge=0, le=12)
floor_sentinel: float = Field(default=0.0)
class OnlineRuntimeConfig(_FrozenConfig):
"""Online-only knobs (``runtime/online.yaml``), consumed at R0.
Holds the injected ``as_of`` clock (no module reads the OS clock for logic),
the top-K cut, the malformed-record policy, and the score presentation.
"""
as_of: datetime
top_k: int = Field(default=100, ge=1)
malformed_record_policy: MalformedRecordPolicy = MalformedRecordPolicy.SKIP
score_presentation: ScorePresentationConfig = Field(
default_factory=ScorePresentationConfig
)
@field_validator("as_of")
@classmethod
def _as_of_must_be_aware(cls, value: datetime) -> datetime:
"""Reject naive datetimes; ``as_of`` must carry an explicit timezone."""
if value.tzinfo is None or value.tzinfo.utcoffset(value) is None:
msg = "online.as_of must be timezone-aware (include an explicit offset)"
raise ValueError(msg)
return value
class OfflineRuntimeConfig(_FrozenConfig):
"""Offline build parameters (``runtime/offline.yaml``).
Holds the injected ``as_of`` clock (so recency-dependent feature
extraction in O14 is byte-stable across rebuilds, mirroring why
:class:`OnlineRuntimeConfig` injects its own), the offline RNG seed, the
pinned SentenceTransformer model id and revision, the KMeans ``k`` for
archetype discovery (O7), the listwise weight-search budget (O9), and
batch sizes for embedding/feature passes.
"""
as_of: datetime
seed: int = Field(default=20240601, ge=0)
st_model_id: str = Field(min_length=1)
st_model_revision: str = Field(min_length=1)
kmeans_k: int = Field(ge=1)
search_budget: int = Field(ge=1)
embedding_batch_size: int = Field(default=256, ge=1)
feature_batch_size: int = Field(default=4096, ge=1)
@field_validator("as_of")
@classmethod
def _as_of_must_be_aware(cls, value: datetime) -> datetime:
"""Reject naive datetimes; ``as_of`` must carry an explicit timezone."""
if value.tzinfo is None or value.tzinfo.utcoffset(value) is None:
msg = "offline.as_of must be timezone-aware (include an explicit offset)"
raise ValueError(msg)
return value
class RedstackConfig(_FrozenConfig):
"""The fully-composed, validated runtime configuration.
Produced by :func:`redstack.config.loader.load_config` from the deterministic
deep-merge of ``base -> runtime/<mode> -> profiles/<profile>``. Exactly one
of :attr:`online` / :attr:`offline` is populated, matching :attr:`run_mode`.
This is the typed object every pipeline and CLI verb receives; raw ``dict``
config never crosses a module boundary.
"""
schema_version: str = Field(min_length=1)
run_mode: RunMode
profile: Profile | None = None
determinism: DeterminismConfig = Field(default_factory=DeterminismConfig)
budget: BudgetConfig = Field(default_factory=BudgetConfig)
paths: PathsConfig = Field(default_factory=PathsConfig)
logging: LoggingConfig = Field(default_factory=LoggingConfig)
online: OnlineRuntimeConfig | None = None
offline: OfflineRuntimeConfig | None = None
@model_validator(mode="after")
def _mode_block_consistency(self) -> RedstackConfig:
"""Exactly the run-mode's block must be present; the other must be absent."""
if self.run_mode is RunMode.ONLINE:
if self.online is None:
raise ValueError("run_mode=online requires an 'online' block")
if self.offline is not None:
raise ValueError("run_mode=online forbids an 'offline' block")
else:
if self.offline is None:
raise ValueError("run_mode=offline requires an 'offline' block")
if self.online is not None:
raise ValueError("run_mode=offline forbids an 'online' block")
return self
# --------------------------------------------------------------------------- #
# Behaviour / authoring seeds (consumed by the offline pipeline). #
# --------------------------------------------------------------------------- #
class ScoringWeightsConfig(_FrozenConfig):
"""Seed/candidate scoring weights — the O9 search *input*, not the contract.
Keys are score-component identifiers (mapped to ``domain.ScoreComponent`` at
the O9/artifact boundary). The frozen, validated weights consumed online
live in ``artifacts/weights/scoring_weights.locked.yaml``, never here, which
prevents "I tweaked the YAML and the leaderboard moved" drift.
"""
layout_version: str = Field(min_length=1)
weights: dict[str, float] = Field(min_length=1)
@field_validator("weights")
@classmethod
def _weights_finite(cls, value: dict[str, float]) -> dict[str, float]:
"""Every seed weight must be a finite real number."""
import math
for component, weight in value.items():
if not math.isfinite(weight):
msg = f"scoring weight for {component!r} is not finite: {weight!r}"
raise ValueError(msg)
return value
class ConceptSeed(_FrozenConfig):
"""Human seed terms for one lexicon concept (mined by O4, expanded by O5)."""
seed_terms: tuple[str, ...] = Field(min_length=1)
@field_validator("seed_terms")
@classmethod
def _terms_non_blank(cls, value: tuple[str, ...]) -> tuple[str, ...]:
"""Seed terms must be non-blank after stripping."""
if any(not term.strip() for term in value):
raise ValueError("lexicon seed_terms must not contain blank entries")
return value
class LexiconSeedConfig(_FrozenConfig):
"""The ``lexicon/lexicon.seed.yaml`` authoring seed: concept -> seed terms."""
concepts: dict[str, ConceptSeed] = Field(min_length=1)
class CompiledLexiconConcept(_FrozenConfig):
"""One concept's matchable surface forms inside the compiled lexicon.
``terms`` is the bag-of-words membership set mined by O4 (single tokens);
``phrases`` is the set of multi-word surface forms admitted for that
concept. Both are consulted for plain substring/token-set matching by
:class:`redstack.engines.lexicon.LexiconEngine`.
"""
terms: frozenset[str] = Field(default_factory=frozenset)
phrases: tuple[str, ...] = Field(default_factory=tuple)
class CompiledLexicon(_FrozenConfig):
"""The O4 ``lexicon_compiled`` artifact: concept id -> matchable surface forms."""
concepts: dict[str, CompiledLexiconConcept] = Field(default_factory=dict)
class AnchorIntent(_FrozenConfig):
"""One ``jd.*`` anchor intent: a stable id, polarity, and the anchor text.
The text is embedded offline (O6) into ``anchor_vectors.npy``; the id is the
key the online ``SemanticEngine`` resolves cosine similarity against.
"""
id: str = Field(min_length=1)
polarity: AnchorPolarity
text: str = Field(min_length=1)
class JdAnchorsConfig(_FrozenConfig):
"""The ``anchors/jd_anchors.yaml`` authoring seed (re-authored per JD)."""
anchors: tuple[AnchorIntent, ...] = Field(min_length=1)
@model_validator(mode="after")
def _ids_unique_and_polarities_present(self) -> JdAnchorsConfig:
"""Anchor ids must be unique and both polarities must be represented."""
ids = [anchor.id for anchor in self.anchors]
if len(set(ids)) != len(ids):
raise ValueError("jd anchor ids must be unique")
polarities = {anchor.polarity for anchor in self.anchors}
if AnchorPolarity.POSITIVE not in polarities:
raise ValueError("jd anchors require at least one positive intent")
if AnchorPolarity.NEGATIVE not in polarities:
raise ValueError("jd anchors require at least one negative intent")
return self
class EligibilityRule(_FrozenConfig):
"""A declarative JD eligibility rule keyed by an eligibility code.
The code string is cross-checked against ``domain.EligibilityCode`` when O6
packages this seed into the artifact (Repository Layout §5). ``penalty`` is
populated for soft penalties (in ``[0, 1]``) and omitted for hard blocks.
"""
code: str = Field(min_length=1)
description: str = Field(min_length=1)
penalty: float | None = Field(default=None, ge=0.0, le=1.0)
class EligibilityRulesConfig(_FrozenConfig):
"""The ``gates/eligibility_rules.yaml`` authoring seed.
Hard blocks carry no penalty (they disqualify); soft penalties must each
declare a bounded penalty weight.
"""
hard_blocks: tuple[EligibilityRule, ...] = Field(default=())
soft_penalties: tuple[EligibilityRule, ...] = Field(default=())
@model_validator(mode="after")
def _shape_and_uniqueness(self) -> EligibilityRulesConfig:
"""Hard blocks omit penalties; soft penalties require them; codes unique."""
for rule in self.hard_blocks:
if rule.penalty is not None:
raise ValueError(f"hard block {rule.code!r} must not carry a penalty")
for rule in self.soft_penalties:
if rule.penalty is None:
raise ValueError(
f"soft penalty {rule.code!r} must declare a penalty weight"
)
codes = [r.code for r in (*self.hard_blocks, *self.soft_penalties)]
if len(set(codes)) != len(codes):
raise ValueError("eligibility rule codes must be unique")
if not codes:
raise ValueError("eligibility rules must define at least one rule")
return self
class HoneypotRule(_FrozenConfig):
"""A human-declared honeypot rule shape keyed by an integrity flag.
The flag string is cross-checked against ``domain.IntegrityFlag`` at the O3
artifact boundary. O3 calibrates the numeric threshold against the census;
this seed declares only the rule's identity and severity.
"""
flag: str = Field(min_length=1)
severity: HoneypotSeverity
description: str = Field(min_length=1)
class HoneypotRulesConfig(_FrozenConfig):
"""The ``integrity/honeypot_rules.yaml`` authoring seed."""
rules: tuple[HoneypotRule, ...] = Field(min_length=1)
@model_validator(mode="after")
def _flags_unique(self) -> HoneypotRulesConfig:
"""Integrity flags must be unique across declared rules."""
flags = [rule.flag for rule in self.rules]
if len(set(flags)) != len(flags):
raise ValueError("honeypot rule flags must be unique")
return self
# --------------------------------------------------------------------------- #
# Calibrated artifact-shaped policy models — consumed *online* by the engines #
# (sourced from compiled O3/O6/O9/O11 artifacts at R0, never hand-authored). #
# --------------------------------------------------------------------------- #
class IntegrityThresholds(_FrozenConfig):
"""Calibrated honeypot thresholds (O3 artifact); consumed by ``IntegrityEngine``.
Sourced from ``artifacts/calibration/integrity_thresholds.json``. Per-flag
``flag_severity``/``flag_weights`` default to empty (the engine falls back
to ``Severity.HARD`` / weight ``0.0`` for any flag absent from the map).
"""
honeypot_threshold: float = Field(ge=0.0, le=1.0)
flag_severity: dict[IntegrityFlag, Severity] = Field(default_factory=dict)
flag_weights: dict[IntegrityFlag, float] = Field(default_factory=dict)
tolerance_experience_years: float = Field(ge=0.0)
duration_date_tolerance_months: float = Field(ge=0.0)
expert_zero_usage_min_count: int = Field(ge=1)
experience_predates_tolerance_years: int = Field(ge=0)
class EligibilityRuleSet(_FrozenConfig):
"""Calibrated eligibility-gate thresholds (O6 artifact); consumed by
``EligibilityEngine``.
Sourced from the compiled ``gates/eligibility_rules.yaml`` artifact — the
JD-derived numeric thresholds behind each hard block / soft penalty.
"""
research_min_semantic_fit: float = Field(ge=0.0, le=1.0)
framework_only_stuffing_min: float = Field(ge=0.0, le=1.0)
framework_only_gap_min: float = Field(ge=0.0, le=1.0)
production_recency_max_months: int = Field(ge=0)
adjacent_domain_min_relevant_credibility: float = Field(ge=0.0, le=1.0)
adjacent_domain_min_negative_fit: float = Field(ge=-1.0, le=1.0)
adjacent_domain_min_nlp_ir_exposure: float = Field(ge=0.0, le=1.0)
closed_source_min_years: float = Field(ge=0.0)
closed_source_max_credible_skills: int = Field(ge=0)
title_chaser_min_hop_rate: float = Field(ge=0.0, le=1.0)
experience_band_min_years: float = Field(ge=0.0)
experience_band_max_years: float = Field(ge=0.0)
def default_eligibility_rules() -> EligibilityRuleSet:
"""The locked ``EligibilityRuleSet`` thresholds.
``gates/eligibility_rules.yaml`` (O6) carries only human descriptions per
code -- O6 never calibrates numeric thresholds, it authors rule *shape*
(no gold-labeled data exists for these). The literal values below are
therefore the single source of truth both the offline (O13a) and online
(R0) paths construct their ``EligibilityEngine`` from, so the two stay in
lockstep by construction rather than by two hand-synced literals.
"""
return EligibilityRuleSet(
research_min_semantic_fit=0.5,
framework_only_stuffing_min=0.6,
framework_only_gap_min=0.3,
production_recency_max_months=18,
# Raised from 0.3: diagnostic evidence across 27 real candidates found
# a clean gap between domain-irrelevant titles (HR Manager, Civil
# Engineer, Accountant, ... — 0.11-0.50) and genuinely ML-titled ones
# (Senior ML Engineer / Applied Scientist / RecSys Engineer —
# 0.80-0.94); 0.6 sits in that gap with margin on both sides.
adjacent_domain_min_relevant_credibility=0.6,
adjacent_domain_min_negative_fit=0.3,
# Same near-zero-noise-vs-any-real-signal floor as the prior
# nine-group skill_match version, now measured over the six groups
# (retr/rank/recsys/ir/nlp/eval) that actually constitute NLP/IR
# exposure -- excluding llm/mle/mlops closes the regression where a
# CV/speech/robotics specialist's genuine-but-irrelevant competency
# in those three groups cleared the old nine-group floor on its own.
adjacent_domain_min_nlp_ir_exposure=0.05,
closed_source_min_years=5.0,
closed_source_max_credible_skills=0,
title_chaser_min_hop_rate=0.5,
experience_band_min_years=2.0,
experience_band_max_years=15.0,
)
class ScoringPolicy(_FrozenConfig):
"""Scoring combination policy (floor + neutral prior); consumed by
``ScoringEngine``."""
floor: float = Field(default=0.0)
neutral_prior: float = Field(default=0.5, ge=0.0, le=1.0)
class BehavioralPolicy(_FrozenConfig):
"""Behavioral-multiplier policy (O11 artifact); consumed by ``BehavioralEngine``."""
family_weights: dict[str, float] = Field(default_factory=dict)
unknown_neutral_base: float = Field(default=0.5, ge=0.0, le=1.0)
m_min: float = Field(ge=0.0, le=1.0)
m_max: float = Field(ge=0.0, le=1.0)
@model_validator(mode="after")
def _bounds_ordered(self) -> BehavioralPolicy:
if self.m_min > self.m_max:
raise ValueError("m_min must not exceed m_max")
return self
class LogisticsPolicy(_FrozenConfig):
"""Logistics-multiplier policy; consumed by ``LogisticsEngine``."""
location_fit_factor: dict[LocationFit, float] = Field(default_factory=dict)
location_default_factor: float = Field(default=1.0)
notice_fit_factor: dict[NoticeFit, float] = Field(default_factory=dict)
notice_default_factor: float = Field(default=1.0)
work_mode_weight: float = Field(default=0.0, ge=0.0, le=1.0)
salary_inversion_factor: float = Field(default=1.0, ge=0.0, le=1.0)
m_min: float = Field(ge=0.0, le=1.0)
m_max: float = Field(ge=0.0, le=1.0)
@model_validator(mode="after")
def _bounds_ordered(self) -> LogisticsPolicy:
if self.m_min > self.m_max:
raise ValueError("m_min must not exceed m_max")
return self
|