API Reference¶

Primary reference for the supported public classes, functions, dataclasses, and operator-facing APIs in Director-AI. Internal helpers, generated protobuf stubs, and compatibility shims are intentionally excluded unless they define a supported integration boundary.

API Selection Model¶

Choose the API by the boundary you control:

Boundary you own	Use	Why
Application code	`guard()`, `score()`, `CoherenceScorer`	Smallest path to reject or annotate one output
Token stream	`StreamingKernel`, `AsyncStreamingKernel`, `/v1/stream`	Route contradictory partial output before completion
Knowledge base	`DocumentIngestionPipeline`, `VectorGroundTruthStore`	Tie verdicts to governed private facts
Shared service	REST server or gRPC server	Centralise auth, tenant binding, metrics, and rollout controls
Evaluation pipeline	`BatchProcessor`, policy evaluation, conformal routing	Compare thresholds, models, and regression gates
Governance surface	`SafetyEvent`, human review, compliance reports	Preserve tenant-safe evidence for review and audits

Pick The Right API First¶

Need	API	Start
Protect an existing SDK client	`guard()`	guard / score / get_score
Score one prompt and answer	`score()`	guard / score / get_score
Route streamed contradictions	`StreamingKernel`	StreamingKernel
Build an agent workflow	`CoherenceAgent`	CoherenceAgent
Process evaluation batches	`BatchProcessor`	BatchProcessor
Serve over HTTP	REST server	REST Server
Serve over gRPC	CoherenceScoring service	gRPC Server
Ground answers in documents	`DocumentIngestionPipeline` and `VectorGroundTruthStore`	Ingestion
Package customer evidence	Customer Model Factory	Customer Model Factory
Enforce finance-sector policy checks	Financial Services Policy API	Financial Services Policy API

For the product and pilot context around these APIs, see Applications and Market Map, Product Overview, and Evaluation Onboarding.

Entry Points¶

Symbol	Module	Purpose
`guard()`	`director_ai`	Wrap an LLM SDK client with coherence scoring
`score()`	`director_ai`	Score a single prompt/response pair
`get_score()`	`director_ai`	Retrieve last score from `on_fail="metadata"`
`VoiceGuard`	`director_ai.integrations.voice`	TTS scoring boundary with opt-in streaming contradiction halt

Core Classes¶

Class	Module	Purpose
`CoherenceScorer`	`director_ai.core.scoring.scorer`	Dual-entropy coherence scoring engine
`StreamingKernel`	`director_ai.core.runtime.streaming`	Opt-in streaming contradiction halt
`AsyncStreamingKernel`	`director_ai.core.runtime.async_streaming`	Async variant of StreamingKernel
`HumanReviewQueue`	`director_ai.core.runtime.human_review`	Durable reviewer approval, retry, and release gate
`NoGoPolicy`	`director_ai.core.guard_control`	Calibrated no-go policy with default irreversibility forecasting
`evaluate_policy_variants()`	`director_ai.core.evaluation.policy`	Controlled profile and threshold comparison on labelled data
`build_causal_attribution_graph()`	`director_ai.core.attribution.causal_graph`	Evidence, claim, halt-trace, and counterfactual attribution DAGs
`AdaptiveThresholdLearner`	`director_ai.core.calibration.adaptive_threshold`	Human-gated Thompson-sampling threshold recommendations
`DefenseUpdatePipeline`	`director_ai.core.defense_genome`	Reviewed defence promotion across feedback, adversarial mining, and registry hot-swap
`AutoRedteamDefenceLoop`	`director_ai.core.defense_genome`	Repeated adversarial-mining cycles with detection-uplift gates before promotion
`build_edge_runtime_readiness()`	`director_ai.core.edge`	Edge/mobile readiness profile for WASM, Rust, ONNX, and smoke evidence
`assess_banking_response()`	`director_ai.core.financial_services`	Evidence, numeric-claim, deposit-limit, complaint, and investment-recommendation checks
`CrossDocumentConsistencyMemory`	`director_ai.core.memory.consistency`	Tenant-scoped long-term consistency checks with retention and delete controls
`DifferentialPrivacyScoreReleaser`	`director_ai.core.federated_privacy.score_release`	Laplace-noised score disclosure with privacy accounting
`FederatedSafetySignalAggregator`	`director_ai.core.federated_privacy.signal_sharing`	Anonymous DP aggregate sharing for tenant-safe guard signals
`NeuroSymbolicVerifier`	`director_ai.core.verification.neuro_symbolic`	Neural score fusion with numeric and formal symbolic checks
`ByzantineFaultTolerantConsensus`	`director_ai.core.scoring.consensus`	PBFT-style quorum over independent verifier votes
`InferenceServerHook`	`director_ai.integrations.inference_server_hooks`	Server-neutral pre-sampling hook for vLLM, TGI, and llama.cpp
`createDirectorAiMiddleware`	`@director-ai/vercel-ai`	Vercel AI SDK `LanguageModelV3Middleware` bridge to `/v1/review`
`AutoGenReplyGuard`	`director_ai.integrations.autogen_swarm`	AutoGen `register_reply()` hook for guarded group-chat messages
`build_guardrails_validator`	`director_ai.integrations.guardrails_ai`	Guardrails AI custom validator for coherence checks
`MetaGuard`	`director_ai.core.meta_guard`	Recursive guard drift monitor with production evasion gates
`SafetyEvent`	`director_ai.core.safety_event`	Tenant-safe telemetry schema and validator for guard decisions
`CoherenceAgent`	`director_ai.core.agent`	Orchestrator: generator + scorer + kernel
`BatchProcessor`	`director_ai.core.runtime.batch`	Concurrent batch scoring

Knowledge & Retrieval¶

Class	Module	Purpose
`GroundTruthStore`	`director_ai.core.retrieval.knowledge`	Key-value fact store (prototype)
`VectorGroundTruthStore`	`director_ai.core.retrieval.vector_store`	Semantic vector store with pluggable backends
`VectorBackend`	`director_ai.core.retrieval.vector_store`	Abstract backend protocol
`DocumentIngestionPipeline`	`director_ai.core.ingestion`	Parse, chunk, update, and delete documents for vector grounding

Configuration¶

Class	Module	Purpose
`DirectorConfig`	`director_ai.core.config`	Env var / YAML / profile configuration

Data Types¶

Class	Module	Purpose
`CoherenceScore`	`director_ai.core.types`	Score result with H_logical, H_factual, evidence, task type, confidence
`ReviewResult`	`director_ai.core.types`	Agent review output
`ScoringEvidence`	`director_ai.core.types`	Retrieved chunks + NLI details
`HaltEvidence`	`director_ai.core.types`	Structured halt reason with evidence
`CounterfactualHaltDiagnostic`	`director_ai.core.types`	Single-fact halt diagnostic
`CausalAttributionGraph`	`director_ai.core.attribution.causal_graph`	DAG representation of scorer and halt causal pathways
`SafetyEvent`	`director_ai.core.safety_event`	Tenant-safe halt and policy event schema
`DirectorSafetySignal`	`director_ai.core.safety_protocol`	Cross-runtime safety protocol envelope
`ConflictAwareKnowledgeGuard`	`director_ai.core.retrieval.conflict_guard`	Pre-ingestion KB conflict checks
`InterlockKernel`	`director_ai.interlock`	Standalone bring-your-own-scorer halt kernel
`TokenEvent`	`director_ai.core.runtime.streaming`	Per-token stream event
`StreamSession`	`director_ai.core.runtime.streaming`	Complete stream session state
`SustainabilityPolicyAdapter`	`director_ai.core.sustainability`	Token, cost, energy, carbon, quota, and forecast policy decisions
`SustainabilityTelemetry`	`director_ai.core.sustainability`	Per-tenant sustainability summaries and threshold alerts
`AgentPassportRegistry`	`director_ai.core.agent_identity`	Signed agent identity, capability policy, revocation, and coherence history
`TrajectoryRollbackManager`	`director_ai.core.trajectory`	Tenant-safe rollback hook registry for preflight halt and escalation outcomes

Verification And Guard APIs¶

Symbol	Module	Purpose
`verify_numeric()`	`director_ai.core.verification.numeric_verifier`	Numeric consistency checks (arithmetic, dates, probabilities)
`NeuroSymbolicVerifier`	`director_ai.core.verification.neuro_symbolic`	Neural + symbolic contradiction gate
`verify_reasoning_chain()`	`director_ai.core.verification.reasoning_verifier`	Reasoning chain logic (non-sequiturs, circularity)
`score_temporal_freshness()`	`director_ai.core.scoring.temporal_freshness`	Staleness risk for date-sensitive claims
`ConsensusScorer`	`director_ai.core.scoring.consensus`	Cross-model factual agreement
`CrossVerifierConsensus`	`director_ai.core.scoring.consensus`	Critical-domain verifier fusion with required coverage and calibrated risk interval
`ByzantineFaultTolerantConsensus`	`director_ai.core.scoring.consensus`	PBFT-style verifier vote quorum
`ConformalPredictor`	`director_ai.core.calibration.conformal`	Calibrated P(hallucination) intervals
`ConformalRoutingPolicy`	`director_ai.core.calibration.conformal`	Conservative allow, human-review, stronger-model, and reject routing from conformal risk bounds
`ConformalRoutingDecision`	`director_ai.core.calibration.conformal`	Evidence-carrying route decision for calibrated uncertainty
`FeedbackLoopDetector`	`director_ai.compliance.feedback_loop_detector`	EU AI Act Art 15(4) feedback loop detection
`LoopMonitor`	`director_ai.agentic.loop_monitor`	Agent loop safety (circular, drift, budget)
`AdversarialTester`	`director_ai.testing.adversarial_suite`	25-pattern adversarial robustness test

Interfaces¶

Interface	Purpose
REST Server	FastAPI endpoints (`/v1/review`, `/v1/health`, `/v1/metrics`, 8 gem endpoints)
gRPC Server	Protocol Buffers service (4 RPC methods)
CLI	23 command-line subcommands

Exceptions¶

Exception	Raised When
`HallucinationError`	`guard()` with `on_fail="raise"` detects low coherence
`KernelHaltError`	SafetyKernel halts the output stream
`ValidationError`	Invalid configuration or input
`DependencyError`	Required optional package missing

Meta-Confidence And Contradiction APIs¶

Symbol	Module	Purpose
`compute_meta_confidence()`	`core.scoring.meta_confidence`	Verdict confidence from margin + signal agreement
`ContradictionTracker`	`core.runtime.contradiction_tracker`	Pairwise cross-turn contradiction matrix
`ContradictionReport`	`core.runtime.contradiction_tracker`	Contradiction summary (worst pair, trend)
`CrossDocumentConsistencyMemory`	`core.memory.consistency`	Durable tenant-scoped contradiction memory across documents
`DifferentialPrivacyScoreReleaser`	`core.federated_privacy.score_release`	Optional DP layer for public score release
`FederatedSafetySignalAggregator`	`core.federated_privacy.signal_sharing`	Anonymous DP aggregate sharing for tenant-safe guard signals

See Meta-Confidence Guide.

Structured Output Verification APIs¶

Function	Module	Purpose
`verify_json()`	`core.verification.json_verifier`	JSON Schema validation + value grounding
`verify_tool_call()`	`core.verification.tool_call_verifier`	Tool existence, arg validation, fabrication detection
`verify_code()`	`core.verification.code_verifier`	Python syntax, import existence, hallucinated API detection

Result types: StructuredVerificationResult, ToolCallResult, CodeCheckResult, FieldVerdict.

See Structured Verification Guide.

Online Calibration APIs¶

Class	Module	Purpose
`FeedbackStore`	`core.calibration.feedback_store`	SQLite-backed human correction store
`OnlineCalibrator`	`core.calibration.online_calibrator`	Threshold sweep + FPR/FNR with Wilson CIs
`CalibrationReport`	`core.calibration.online_calibrator`	Calibration metrics dataclass
`HumanReviewQueue`	`core.runtime.human_review`	Reviewer case queue with append-only decisions
`AdaptiveThresholdLearner`	`core.calibration.adaptive_threshold`	Human-gated Thompson-sampling threshold recommender

See Online Calibration Guide.

Compliance APIs¶

Class	Module	Purpose
`AuditLog`	`compliance.audit_log`	SQLite audit trail for every scored interaction
`AuditEntry`	`compliance.audit_log`	Single interaction record
`ComplianceReporter`	`compliance.reporter`	Article 15 report generator
`Article15Report`	`compliance.reporter`	Structured report with metrics, drift, incidents
`Article15TemplateContext`	`compliance.reporter`	Operator-supplied Article 15 technical documentation context
`Soc2IsoReadinessReport`	`compliance.readiness`	Tenant-safe SOC 2 / ISO 27001 readiness report with SOC 2 Type I path guidance
`Soc2IsoControl`	`compliance.readiness`	One readiness control mapped to SOC 2, ISO 27001, and optional HIPAA Security Rule evidence
`HipaaDocumentationPacket`	`compliance.readiness`	Tenant-safe HIPAA documentation readiness packet
`HipaaDeploymentObligation`	`compliance.readiness`	Operator-owned HIPAA obligation with 45 CFR Part 164 references
`build_soc2_iso_readiness_report()`	`compliance.readiness`	Default or operator-supplied readiness report builder
`build_hipaa_documentation_packet()`	`compliance.readiness`	Default or operator-supplied HIPAA documentation packet builder
`build_forensics_report()`	`core.observability.forensics`	Tenant-safe scorer-miss report over eval records plus reviewer labels
`ForensicsReport`	`core.observability.forensics`	Miss counts, per-scorer/model/domain grouping, and per-case operator actions
`build_observability_operations_report()`	`ui.safety_dashboard`	Tenant-safe halt forensics, drift alerts, controls, and compliance-export packet
`ComplianceExportRef`	`ui.safety_dashboard`	Compliance export reference included without serialising artefact contents
`DriftDetector`	`compliance.drift_detector`	Statistical drift detection (two-proportion z-test)
`DriftResult`	`compliance.drift_detector`	Drift analysis result with z-score, p-value, severity

See Compliance Reporting Guide.

Enterprise Moderation¶

Class	Module	Purpose
`ContentModerator`	`enterprise.moderation`	One-call PII redaction plus toxicity allow/warn/block wrapper
`ContentModerationResult`	`enterprise.moderation`	Tenant-safe moderation decision and metadata
`ModerationAction`	`enterprise.moderation`	`allow`, `redact`, `warn`, or `block` decision enum
`CustomRuleset`	`enterprise.rules_dsl`	Strict JSON/YAML custom policy loader for operator-owned rules
`CustomRule`	`enterprise.rules_dsl`	Validated compiled-rule input record
`RulesDslError`	`enterprise.rules_dsl`	Stable validation error for invalid ruleset documents
`build_trust_console_report()`	`ui.safety_dashboard`	Tenant-safe Trust Console JSON/Markdown report builder
`TrustControl`	`ui.safety_dashboard`	Readiness control row for customer-facing trust reports
`build_observability_operations_report()`	`ui.safety_dashboard`	Tenant-safe operations packet for halt forensics, drift alerts, and compliance exports
`ComplianceExportRef`	`ui.safety_dashboard`	Operator-owned compliance export reference

See Enterprise API.

Import Patterns¶

# Top-level convenience imports
from director_ai import guard, score, get_score
from director_ai import CoherenceScorer, StreamingKernel, CoherenceAgent
from director_ai import InferenceHookRequest, build_inference_server_hook

# v3.10.0: Structured verification (stdlib only, no torch)
from director_ai import verify_json, verify_tool_call, verify_code

# v3.10.0: Online calibration
from director_ai import FeedbackStore, OnlineCalibrator, CalibrationReport

# v3.10.0: EU AI Act compliance
from director_ai import AuditLog, AuditEntry, ComplianceReporter, Article15Report, Article15TemplateContext

# Commercial readiness
from director_ai import build_soc2_iso_readiness_report

# v3.10.0: Meta-confidence
from director_ai import compute_meta_confidence, ContradictionTracker

# Direct module imports (for type hints and advanced use)
from director_ai.core.config import DirectorConfig
from director_ai.core.types import CoherenceScore, ReviewResult
from director_ai.core.retrieval.vector_store import VectorGroundTruthStore, ChromaBackend
from director_ai.core.runtime.batch import BatchProcessor

# Enterprise (lazy-loaded)
from director_ai.enterprise import TenantRouter, Policy, AuditLogger, CustomRuleset