feat: Intelligence Pipeline v3 — full implementation

Multi-stage evidence-grounded inference architecture replacing the monolithic 9B model extraction pipeline. CPU-first specialist services handle routine extraction while the 9B vLLM model is preserved for semantic adjudication of ambiguous cases. Key components: - Capability-aware inference gateway (OpenAI-compatible + Ollama) - Endpoint registry with DB migrations and REST API - Sentence-aware document segmenter (property tests) - Deterministic financial parsing with offset integrity - Symbol resolution with ambiguity detection - Specialist service (GLiNER2, dynamic batching, K8s deployment) - Company-specific sentiment (FinBERT, calibration) - Retrieval-based novelty and duplicate detection - Confidence calibration pipeline - Deterministic routing engine (property tests) - 9B adjudication layer with VRAM gating - Stock-specific impact model (features, labels, baseline, trained) - Pipeline orchestrator (state machine, queues, leases, feature flags) - Bounded parallelism (async workers, semaphore, load shedding) - Observability (tracing, metrics, alerts) - Compatibility adapter (v3→v2 golden mapping tests) - Shadow/canary promotion framework - Active learning and fine-tuning pipeline Test results: 1,161 tests pass, ruff lint clean. All 282 spec tasks completed.
2026-07-13 02:14:59 +00:00
parent 84634a365e
commit a72f336ad1
227 changed files with 50403 additions and 0 deletions
@@ -0,0 +1,142 @@
+"""Conservative confidence defaults for underrepresented classes.
+
+When calibration data is insufficient for a specific document type or
+event class, returns conservative values (0.3-0.5) and marks the result
+as under-calibrated per Requirement 10.7.
+"""
+
+from __future__ import annotations
+
+import logging
+
+from services.intelligence_pipeline_v3.confidence.models import ConfidenceResult
+
+logger = logging.getLogger(__name__)
+
+# Conservative default probabilities by document type.
+# These are intentionally low (0.3-0.5) to avoid overconfidence
+# when insufficient calibration data exists.
+_DOCUMENT_TYPE_DEFAULTS: dict[str, float] = {
+    "news": 0.45,
+    "filing": 0.40,
+    "transcript": 0.40,
+    "press_release": 0.45,
+    "macro_event": 0.35,
+    "unknown": 0.30,
+}
+
+# Conservative default probabilities by event class.
+# More complex or rare event types get lower defaults.
+_EVENT_CLASS_DEFAULTS: dict[str, float] = {
+    "earnings_beat": 0.50,
+    "earnings_miss": 0.50,
+    "guidance_raise": 0.45,
+    "guidance_cut": 0.45,
+    "merger_acquisition": 0.40,
+    "product_launch": 0.45,
+    "regulatory_action": 0.40,
+    "management_change": 0.45,
+    "legal_proceeding": 0.40,
+    "supply_chain": 0.35,
+    "rating_change": 0.45,
+    "dividend_change": 0.45,
+    "buyback": 0.45,
+    "macro_policy": 0.35,
+    "geopolitical": 0.30,
+    "sector_rotation": 0.35,
+    "unknown": 0.30,
+}
+
+# Features used when returning conservative defaults
+_DEFAULT_FEATURES_USED = [
+    "document_type_prior",
+    "event_class_prior",
+]
+
+
+def get_default_confidence(
+    document_type: str,
+    event_class: str,
+) -> ConfidenceResult:
+    """Return a conservative confidence result for underrepresented classes.
+
+    Used when calibration data is insufficient for the given document type
+    and event class combination. Returns conservative probabilities (0.3-0.5)
+    and marks the result as under-calibrated.
+
+    Parameters
+    ----------
+    document_type
+        The document type (news, filing, transcript, etc.).
+    event_class
+        The classified event type (earnings_beat, merger_acquisition, etc.).
+
+    Returns
+    -------
+    ConfidenceResult
+        A conservative confidence result with under_calibrated=True.
+    """
+    doc_default = _DOCUMENT_TYPE_DEFAULTS.get(
+        document_type, _DOCUMENT_TYPE_DEFAULTS["unknown"]
+    )
+    event_default = _EVENT_CLASS_DEFAULTS.get(
+        event_class, _EVENT_CLASS_DEFAULTS["unknown"]
+    )
+
+    # Take the minimum of document and event defaults for extra conservatism
+    probability = min(doc_default, event_default)
+
+    logger.debug(
+        "Using conservative default confidence: doc_type=%s (%.2f), event=%s (%.2f) -> %.2f",
+        document_type,
+        doc_default,
+        event_class,
+        event_default,
+        probability,
+    )
+
+    return ConfidenceResult(
+        probability=probability,
+        features_used=_DEFAULT_FEATURES_USED,
+        is_calibrated=False,
+        under_calibrated=True,
+        calibration_version="conservative-default-v1",
+    )
+
+
+def is_underrepresented(
+    document_type: str,
+    event_class: str,
+    min_samples: int = 30,
+    known_counts: dict[tuple[str, str], int] | None = None,
+) -> bool:
+    """Check if a document_type + event_class combination is underrepresented.
+
+    Parameters
+    ----------
+    document_type
+        The document type.
+    event_class
+        The event class.
+    min_samples
+        Minimum number of calibration samples to consider a class well-represented.
+    known_counts
+        Optional mapping of (doc_type, event_class) -> sample count.
+        If None, treats any unknown combination as underrepresented.
+
+    Returns
+    -------
+    bool
+        True if the class has insufficient calibration data.
+    """
+    if known_counts is None:
+        # Without explicit counts, use heuristic: unknown types are underrepresented
+        if document_type not in _DOCUMENT_TYPE_DEFAULTS:
+            return True
+        if event_class not in _EVENT_CLASS_DEFAULTS:
+            return True
+        return False
+
+    key = (document_type, event_class)
+    count = known_counts.get(key, 0)
+    return count < min_samples