Namespaces
namespace	byte_stream_split
	BYTE_STREAM_SPLIT encoding functions for float and double types.

namespace	commercial
	Commercial licensing and evaluation-tier usage enforcement.

namespace	crypto

namespace	delta

namespace	detail
	Internal implementation details for dictionary encoding.

namespace	detail_mmap

namespace	detail_mmap_reader

namespace	detail_reader

namespace	dora

namespace	eu_ai_act

namespace	gdpr

namespace	mifid2

namespace	regulatory

namespace	risk

namespace	simd
	Platform-optimized SIMD routines for common vector operations.

namespace	thrift

namespace	xxhash
	xxHash64 hashing functions for Parquet bloom filter support.

namespace	z_order
	Z-order curve (Morton code) utilities for spatial sort keys.

Classes
class	Arena
	Bump-pointer arena allocator for batch Parquet reads. More...

class	ArrowExporter
	Exports Signet Forge tensors and columns as Arrow C Data Interface structs. More...

class	ArrowImporter
	Imports Arrow C Data Interface arrays into Signet TensorView or OwnedTensor. More...

struct	Art15Metrics
	Computed accuracy, robustness, and drift metrics per EU AI Act Art.15. More...

class	Art15MetricsCalculator
	Computes Art.15 accuracy, robustness, and drift metrics from inference records. More...

class	AuditChainVerifier
	Verifies hash chain integrity. More...

class	AuditChainWriter
	Builds SHA-256 hash chains during Parquet writes. More...

struct	AuditMetadata
	Chain summary stored in Parquet key-value metadata. More...

class	BatchTensorBuilder
	Builds a single contiguous 2D tensor from multiple column tensors, suitable for passing to an ML inference engine (ONNX Runtime, etc.). More...

struct	BufferInfo
	Simple C-contiguous buffer descriptor for Python interop. More...

struct	ClockSyncStatus
	NTP/PTP clock synchronization status for MiFID II RTS 25 Art.3. More...

class	CodecRegistry
	Thread-safe singleton registry of compression codec implementations. More...

struct	Column
	Typed column descriptor for the Schema::build() variadic API. More...

class	ColumnBatch
	A column-major batch of feature rows for ML inference and WAL serialization. More...

struct	ColumnDesc
	Describes a single column in a ColumnBatch schema. More...

struct	ColumnDescriptor
	Descriptor for a single column in a Parquet schema. More...

struct	ColumnFileStats
	Per-column statistics from ParquetReader::file_stats(). More...

struct	ColumnIndex
	Per-page min/max statistics for predicate pushdown. More...

class	ColumnIndexBuilder
	Builder that accumulates per-page statistics during column writing. More...

class	ColumnReader
	PLAIN-encoded Parquet column decoder. More...

class	ColumnStatistics
	Per-column-chunk statistics tracker. More...

class	ColumnToTensor
	Provides static methods to convert Parquet column data into tensor form. More...

class	ColumnWriter
	PLAIN encoding writer for a single Parquet column. More...

struct	ColumnWriteStats
	Per-column statistics produced by ParquetWriter::close(). More...

struct	ComplianceReport
	The generated compliance report returned to the caller. More...

class	CompressionCodec
	Abstract base class for all compression/decompression codecs. More...

class	DataClassificationOntology
	A named collection of data classification rules forming a formal ontology. More...

struct	DataClassificationRule
	Per-field data classification and handling policy. More...

class	DecisionLogReader
	Reads AI decision log Parquet files and verifies hash chain integrity. More...

class	DecisionLogWriter
	Writes AI trading decision records to Parquet files with cryptographic hash chaining for tamper-evident audit trails. More...

struct	DecisionRecord
	A single AI-driven trading decision with full provenance. More...

class	Dequantizer
	Dequantizes INT8/INT4 quantized vectors back to float32. More...

class	DictionaryDecoder
	Dictionary decoder for Parquet PLAIN_DICTIONARY / RLE_DICTIONARY encoding. More...

class	DictionaryEncoder
	Dictionary encoder for Parquet PLAIN_DICTIONARY / RLE_DICTIONARY encoding. More...

struct	DLDataType
	DLPack data type descriptor. More...

struct	DLDevice
	DLPack device descriptor (type + ordinal). More...

struct	DLManagedTensor
	DLPack managed tensor – the exchange object for from_dlpack(). More...

struct	DLTensor
	DLPack tensor descriptor (non-owning). More...

struct	DreadScore
	DREAD risk quantification — 5 factors scored 1..10. More...

struct	Error
	Lightweight error value carrying an ErrorCode and a human-readable message. More...

class	EUAIActReporter
	EU AI Act compliance report generator (Regulation (EU) 2024/1689). More...

class	EventBus
	Multi-tier event bus for routing SharedColumnBatch events. More...

struct	EventBusOptions
	Configuration options for EventBus. More...

class	expected
	A lightweight result type that holds either a success value of type `T` or an Error. More...

class	expected< void >
	Specialization of expected for void — used for operations that return success or error only. More...

struct	FeatureGroupDef
	Schema definition for a single feature group. More...

class	FeatureReader
	Point-in-time correct ML feature store reader over Parquet files. More...

struct	FeatureReaderOptions
	Configuration options for FeatureReader::open(). More...

struct	FeatureVector
	A single versioned observation for one entity. More...

class	FeatureWriter
	Append-only writer for a single feature group. More...

struct	FeatureWriterOptions
	Configuration options for FeatureWriter::create(). More...

struct	FileStats
	Aggregate file-level statistics returned by ParquetReader::file_stats(). More...

struct	HashChainEntry
	A single link in the cryptographic hash chain. More...

class	HumanOverrideLogReader
	Reads human override log Parquet files and verifies hash chain integrity. More...

class	HumanOverrideLogWriter
	Writes human override events to Parquet files with cryptographic hash chaining for tamper-evident audit trails. More...

struct	HumanOverrideRecord
	A single human oversight event with full provenance. More...

struct	HybridQueryOptions
	Per-query filter options passed to HybridReader::read(). More...

class	HybridReader
	Reads StreamRecords across historical Parquet files and (optionally) a live StreamingSink ring buffer snapshot. More...

struct	HybridReaderOptions
	Options for constructing a HybridReader via HybridReader::create(). More...

struct	IncidentPlaybook
	An ordered sequence of response steps for a specific incident type. More...

class	IncidentResponseTracker
	Tracks execution progress of a playbook during an active incident. More...

class	InferenceLogReader
	Reads ML inference log Parquet files and verifies hash chain integrity. More...

class	InferenceLogWriter
	Writes ML inference records to Parquet files with cryptographic hash chaining for tamper-evident audit trails. More...

struct	InferenceRecord
	A single ML inference event with full operational metadata. More...

class	LogRetentionManager
	Manages log file lifecycle: retention, archival, and deletion. More...

class	MappedSegment
	RAII cross-platform memory-mapped file segment. More...

class	MiFID2Reporter
	MiFID II RTS 24 algorithmic trading compliance report generator. More...

struct	Mitigation
	A specific mitigation control for a threat. More...

class	MmapParquetReader

class	MmapReader
	Low-level memory-mapped file handle. More...

class	MpmcRing
	Lock-free bounded multi-producer multi-consumer ring buffer. More...

class	MpscRingBuffer
	Multiple-producer single-consumer (MPSC) bounded ring buffer. More...

struct	native_type_of
	Maps a Parquet PhysicalType back to its corresponding C++ native type. More...

struct	native_type_of< PhysicalType::BOOLEAN >

struct	native_type_of< PhysicalType::BYTE_ARRAY >

struct	native_type_of< PhysicalType::DOUBLE >

struct	native_type_of< PhysicalType::FLOAT >

struct	native_type_of< PhysicalType::INT32 >

struct	native_type_of< PhysicalType::INT64 >

class	NumpyBridge
	Exports and imports Signet tensors via DLPack, enabling zero-copy interoperability with PyTorch, NumPy, JAX, and other DLPack-aware frameworks. More...

struct	OffsetIndex
	Page locations for random access within a column chunk. More...

struct	OnnxInputSet
	A set of named ONNX tensors for multi-input model inference. More...

struct	OnnxTensorInfo
	Contains all information needed to create an OrtValue externally. More...

class	OverrideRateMonitor
	Sliding-window override rate monitor — EU AI Act Art.14(5). More...

struct	OverrideRateMonitorOptions
	Options for the override rate monitor. More...

class	OwnedTensor
	An owning tensor that manages its own memory via a std::vector<uint8_t> buffer. More...

struct	PageLocation
	File offset and size descriptor for a single data page. More...

struct	parquet_type_of
	Maps a C++ type to its corresponding Parquet PhysicalType at compile time. More...

struct	parquet_type_of< bool >

struct	parquet_type_of< double >

struct	parquet_type_of< float >

struct	parquet_type_of< int32_t >

struct	parquet_type_of< int64_t >

struct	parquet_type_of< std::string >

class	ParquetReader
	Parquet file reader with typed column access and full encoding support. More...

class	ParquetWriter
	Streaming Parquet file writer with row-based and column-based APIs. More...

class	PlaybookRegistry
	Registry of incident response playbooks indexed by incident type. More...

struct	PlaybookStep
	A single step in an incident response playbook. More...

struct	QuantizationParams
	Parameters that fully describe a quantization mapping. More...

class	QuantizedVectorReader
	Reads quantized page data (FIXED_LEN_BYTE_ARRAY) and dequantizes to float32 on demand. More...

class	QuantizedVectorWriter
	Accumulates float32 vectors, quantizes them, and produces FIXED_LEN_BYTE_ARRAY page data suitable for Parquet column chunks. More...

class	Quantizer
	Quantizes float32 vectors to INT8 or INT4 representation. More...

struct	RegulatoryChange
	A tracked regulatory change record. More...

class	RegulatoryChangeMonitor
	Registry and tracker for regulatory changes affecting the system. More...

struct	ReportOptions
	Query and formatting parameters for compliance report generation. More...

struct	RetentionPolicy
	Retention policy configuration for log lifecycle management. More...

struct	RetentionSummary
	Summary of a retention enforcement pass. More...

class	RleDecoder
	Streaming decoder for the Parquet RLE/Bit-Packing Hybrid scheme. More...

class	RleEncoder
	Streaming encoder for the Parquet RLE/Bit-Packing Hybrid scheme. More...

class	RowLineageTracker
	Per-row lineage tracking inspired by Iceberg V3-style data governance. More...

class	Schema
	Immutable schema description for a Parquet file. More...

class	SchemaBuilder
	Fluent builder for constructing a Schema one column at a time. More...

class	SnappyCodec
	Bundled Snappy compression codec (header-only, no external dependency). More...

class	SplitBlockBloomFilter
	Parquet-spec Split Block Bloom Filter for probabilistic set membership. More...

class	SpscRingBuffer
	Lock-free single-producer single-consumer (SPSC) bounded ring buffer. More...

class	StreamingSink
	Background-thread Parquet compaction sink fed by a lock-free ring buffer. More...

struct	StreamRecord
	A single record flowing through the StreamingSink pipeline. More...

struct	TensorShape
	Describes the shape of a tensor as a vector of dimension sizes. More...

class	TensorView
	A lightweight, non-owning view into a contiguous block of typed memory, interpreted as a multi-dimensional tensor. More...

struct	ThreatEntry
	A single identified threat in the threat model. More...

struct	ThreatModel
	A threat model for a specific component or the entire system. More...

struct	ThreatModelAnalysis
	Analysis result from validating a threat model. More...

class	ThreatModelAnalyzer
	Validates threat model coverage and produces audit-ready JSON. More...

struct	VectorColumnSpec
	Configuration for a vector column: dimensionality and element precision. More...

class	VectorReader
	Reads FIXED_LEN_BYTE_ARRAY page data back into float vectors. More...

class	VectorWriter
	Buffers float vectors and encodes them as FIXED_LEN_BYTE_ARRAY PLAIN data. More...

struct	WalEntry
	A single decoded WAL record returned by WalReader::next() or read_all(). More...

class	WalManager
	Manages multiple rolling WAL segment files in a directory. More...

struct	WalManagerOptions
	Configuration options for WalManager::open(). More...

struct	WalMmapOptions
	Configuration options for WalMmapWriter::open(). More...

class	WalMmapWriter
	High-performance WAL writer using a ring of N memory-mapped segments. More...

class	WalReader
	Sequential WAL file reader for crash recovery and replay. More...

class	WalWriter
	Append-only Write-Ahead Log writer with CRC-32 integrity per record. More...

struct	WalWriterOptions
	Configuration options for WalWriter::open(). More...

struct	WriterOptions
	Configuration options for ParquetWriter. More...

struct	WriteStats
	File-level write statistics returned by ParquetWriter::close(). More...

Typedefs
using	TDT = TensorDataType
	Convenience alias for TensorDataType (shorter schema declarations).

using	SharedColumnBatch = std::shared_ptr< ColumnBatch >
	Thread-safe shared pointer to a ColumnBatch – the unit transferred between producer and consumer threads via EventBus.

using	WalRecord = WalEntry
	Alias so callers can use either WalEntry or WalRecord.

template<PhysicalType PT>
using	native_type_of_t = typename native_type_of< PT >::type
	Convenience alias: `native_type_of_t<PhysicalType::INT64>` == `int64_t`.

Enumerations
enum class	ReportFormat { JSON , NDJSON , CSV }
	Output serialization format for compliance reports. More...

enum class	TimestampGranularity { NANOS , MICROS , MILLIS }
	Timestamp granularity for MiFID II RTS 24 Art.2(2) compliance. More...

enum class	ComplianceStandard { MIFID2_RTS24 , EU_AI_ACT_ART12 , EU_AI_ACT_ART13 , EU_AI_ACT_ART19 }
	Which regulatory standard a compliance report satisfies. More...

enum class	DataClassification : int32_t { PUBLIC = 0 , INTERNAL = 1 , RESTRICTED = 2 , HIGHLY_RESTRICTED = 3 }
	Data confidentiality level per DORA Art.8 + ISO 27001 Annex A. More...

enum class	DataSensitivity : int32_t { NEUTRAL = 0 , PSEUDONYMISED = 1 , ANONYMISED = 2 , PII = 3 , FINANCIAL_PII = 4 , BIOMETRIC = 5 , HEALTH = 6 }
	Data sensitivity per GDPR Art.9 special categories. More...

enum class	RegulatoryRegime : int32_t { NONE = 0 , GDPR = 1 , MIFID2 = 2 , DORA = 3 , EU_AI_ACT = 4 , SOX = 5 , SEC_17A4 = 6 , PCI_DSS = 7 , HIPAA = 8 }
	Regulatory regime(s) applicable to the data. More...

enum class	DecisionType : int32_t { SIGNAL = 0 , ORDER_NEW = 1 , ORDER_CANCEL = 2 , ORDER_MODIFY = 3 , POSITION_OPEN = 4 , POSITION_CLOSE = 5 , RISK_OVERRIDE = 6 , NO_ACTION = 7 }
	Classification of the AI-driven trading decision. More...

enum class	RiskGateResult : int32_t { PASSED = 0 , REJECTED = 1 , MODIFIED = 2 , THROTTLED = 3 }
	Outcome of the pre-trade risk gate evaluation. More...

enum class	OrderType : int32_t { MARKET = 0 , LIMIT = 1 , STOP = 2 , STOP_LIMIT = 3 , PEGGED = 4 , OTHER = 99 }
	Order type classification for MiFID II RTS 24 Annex I Table 2 Field 7. More...

enum class	TimeInForce : int32_t { DAY = 0 , GTC = 1 , IOC = 2 , FOK = 3 , GTD = 4 , OTHER = 99 }
	Time-in-force classification for MiFID II RTS 24 Annex I Table 2 Field 8. More...

enum class	BuySellIndicator : int32_t { BUY = 0 , SELL = 1 , SHORT_SELL = 2 }
	Buy/sell direction for MiFID II RTS 24 Annex I Table 2 Field 6. More...

enum class	OverrideSource : int32_t { ALGORITHMIC = 0 , HUMAN = 1 , AUTOMATED = 2 }
	Source of a decision or override — EU AI Act Art.14(4). More...

enum class	OverrideAction : int32_t { APPROVE = 0 , MODIFY = 1 , REJECT = 2 , ESCALATE = 3 , HALT = 4 }
	What action the human override took — EU AI Act Art.14(4). More...

enum class	HaltReason : int32_t { MANUAL = 0 , SAFETY_THRESHOLD = 1 , ANOMALY_DETECTED = 2 , REGULATORY = 3 , MAINTENANCE = 4 , EXTERNAL = 5 }
	Reason for system halt — EU AI Act Art.14(4) "stop button". More...

enum class	IncidentPhase : int32_t { PREPARATION = 0 , DETECTION = 1 , CONTAINMENT = 2 , ERADICATION = 3 , RECOVERY = 4 , LESSONS_LEARNED = 5 }
	NIST SP 800-61 incident response lifecycle phases. More...

enum class	IncidentSeverity : int32_t { P4_LOW = 0 , P3_MEDIUM = 1 , P2_HIGH = 2 , P1_CRITICAL = 3 }
	Incident severity per DORA Art.10(1) classification. More...

enum class	EscalationLevel : int32_t { L1_OPERATIONS = 0 , L2_ENGINEERING = 1 , L3_MANAGEMENT = 2 , L4_REGULATORY = 3 }
	Escalation hierarchy for incident routing. More...

enum class	NotificationChannel : int32_t { INTERNAL_LOG = 0 , EMAIL = 1 , PAGER = 2 , REGULATORY = 3 }
	Notification channel for incident communications. More...

enum class	InferenceType : int32_t { CLASSIFICATION = 0 , REGRESSION = 1 , EMBEDDING = 2 , GENERATION = 3 , RANKING = 4 , ANOMALY = 5 , RECOMMENDATION = 6 , CUSTOM = 255 }
	Classification of the ML inference operation. More...

enum class	QuantizationScheme : int32_t { SYMMETRIC_INT8 = 0 , ASYMMETRIC_INT8 = 1 , SYMMETRIC_INT4 = 2 }
	Identifies the quantization method used for vector compression. More...

enum class	RegulatoryChangeType : int32_t { NEW_REGULATION = 0 , AMENDMENT = 1 , GUIDANCE = 2 , TECHNICAL_STANDARD = 3 , ENFORCEMENT = 4 , DEPRECATION = 5 }
	Type of regulatory change being tracked. More...

enum class	RegulatoryImpact : int32_t { NONE = 0 , INFORMATIONAL = 1 , LOW = 2 , MEDIUM = 3 , HIGH = 4 , CRITICAL = 5 }
	Impact level of a regulatory change on the system. More...

enum class	ChangeComplianceStatus : int32_t { NOT_ASSESSED = 0 , ASSESSED = 1 , IN_PROGRESS = 2 , IMPLEMENTED = 3 , VERIFIED = 4 , NOT_APPLICABLE = 5 }
	Compliance status for a tracked regulatory change. More...

enum class	TensorDataType : int32_t { FLOAT32 = 0 , FLOAT64 = 1 , INT32 = 2 , INT64 = 3 , INT8 = 4 , UINT8 = 5 , INT16 = 6 , FLOAT16 = 7 , BOOL = 8 }
	Element data type for tensor storage, mapping to ONNX/PyTorch/TF type enums. More...

enum class	StrideCategory : int32_t { SPOOFING = 0 , TAMPERING = 1 , REPUDIATION = 2 , INFORMATION_DISCLOSURE = 3 , DENIAL_OF_SERVICE = 4 , ELEVATION_OF_PRIVILEGE = 5 }
	Microsoft STRIDE threat categories. More...

enum class	ThreatSeverity : int32_t { LOW = 0 , MEDIUM = 1 , HIGH = 2 , CRITICAL = 3 }
	Threat severity classification per NIST SP 800-30. More...

enum class	MitigationStatus : int32_t { NOT_MITIGATED = 0 , PARTIAL = 1 , MITIGATED = 2 , ACCEPTED = 3 , TRANSFERRED = 4 }
	Mitigation status for a threat. More...

enum class	VectorElementType : int32_t { FLOAT32 = 0 , FLOAT64 = 1 , FLOAT16 = 2 }
	Specifies the numerical precision of each element within a vector column. More...

enum class	WalLifecycleMode : uint8_t { Development = 0 , Benchmark = 1 , Production = 2 }
	Controls safety guardrails for WAL segment lifecycle operations. More...

enum class	ErrorCode { OK = 0 , IO_ERROR , INVALID_FILE , CORRUPT_FOOTER , CORRUPT_PAGE , CORRUPT_DATA , INVALID_ARGUMENT , UNSUPPORTED_ENCODING , UNSUPPORTED_COMPRESSION , UNSUPPORTED_TYPE , SCHEMA_MISMATCH , OUT_OF_RANGE , THRIFT_DECODE_ERROR , ENCRYPTION_ERROR , HASH_CHAIN_BROKEN , LICENSE_ERROR , LICENSE_LIMIT_EXCEEDED , INTERNAL_ERROR }
	Error codes returned by all Signet Forge operations. More...

enum class	OnnxTensorType : int32_t { UNDEFINED = 0 , FLOAT = 1 , UINT8 = 2 , INT8 = 3 , UINT16 = 4 , INT16 = 5 , INT32 = 6 , INT64 = 7 , STRING = 8 , BOOL = 9 , FLOAT16 = 10 , DOUBLE = 11 , UINT32 = 12 , UINT64 = 13 , BFLOAT16 = 16 }
	ONNX tensor element data types, mirroring OrtTensorElementDataType. More...

enum class	PhysicalType : int32_t { BOOLEAN = 0 , INT32 = 1 , INT64 = 2 , INT96 = 3 , FLOAT = 4 , DOUBLE = 5 , BYTE_ARRAY = 6 , FIXED_LEN_BYTE_ARRAY = 7 }
	Parquet physical (storage) types as defined in parquet.thrift. More...

enum class	LogicalType : int32_t { NONE = 0 , STRING = 1 , ENUM = 2 , UUID = 3 , DATE = 4 , TIME_MS = 5 , TIME_US = 6 , TIME_NS = 7 , TIMESTAMP_MS = 8 , TIMESTAMP_US = 9 , TIMESTAMP_NS = 10 , DECIMAL = 11 , JSON = 12 , BSON = 13 , FLOAT16 = 14 , FLOAT32_VECTOR = 100 }
	Parquet logical types (from parquet.thrift LogicalType union). More...

enum class	ConvertedType : int32_t { NONE = -1 , UTF8 = 0 , MAP = 1 , MAP_KEY_VALUE = 2 , LIST = 3 , ENUM = 4 , DECIMAL = 5 , DATE = 6 , TIME_MILLIS = 7 , TIME_MICROS = 8 , TIMESTAMP_MILLIS = 9 , TIMESTAMP_MICROS = 10 , UINT_8 = 11 , UINT_16 = 12 , UINT_32 = 13 , UINT_64 = 14 , INT_8 = 15 , INT_16 = 16 , INT_32 = 17 , INT_64 = 18 , JSON = 19 , BSON = 20 , INTERVAL = 21 }
	Legacy Parquet converted types for backward compatibility with older readers. More...

enum class	Encoding : int32_t { PLAIN = 0 , PLAIN_DICTIONARY = 2 , RLE = 3 , BIT_PACKED = 4 , DELTA_BINARY_PACKED = 5 , DELTA_LENGTH_BYTE_ARRAY = 6 , DELTA_BYTE_ARRAY = 7 , RLE_DICTIONARY = 8 , BYTE_STREAM_SPLIT = 9 }
	Parquet page encoding types. More...

enum class	Compression : int32_t { UNCOMPRESSED = 0 , SNAPPY = 1 , GZIP = 2 , LZO = 3 , BROTLI = 4 , LZ4 = 5 , ZSTD = 6 , LZ4_RAW = 7 }
	Parquet compression codecs. More...

enum class	PageType : int32_t { DATA_PAGE = 0 , INDEX_PAGE = 1 , DICTIONARY_PAGE = 2 , DATA_PAGE_V2 = 3 }
	Parquet page types within a column chunk. More...

enum class	Repetition : int32_t { REQUIRED = 0 , OPTIONAL = 1 , REPEATED = 2 }
	Parquet field repetition types (nullability / cardinality). More...

DLPack type definitions (matching dlpack.h v0.8)
Self-contained DLPack struct definitions for zero-dependency interop.
enum class	DLDeviceType : int32_t { kDLCPU = 1 , kDLCUDA = 2 , kDLCUDAHost = 3 , kDLROCM = 10 , kDLMetal = 8 , kDLVulkan = 7 }
	DLPack device type, matching DLDeviceType from dlpack.h. More...

enum class	DLDataTypeCode : uint8_t { kDLInt = 0 , kDLUInt = 1 , kDLFloat = 2 , kDLBfloat = 4 }
	DLPack data type code, matching DLDataTypeCode from dlpack.h. More...

Functions
int64_t	now_ns ()
	Return the current time as nanoseconds since the Unix epoch (UTC).

std::string	hash_to_hex (const std::array< uint8_t, 32 > &hash)
	Convert a 32-byte SHA-256 hash to a lowercase hexadecimal string (64 chars).

expected< std::array< uint8_t, 32 > >	hex_to_hash (const std::string &hex)
	Convert a 64-character lowercase hex string back to a 32-byte hash.

std::string	generate_chain_id ()
	Generate a simple chain identifier based on the current timestamp.

expected< AuditMetadata >	build_audit_metadata (const AuditChainWriter &writer, const std::string &chain_id)
	Build an AuditMetadata from a populated AuditChainWriter.

expected< std::vector< HashChainEntry > >	deserialize_and_verify_chain (const uint8_t *chain_data, size_t chain_size)
	Deserialize and verify a chain from serialized bytes in one call.

SharedColumnBatch	make_column_batch (std::vector< ColumnDesc > schema, size_t reserve_rows=64)
	Convenience factory: create a shared batch with a given schema.

Schema	decision_log_schema ()
	Build the Parquet schema for AI decision log files.

Schema	human_override_log_schema ()
	Build the Parquet schema for human override log files.

Schema	inference_log_schema ()
	Build the Parquet schema for ML inference log files.

constexpr size_t	tensor_element_size (TensorDataType dtype) noexcept
	Returns the byte size of a single element of the given tensor data type.

const char *	tensor_dtype_name (TensorDataType dtype) noexcept
	Returns a human-readable name for a TensorDataType.

float	f16_to_f32 (uint16_t h) noexcept
	Convert a 16-bit IEEE 754 half-precision value to a 32-bit float.

uint16_t	f32_to_f16 (float val) noexcept
	Convert a 32-bit float to a 16-bit IEEE 754 half-precision value.

SchemaBuilder &	add_vector_column (SchemaBuilder &builder, const std::string &name, uint32_t dimension, VectorElementType elem=VectorElementType::FLOAT32)
	Add a vector column to a SchemaBuilder.

void	append_le32 (std::vector< uint8_t > &buf, uint32_t val)
	Append a uint32_t in little-endian byte order to a byte buffer.

void	append_le64 (std::vector< uint8_t > &buf, uint64_t val)
	Append a uint64_t in little-endian byte order to a byte buffer.

expected< std::vector< uint8_t > >	compress (Compression codec, const uint8_t *data, size_t size)
	Compress data using the specified codec via the global CodecRegistry.

expected< std::vector< uint8_t > >	decompress (Compression codec, const uint8_t *data, size_t size, size_t uncompressed_size)
	Decompress data using the specified codec via the global CodecRegistry.

Compression	auto_select_compression (const uint8_t *sample_data, size_t sample_size)
	Automatically select the best available compression codec.

void	register_snappy_codec ()
	Register the bundled Snappy codec with the global CodecRegistry.

size_t	encode_varint (std::vector< uint8_t > &buf, uint64_t value)
	Encode an unsigned varint (LEB128) into a byte buffer.

uint64_t	decode_varint (const uint8_t *data, size_t &pos, size_t size)
	Decode an unsigned varint (LEB128) from a byte buffer.

void	bit_pack_8 (std::vector< uint8_t > &out, const uint64_t *values, int bit_width)
	Pack exactly 8 values at the given bit width into a byte buffer.

void	bit_unpack_8 (const uint8_t src, uint64_t values, int bit_width)
	Unpack exactly 8 values at the given bit width from a byte buffer.

expected< BufferInfo >	to_buffer_info (const TensorView &tensor)
	Create a BufferInfo from a TensorView for Python buffer protocol export.

expected< OnnxInputSet >	prepare_inputs_for_onnx (const std::vector< std::pair< std::string, TensorView > > &inputs)
	Prepare a batch of named TensorViews for ONNX Runtime inference.

const char *	onnx_type_name (OnnxTensorType t)
	Return a human-readable string for an OnnxTensorType value.

expected< size_t >	validate_mmap_page_value_count (int64_t num_values, const char *context)

expected< size_t >	validate_page_value_count (int64_t num_values, const char *context)

bool	has_encrypted_page_header_prefix (const uint8_t *data, size_t size) noexcept

uint32_t	load_le32 (const uint8_t *data) noexcept

template<typename T >
std::vector< uint8_t >	to_le_bytes (T value)
	Convert an arithmetic value to its little-endian byte representation.

std::vector< uint8_t >	to_le_bytes (const std::string &value)
	Overload for std::string – returns raw bytes (no endian conversion needed).

template<typename T >
T	from_le_bytes (const std::vector< uint8_t > &bytes)
	Reconstruct an arithmetic value from its little-endian byte representation.

Format string mappings
Conversion functions between Parquet/Tensor types and Arrow format strings.
const char *	parquet_to_arrow_format (PhysicalType pt)
	Map a Parquet PhysicalType to an Arrow format string.

const char *	tensor_dtype_to_arrow_format (TensorDataType dtype)
	Map a TensorDataType to an Arrow format string.

expected< TensorDataType >	arrow_format_to_tensor_dtype (const char *format)
	Map an Arrow format string to a TensorDataType.

expected< TensorDataType >	physical_to_tensor_dtype (PhysicalType pt)
	Map a PhysicalType to a TensorDataType (for column export).

size_t	physical_type_byte_size (PhysicalType pt)
	Return the byte size for a PhysicalType (primitive types only).

Type conversion: TensorDataType <-> DLDataType
DLDataType	to_dlpack_dtype (TensorDataType dtype)
	Convert a Signet TensorDataType to a DLPack DLDataType.

expected< TensorDataType >	from_dlpack_dtype (DLDataType dl_dtype)
	Convert a DLPack DLDataType back to a Signet TensorDataType.

Type conversion: TensorDataType <-> OnnxTensorType
OnnxTensorType	to_onnx_type (TensorDataType dtype)
	Convert a Signet TensorDataType to the corresponding OnnxTensorType.

expected< TensorDataType >	from_onnx_type (OnnxTensorType ort_type)
	Convert an OnnxTensorType back to a Signet TensorDataType.

Zero-copy tensor export for ONNX Runtime
expected< OnnxTensorInfo >	prepare_for_onnx (const TensorView &tensor)
	Prepare a TensorView for ONNX Runtime consumption (zero-copy).

expected< OnnxTensorInfo >	prepare_for_onnx (const OwnedTensor &tensor)
	Prepare an OwnedTensor for ONNX Runtime consumption (zero-copy).

Variables
constexpr size_t	HASH_CHAIN_ENTRY_SIZE = 112
	Chain summary stored in Parquet key-value metadata.

constexpr uint8_t	kEncryptedPageHeaderMagic [4] = {'S', 'P', 'H', '1'}

template<typename T >
constexpr PhysicalType	parquet_type_of_v = parquet_type_of<T>::value
	Convenience variable template: `parquet_type_of_v<double>` == `PhysicalType::DOUBLE`.

constexpr int32_t	PARQUET_VERSION = 2
	Parquet format version written to the file footer.

constexpr const char *	SIGNET_CREATED_BY = "SignetStack signet-forge version 0.1.0"
	Default "created_by" string embedded in every Parquet footer.

constexpr uint32_t	PARQUET_MAGIC = 0x31524150
	"PAR1" magic bytes (little-endian uint32) — marks a standard Parquet file.

constexpr uint32_t	PARQUET_MAGIC_ENCRYPTED = 0x45524150
	"PARE" magic bytes (little-endian uint32) — marks a Parquet file with an encrypted footer.

Typedef Documentation

◆ native_type_of_t

template<PhysicalType PT>

using signet::forge::native_type_of_t = typedef typename native_type_of<PT>::type

Convenience alias: native_type_of_t<PhysicalType::INT64> == int64_t.

Definition at line 198 of file types.hpp.

◆ SharedColumnBatch

using signet::forge::SharedColumnBatch = typedef std::shared_ptr<ColumnBatch>

Thread-safe shared pointer to a ColumnBatch – the unit transferred between producer and consumer threads via EventBus.

Definition at line 400 of file column_batch.hpp.

◆ TDT

using signet::forge::TDT = typedef TensorDataType

Convenience alias for TensorDataType (shorter schema declarations).

Definition at line 46 of file column_batch.hpp.

◆ WalRecord

using signet::forge::WalRecord = typedef WalEntry

Alias so callers can use either WalEntry or WalRecord.

Definition at line 207 of file wal.hpp.

Enumeration Type Documentation

◆ BuySellIndicator

enum class signet::forge::BuySellIndicator : int32_t

strong

Buy/sell direction for MiFID II RTS 24 Annex I Table 2 Field 6.

Enumerator
BUY
SELL
SHORT_SELL	Short selling (RTS 24 Annex I Field 16)

Definition at line 90 of file decision_log.hpp.

◆ ChangeComplianceStatus

enum class signet::forge::ChangeComplianceStatus : int32_t

strong

Compliance status for a tracked regulatory change.

Enumerator
NOT_ASSESSED	Impact assessment not yet performed.
ASSESSED	Impact assessed, action plan pending.
IN_PROGRESS	Implementation underway.
IMPLEMENTED	Changes implemented.
VERIFIED	Compliance verified by review/testing.
NOT_APPLICABLE	Change does not apply to this system.

Definition at line 65 of file regulatory_monitor.hpp.

◆ ComplianceStandard

enum class signet::forge::ComplianceStandard

strong

Which regulatory standard a compliance report satisfies.

Enumerator
MIFID2_RTS24	MiFID II RTS 24 — algorithmic trading records.
EU_AI_ACT_ART12	EU AI Act Article 12 — operational logging.
EU_AI_ACT_ART13	EU AI Act Article 13 — transparency disclosure.
EU_AI_ACT_ART19	EU AI Act Article 19 — conformity assessment summary.

Definition at line 43 of file compliance_types.hpp.

◆ Compression

enum class signet::forge::Compression : int32_t

strong

Parquet compression codecs.

Snappy is bundled (header-only); ZSTD, LZ4, and Gzip require linking external libraries enabled via CMake options.

See also: WriterOptions::compression, WriterOptions::auto_compression

Enumerator
UNCOMPRESSED	No compression.
SNAPPY	Snappy compression (bundled, header-only).
GZIP	Gzip/deflate compression (requires SIGNET_ENABLE_GZIP).
LZO	LZO compression (not currently supported).
BROTLI	Brotli compression (not currently supported).
LZ4	LZ4 block compression (requires SIGNET_ENABLE_LZ4).
ZSTD	Zstandard compression (requires SIGNET_ENABLE_ZSTD).
LZ4_RAW	LZ4 raw (unframed) block compression.

Definition at line 115 of file types.hpp.

◆ ConvertedType

enum class signet::forge::ConvertedType : int32_t

strong

Legacy Parquet converted types for backward compatibility with older readers.

Prefer LogicalType for new code. ConvertedType is written to the Thrift footer only when a corresponding LogicalType mapping exists (e.g. STRING → UTF8).

See also: LogicalType

Enumerator
NONE	No converted type annotation.
UTF8	UTF-8 encoded string.
MAP	Map (nested group).
MAP_KEY_VALUE	Map key-value pair.
LIST	List (nested group).
ENUM	Enum string.
DECIMAL	Fixed-point decimal.
DATE	Date (days since epoch).
TIME_MILLIS	Time in milliseconds.
TIME_MICROS	Time in microseconds.
TIMESTAMP_MILLIS	Timestamp in milliseconds.
TIMESTAMP_MICROS	Timestamp in microseconds.
UINT_8	Unsigned 8-bit integer.
UINT_16	Unsigned 16-bit integer.
UINT_32	Unsigned 32-bit integer.
UINT_64	Unsigned 64-bit integer.
INT_8	Signed 8-bit integer.
INT_16	Signed 16-bit integer.
INT_32	Signed 32-bit integer.
INT_64	Signed 64-bit integer.
JSON	JSON document.
BSON	BSON document.
INTERVAL	Time interval.

Definition at line 67 of file types.hpp.

◆ DataClassification

enum class signet::forge::DataClassification : int32_t

strong

Data confidentiality level per DORA Art.8 + ISO 27001 Annex A.

Enumerator
PUBLIC	No confidentiality requirement.
INTERNAL	Business-internal, not for external sharing.
RESTRICTED	Regulated data (GDPR, FCA, MiFID II)
HIGHLY_RESTRICTED	Cryptographic keys, trading secrets, PII.

Definition at line 46 of file data_classification.hpp.

◆ DataSensitivity

enum class signet::forge::DataSensitivity : int32_t

strong

Data sensitivity per GDPR Art.9 special categories.

Enumerator
NEUTRAL	No special sensitivity.
PSEUDONYMISED	Identifiable only with additional key (Art.25)
ANONYMISED	Irreversibly de-identified (Art.4(1))
PII	Personally Identifiable Information.
FINANCIAL_PII	Financial account data, trading activity.
BIOMETRIC	Biometric data (Art.9 special category)
HEALTH	Health/genetic data (Art.9 special category)

Definition at line 54 of file data_classification.hpp.

◆ DecisionType

enum class signet::forge::DecisionType : int32_t

strong

Classification of the AI-driven trading decision.

Covers the full lifecycle of order management decisions that must be logged under MiFID II RTS 24 and EU AI Act Article 12.

Enumerator
SIGNAL	Raw model signal/prediction.
ORDER_NEW	Decision to submit a new order.
ORDER_CANCEL	Decision to cancel an existing order.
ORDER_MODIFY	Decision to modify an existing order.
POSITION_OPEN	Decision to open a position.
POSITION_CLOSE	Decision to close a position.
RISK_OVERRIDE	Risk gate override/rejection.
NO_ACTION	Model evaluated but no action taken.

Definition at line 48 of file decision_log.hpp.

◆ DLDataTypeCode

enum class signet::forge::DLDataTypeCode : uint8_t

strong

DLPack data type code, matching DLDataTypeCode from dlpack.h.

Enumerator
kDLInt	Signed integer.
kDLUInt	Unsigned integer.
kDLFloat	IEEE floating point.
kDLBfloat	Brain floating point (bfloat16)

Definition at line 50 of file numpy_bridge.hpp.

◆ DLDeviceType

enum class signet::forge::DLDeviceType : int32_t

strong

DLPack device type, matching DLDeviceType from dlpack.h.

Only kDLCPU and kDLCUDAHost are supported for import by NumpyBridge. Other device types are defined for completeness and forward compatibility.

Enumerator
kDLCPU	System main memory.
kDLCUDA	NVIDIA CUDA GPU memory.
kDLCUDAHost	CUDA pinned host memory.
kDLROCM	AMD ROCm GPU memory.
kDLMetal	Apple Metal GPU memory.
kDLVulkan	Vulkan GPU memory.

Definition at line 40 of file numpy_bridge.hpp.

◆ Encoding

enum class signet::forge::Encoding : int32_t

strong

Parquet page encoding types.

Each data page stores values using one of these encodings. The writer selects encoding per-column (or auto-selects based on data characteristics).

See also: WriterOptions::default_encoding, WriterOptions::auto_encoding

Enumerator
PLAIN	Values stored back-to-back in their native binary layout.
PLAIN_DICTIONARY	Legacy dictionary encoding (Parquet 1.0).
RLE	Run-length / bit-packed hybrid (used for booleans and def/rep levels).
BIT_PACKED	Deprecated — superseded by RLE.
DELTA_BINARY_PACKED	Delta encoding for INT32/INT64 (compact for sorted/sequential data).
DELTA_LENGTH_BYTE_ARRAY	Delta-encoded lengths + concatenated byte arrays.
DELTA_BYTE_ARRAY	Incremental/prefix encoding for byte arrays.
RLE_DICTIONARY	Modern dictionary encoding (Parquet 2.0) — dict page + RLE indices.
BYTE_STREAM_SPLIT	Byte-stream split for FLOAT/DOUBLE (transposes byte lanes for better compression).

Definition at line 98 of file types.hpp.

◆ ErrorCode

enum class signet::forge::ErrorCode

strong

Error codes returned by all Signet Forge operations.

Every function in the library that can fail returns an expected<T> whose error payload carries one of these codes together with a human-readable message string. Codes are grouped by subsystem so that callers can pattern-match on categories (I/O, format corruption, unsupported features, licensing) without inspecting the message text.

Enumerator
OK	Operation completed successfully (no error).
IO_ERROR	A file-system or stream I/O operation failed (open, read, write, rename).
INVALID_FILE	The file is not a valid Parquet file (e.g. missing or wrong magic bytes).
CORRUPT_FOOTER	The Parquet footer (FileMetaData) is missing, truncated, or malformed.
CORRUPT_PAGE	A data page failed integrity checks (bad CRC, truncated, or exceeds size limits).
CORRUPT_DATA	Decoded data is corrupt or inconsistent (e.g. out-of-range dictionary index).
INVALID_ARGUMENT	A caller-supplied argument is outside the valid range or violates a precondition.
UNSUPPORTED_ENCODING	The file uses an encoding not supported by this build (e.g. BYTE_STREAM_SPLIT on integers).
UNSUPPORTED_COMPRESSION	The file uses a compression codec not linked into this build (ZSTD, LZ4, Gzip).
UNSUPPORTED_TYPE	The file contains a Parquet physical or logical type that is not implemented.
SCHEMA_MISMATCH	The requested column name or type does not match the file schema.
OUT_OF_RANGE	An index, offset, or size value is outside the valid range.
THRIFT_DECODE_ERROR	The Thrift Compact Protocol decoder encountered invalid or malicious input.
ENCRYPTION_ERROR	An encryption or decryption operation failed (bad key, tampered ciphertext, PME error).
HASH_CHAIN_BROKEN	The cryptographic audit hash chain is broken, indicating data tampering.
LICENSE_ERROR	The commercial license is missing, invalid, or the build is misconfigured.
LICENSE_LIMIT_EXCEEDED	An evaluation-tier usage limit has been exceeded (rows, users, nodes, or time).
INTERNAL_ERROR	An unexpected internal error that does not fit any other category.

Definition at line 49 of file error.hpp.

◆ EscalationLevel

enum class signet::forge::EscalationLevel : int32_t

strong

Escalation hierarchy for incident routing.

Enumerator
L1_OPERATIONS	First-line operations team.
L2_ENGINEERING	Engineering / DevOps on-call.
L3_MANAGEMENT	Management / CISO notification.
L4_REGULATORY	Regulatory authority notification (DORA Art.19)

Definition at line 64 of file incident_response.hpp.

◆ HaltReason

enum class signet::forge::HaltReason : int32_t

strong

Reason for system halt — EU AI Act Art.14(4) "stop button".

Enumerator
MANUAL	Operator manually halted the system.
SAFETY_THRESHOLD	Override rate exceeded safety threshold.
ANOMALY_DETECTED	Anomalous behavior detected.
REGULATORY	Regulatory or compliance-driven halt.
MAINTENANCE	Scheduled maintenance halt.
EXTERNAL	External event (market halt, circuit breaker)

Definition at line 76 of file human_oversight.hpp.

◆ IncidentPhase

enum class signet::forge::IncidentPhase : int32_t

strong

NIST SP 800-61 incident response lifecycle phases.

Enumerator
PREPARATION	Pre-incident readiness.
DETECTION	Anomaly detection / alert triage.
CONTAINMENT	Limit blast radius.
ERADICATION	Remove root cause.
RECOVERY	Restore normal operations.
LESSONS_LEARNED	Post-incident review (DORA Art.13)

Definition at line 46 of file incident_response.hpp.

◆ IncidentSeverity

enum class signet::forge::IncidentSeverity : int32_t

strong

Incident severity per DORA Art.10(1) classification.

Enumerator
P4_LOW	Minor, no customer impact.
P3_MEDIUM	Limited impact, workaround available.
P2_HIGH	Significant impact, SLA breach risk.
P1_CRITICAL	Major outage, data loss, regulatory notification required.

Definition at line 56 of file incident_response.hpp.

◆ InferenceType

enum class signet::forge::InferenceType : int32_t

strong

Classification of the ML inference operation.

Covers common ML workloads from classical models to LLM generation.

Enumerator
CLASSIFICATION	Binary or multi-class classification.
REGRESSION	Continuous value prediction.
EMBEDDING	Vector embedding computation.
GENERATION	LLM text generation.
RANKING	Ranking/scoring of candidates.
ANOMALY	Anomaly/outlier detection.
RECOMMENDATION	Recommendation system inference.
CUSTOM	Application-specific inference type.

Definition at line 47 of file inference_log.hpp.

◆ LogicalType

enum class signet::forge::LogicalType : int32_t

strong

Parquet logical types (from parquet.thrift LogicalType union).

Logical types add semantic meaning on top of a PhysicalType. For example, a STRING column is stored as BYTE_ARRAY but interpreted as UTF-8 text.

See also: PhysicalType, ColumnDescriptor

Enumerator
NONE	No logical annotation — raw physical type.
STRING	UTF-8 string (stored as BYTE_ARRAY).
ENUM	Enum string (stored as BYTE_ARRAY).
UUID	RFC 4122 UUID (stored as FIXED_LEN_BYTE_ARRAY(16)).
DATE	Calendar date — INT32, days since 1970-01-01.
TIME_MS	Time of day — INT32, milliseconds since midnight.
TIME_US	Time of day — INT64, microseconds since midnight.
TIME_NS	Time of day — INT64, nanoseconds since midnight.
TIMESTAMP_MS	Timestamp — INT64, milliseconds since Unix epoch.
TIMESTAMP_US	Timestamp — INT64, microseconds since Unix epoch.
TIMESTAMP_NS	Timestamp — INT64, nanoseconds since Unix epoch.
DECIMAL	Fixed-point decimal (INT32/INT64/FIXED_LEN_BYTE_ARRAY).
JSON	JSON document (stored as BYTE_ARRAY).
BSON	BSON document (stored as BYTE_ARRAY).
FLOAT16	IEEE 754 half-precision float (FIXED_LEN_BYTE_ARRAY(2)).
FLOAT32_VECTOR	ML embedding vector — FIXED_LEN_BYTE_ARRAY(dim*4). Signet AI-native extension; stored as standard Parquet types with logical annotation only.

Definition at line 41 of file types.hpp.

◆ MitigationStatus

enum class signet::forge::MitigationStatus : int32_t

strong

Mitigation status for a threat.

Enumerator
NOT_MITIGATED	No mitigation in place.
PARTIAL	Some controls, residual risk remains.
MITIGATED	Fully mitigated by implemented controls.
ACCEPTED	Risk accepted per organizational policy.
TRANSFERRED	Risk transferred (insurance, third-party)

Definition at line 68 of file threat_model.hpp.

◆ NotificationChannel

enum class signet::forge::NotificationChannel : int32_t

strong

Notification channel for incident communications.

Enumerator
INTERNAL_LOG	System log only.
EMAIL	Email to responsible parties.
PAGER	PagerDuty / on-call alert.
REGULATORY	Formal regulatory notification (DORA Art.19(1))

Definition at line 72 of file incident_response.hpp.

◆ OnnxTensorType

enum class signet::forge::OnnxTensorType : int32_t

strong

ONNX tensor element data types, mirroring OrtTensorElementDataType.

Numeric values match the ONNX Runtime C API's OrtTensorElementDataType enum exactly, so they can be cast directly via static_cast<> when constructing OrtValues.

Note: Not all types have Signet TensorDataType equivalents. Use from_onnx_type() to convert, which returns an error for unsupported types (STRING, UINT16, UINT32, UINT64, BFLOAT16).

See also: to_onnx_type, from_onnx_type, onnx_type_name

Enumerator
UNDEFINED	No type (invalid / uninitialized)
FLOAT	32-bit IEEE float (float32)
UINT8	8-bit unsigned integer
INT8	8-bit signed integer
UINT16	16-bit unsigned integer
INT16	16-bit signed integer
INT32	32-bit signed integer
INT64	64-bit signed integer
STRING	Variable-length string.
BOOL	Boolean (1 byte)
FLOAT16	16-bit IEEE float (float16)
DOUBLE	64-bit IEEE float (float64)
UINT32	32-bit unsigned integer
UINT64	64-bit unsigned integer
BFLOAT16	Brain floating-point (bfloat16)

Definition at line 39 of file onnx_bridge.hpp.

◆ OrderType

enum class signet::forge::OrderType : int32_t

strong

Order type classification for MiFID II RTS 24 Annex I Table 2 Field 7.

Enumerator
MARKET	Market order.
LIMIT	Limit order.
STOP	Stop order.
STOP_LIMIT	Stop-limit order.
PEGGED	Pegged order.
OTHER	Other order type.

Definition at line 70 of file decision_log.hpp.

◆ OverrideAction

enum class signet::forge::OverrideAction : int32_t

strong

What action the human override took — EU AI Act Art.14(4).

Enumerator
APPROVE	Human approved the AI system's output as-is.
MODIFY	Human modified the AI system's output.
REJECT	Human rejected the AI system's output entirely.
ESCALATE	Human escalated to a higher authority.
HALT	Human triggered system halt ("stop button")

Definition at line 67 of file human_oversight.hpp.

◆ OverrideSource

enum class signet::forge::OverrideSource : int32_t

strong

Source of a decision or override — EU AI Act Art.14(4).

Tracks whether an output was produced algorithmically or overridden by a human.

Enumerator
ALGORITHMIC	Original AI system output (no human intervention)
HUMAN	Human operator override.
AUTOMATED	Automated safety system override (e.g. risk gate)

Definition at line 60 of file human_oversight.hpp.

◆ PageType

enum class signet::forge::PageType : int32_t

strong

Parquet page types within a column chunk.

Enumerator
DATA_PAGE	Data page (Parquet 1.0 format).
INDEX_PAGE	Index page (reserved, not used by Signet).
DICTIONARY_PAGE	Dictionary page — contains the value dictionary for RLE_DICTIONARY columns.
DATA_PAGE_V2	Data page v2 (Parquet 2.0 format with separate rep/def level sections).

Definition at line 127 of file types.hpp.

◆ PhysicalType

enum class signet::forge::PhysicalType : int32_t

strong

Parquet physical (storage) types as defined in parquet.thrift.

Every column in a Parquet file stores values in one of these physical representations. The mapping from C++ types is provided by parquet_type_of.

See also: LogicalType, parquet_type_of, ColumnDescriptor

Enumerator
BOOLEAN	1-bit boolean, bit-packed in pages.
INT32	32-bit signed integer (little-endian).
INT64	64-bit signed integer (little-endian).
INT96	96-bit value (deprecated — legacy Impala timestamps).
FLOAT	IEEE 754 single-precision float.
DOUBLE	IEEE 754 double-precision float.
BYTE_ARRAY	Variable-length byte sequence (strings, binary).
FIXED_LEN_BYTE_ARRAY	Fixed-length byte array (UUID, vectors, decimals).

Definition at line 20 of file types.hpp.

◆ QuantizationScheme

enum class signet::forge::QuantizationScheme : int32_t

strong

Identifies the quantization method used for vector compression.

See also: QuantizationParams, Quantizer, Dequantizer

Enumerator
SYMMETRIC_INT8	`value = round(float / scale)`, range [-127, 127].
ASYMMETRIC_INT8	`value = round((float - zero_point) / scale)`, range [0, 255].
SYMMETRIC_INT4	`value = round(float / scale)`, range [-7, 7], nibble-packed.

Definition at line 60 of file quantized_vector.hpp.

◆ RegulatoryChangeType

enum class signet::forge::RegulatoryChangeType : int32_t

strong

Type of regulatory change being tracked.

Enumerator
NEW_REGULATION	Entirely new regulation enacted.
AMENDMENT	Existing regulation amended.
GUIDANCE	Supervisory guidance or interpretation.
TECHNICAL_STANDARD	RTS/ITS (Regulatory/Implementing Technical Standards)
ENFORCEMENT	Enforcement action or precedent.
DEPRECATION	Regulation repealed or superseded.

Definition at line 45 of file regulatory_monitor.hpp.

◆ RegulatoryImpact

enum class signet::forge::RegulatoryImpact : int32_t

strong

Impact level of a regulatory change on the system.

Enumerator
NONE	No system impact.
INFORMATIONAL	Awareness only, no action needed.
LOW	Minor documentation update.
MEDIUM	Code/configuration changes required.
HIGH	Significant architectural changes.
CRITICAL	Immediate action required (compliance deadline)

Definition at line 55 of file regulatory_monitor.hpp.

◆ RegulatoryRegime

enum class signet::forge::RegulatoryRegime : int32_t

strong

Regulatory regime(s) applicable to the data.

Enumerator
NONE
GDPR	EU General Data Protection Regulation.
MIFID2	Markets in Financial Instruments Directive II.
DORA	Digital Operational Resilience Act.
EU_AI_ACT	EU Artificial Intelligence Act.
SOX	Sarbanes-Oxley Act.
SEC_17A4	SEC Rule 17a-4 (records retention)
PCI_DSS	Payment Card Industry Data Security Standard.
HIPAA	Health Insurance Portability and Accountability Act.

Definition at line 65 of file data_classification.hpp.

◆ Repetition

enum class signet::forge::Repetition : int32_t

strong

Parquet field repetition types (nullability / cardinality).

See also: ColumnDescriptor::repetition

Enumerator
REQUIRED	Exactly one value per row (non-nullable).
OPTIONAL	Zero or one value per row (nullable).
REPEATED	Zero or more values per row (list).

Definition at line 140 of file types.hpp.

◆ ReportFormat

enum class signet::forge::ReportFormat

strong

Output serialization format for compliance reports.

Enumerator
JSON	Pretty-printed JSON object (default)
NDJSON	Newline-delimited JSON — one record per line (streaming-friendly)
CSV	Comma-separated values with header row.

Definition at line 25 of file compliance_types.hpp.

◆ RiskGateResult

enum class signet::forge::RiskGateResult : int32_t

strong

Outcome of the pre-trade risk gate evaluation.

Records whether the order passed, was rejected, modified, or throttled by the risk management system.

Enumerator
PASSED	All risk checks passed.
REJECTED	Order rejected by risk gate.
MODIFIED	Order modified by risk gate (e.g., size reduced)
THROTTLED	Order delayed by rate limiting.

Definition at line 62 of file decision_log.hpp.

◆ StrideCategory

enum class signet::forge::StrideCategory : int32_t

strong

Microsoft STRIDE threat categories.

Enumerator
SPOOFING	Authentication bypass, identity impersonation.
TAMPERING	Unauthorized data modification.
REPUDIATION	Denying actions without proof.
INFORMATION_DISCLOSURE	Unauthorized data exposure.
DENIAL_OF_SERVICE	Resource exhaustion, availability attacks.
ELEVATION_OF_PRIVILEGE	Gaining unauthorized access levels.

Definition at line 50 of file threat_model.hpp.

◆ TensorDataType

enum class signet::forge::TensorDataType : int32_t

strong

Element data type for tensor storage, mapping to ONNX/PyTorch/TF type enums.

Enumerator
FLOAT32	IEEE 754 single-precision (4 bytes)
FLOAT64	IEEE 754 double-precision (8 bytes)
INT32	Signed 32-bit integer.
INT64	Signed 64-bit integer.
INT8	Signed 8-bit integer.
UINT8	Unsigned 8-bit integer.
INT16	Signed 16-bit integer.
FLOAT16	IEEE 754 half-precision (2 bytes)
BOOL	Boolean (1 byte)

Definition at line 148 of file tensor_bridge.hpp.

◆ ThreatSeverity

enum class signet::forge::ThreatSeverity : int32_t

strong

Threat severity classification per NIST SP 800-30.

Enumerator
LOW	DREAD composite < 4.0.
MEDIUM	DREAD composite 4.0 - 6.9.
HIGH	DREAD composite 7.0 - 8.9.
CRITICAL	DREAD composite >= 9.0.

Definition at line 60 of file threat_model.hpp.

◆ TimeInForce

enum class signet::forge::TimeInForce : int32_t

strong

Time-in-force classification for MiFID II RTS 24 Annex I Table 2 Field 8.

Enumerator
DAY	Day order (valid until end of trading day)
GTC	Good-Till-Cancelled.
IOC	Immediate-Or-Cancel.
FOK	Fill-Or-Kill.
GTD	Good-Till-Date.
OTHER	Other.

Definition at line 80 of file decision_log.hpp.

◆ TimestampGranularity

enum class signet::forge::TimestampGranularity

strong

Timestamp granularity for MiFID II RTS 24 Art.2(2) compliance.

Controls the sub-second precision emitted in ISO 8601 timestamp fields. RTS 24 requires nanosecond precision for high-frequency trading; lower granularities may be appropriate for non-HFT reporting regimes.

Enumerator
NANOS	9 sub-second digits (default, MiFID II HFT compliant)
MICROS	6 sub-second digits
MILLIS	3 sub-second digits

Definition at line 36 of file compliance_types.hpp.

◆ VectorElementType

enum class signet::forge::VectorElementType : int32_t

strong

Specifies the numerical precision of each element within a vector column.

Enumerator
FLOAT32	IEEE 754 single-precision (4 bytes per element)
FLOAT64	IEEE 754 double-precision (8 bytes per element)
FLOAT16	IEEE 754 half-precision (2 bytes per element)

Definition at line 47 of file vector_type.hpp.

◆ WalLifecycleMode

enum class signet::forge::WalLifecycleMode : uint8_t

strong

Controls safety guardrails for WAL segment lifecycle operations.

In Production mode, destructive operations like reset_on_open are denied.

Enumerator
Development	Permissive: allows reset_on_open.
Benchmark	Same as Development, for benchmark harnesses.
Production	Strict: reset_on_open is rejected.

Definition at line 766 of file wal.hpp.

Function Documentation

◆ add_vector_column()

SchemaBuilder & signet::forge::add_vector_column	(	SchemaBuilder &	builder,
		const std::string &	name,
		uint32_t	dimension,
		VectorElementType	elem = `VectorElementType::FLOAT32`
	)

inline

Add a vector column to a SchemaBuilder.

Creates a FIXED_LEN_BYTE_ARRAY column with FLOAT32_VECTOR logical type, type_length set to dimension * element_size.

Usage: auto schema = add_vector_column( Schema::builder("embeddings") .column<int64_t>("id") .column<std::string>("text"), "embedding", 768) .build();

Parameters

builder	The SchemaBuilder to add the column to.
name	Column name.
dimension	Number of elements per vector (e.g. 768).
elem	Element type (default FLOAT32).

Returns: Reference to the builder for chaining.

Definition at line 788 of file vector_type.hpp.

◆ append_le32()

void signet::forge::append_le32	(	std::vector< uint8_t > &	buf,
		uint32_t	val
	)

inline

Append a uint32_t in little-endian byte order to a byte buffer.

Parameters

buf	The destination byte buffer.
val	The 32-bit value to append (4 bytes).

Definition at line 32 of file column_writer.hpp.

◆ append_le64()

void signet::forge::append_le64	(	std::vector< uint8_t > &	buf,
		uint64_t	val
	)

inline

Append a uint64_t in little-endian byte order to a byte buffer.

Parameters

buf	The destination byte buffer.
val	The 64-bit value to append (8 bytes).

Definition at line 42 of file column_writer.hpp.

◆ arrow_format_to_tensor_dtype()

expected< TensorDataType > signet::forge::arrow_format_to_tensor_dtype ( const char * format )

inline

Map an Arrow format string to a TensorDataType.

Supports the standard single-character format codes for primitive numeric types and booleans. Multi-character format codes (e.g. "tss:" for timestamp) are not supported and return UNSUPPORTED_TYPE.

Parameters

format Arrow format string (must not be null or empty).

Returns: The corresponding TensorDataType on success, or an UNSUPPORTED_TYPE error for unrecognized format strings.

See also: tensor_dtype_to_arrow_format

Definition at line 218 of file arrow_bridge.hpp.

◆ auto_select_compression()

Compression signet::forge::auto_select_compression	(	const uint8_t *	sample_data,
		size_t	sample_size
	)

inline

Automatically select the best available compression codec.

Priority order: ZSTD > Snappy > LZ4_RAW > LZ4 > UNCOMPRESSED.

ZSTD gives the best compression ratio for most Parquet workloads.
Snappy is the fastest but has moderate compression.
LZ4 is a good middle ground between speed and ratio.

Parameters

sample_data	Reserved for future data-aware heuristics (e.g. entropy estimation). Currently unused.
sample_size	Reserved for future use. Currently unused.

Returns: The Compression enumerator for the highest-priority codec that is registered, or Compression::UNCOMPRESSED if none are available.

See also: CodecRegistry::has(), compress()

Definition at line 270 of file codec.hpp.

◆ bit_pack_8()

void signet::forge::bit_pack_8	(	std::vector< uint8_t > &	out,
		const uint64_t *	values,
		int	bit_width
	)

inline

Pack exactly 8 values at the given bit width into a byte buffer.

Each value occupies bit_width bits, packed LSB-first in little-endian byte order. Appends exactly bit_width bytes to out (since 8 values * bit_width bits = bit_width bytes). If bit_width is 0, no bytes are emitted (all values are implicitly zero).

Parameters

out	Output byte buffer to append packed bytes to.
values	Pointer to exactly 8 unsigned values to pack.
bit_width	Bits per value (0–64).

See also: bit_unpack_8

Definition at line 116 of file rle.hpp.

◆ bit_unpack_8()

void signet::forge::bit_unpack_8	(	const uint8_t *	src,
		uint64_t *	values,
		int	bit_width
	)

inline

Unpack exactly 8 values at the given bit width from a byte buffer.

Reverses the packing performed by bit_pack_8(). Reads bit_width bytes from src and unpacks 8 values of bit_width bits each, stored LSB-first. If bit_width is 0, all output values are set to zero.

Parameters

src	Pointer to at least `bit_width` bytes of packed data.
values	Output array of exactly 8 unsigned values.
bit_width	Bits per value (0–64).

See also: bit_pack_8

Definition at line 154 of file rle.hpp.

◆ build_audit_metadata()

expected< AuditMetadata > signet::forge::build_audit_metadata	(	const AuditChainWriter &	writer,
		const std::string &	chain_id
	)

inline

Build an AuditMetadata from a populated AuditChainWriter.

Extracts the chain summary from a writer that has accumulated entries. The chain_id must be provided by the caller (use generate_chain_id() to create one, or reuse an existing ID for chain continuation).

Parameters

writer	The chain writer to extract metadata from.
chain_id	Unique chain identifier for this audit trail.

Returns: The populated AuditMetadata, or an error if the writer has no entries.

Definition at line 967 of file audit_chain.hpp.

◆ compress()

expected< std::vector< uint8_t > > signet::forge::compress	(	Compression	codec,
		const uint8_t *	data,
		size_t	size
	)

inline

Compress data using the specified codec via the global CodecRegistry.

For Compression::UNCOMPRESSED, returns a verbatim copy of the input without consulting the registry.

Parameters

codec	The compression type to use.
data	Pointer to the raw input bytes.
size	Number of bytes to compress.

Returns: The compressed byte buffer, or an Error if the codec is not registered or compression fails.

See also: decompress(), auto_select_compression()

Definition at line 183 of file codec.hpp.

◆ decision_log_schema()

Schema signet::forge::decision_log_schema ( )

inline

Build the Parquet schema for AI decision log files.

Columns: timestamp_ns INT64 (TIMESTAMP_NS) – Decision timestamp strategy_id INT32 – Strategy identifier model_version BYTE_ARRAY (STRING) – Model version hash decision_type INT32 – DecisionType enum value input_features BYTE_ARRAY (STRING) – JSON array of floats model_output DOUBLE – Primary model output confidence DOUBLE – Model confidence [0,1] risk_result INT32 – RiskGateResult enum value order_id BYTE_ARRAY (STRING) – Associated order ID symbol BYTE_ARRAY (STRING) – Trading symbol price DOUBLE – Decision price quantity DOUBLE – Decision quantity venue BYTE_ARRAY (STRING) – Execution venue notes BYTE_ARRAY (STRING) – Free-text notes chain_seq INT64 – Hash chain sequence number chain_hash BYTE_ARRAY (STRING) – Hex entry hash prev_hash BYTE_ARRAY (STRING) – Hex previous hash

Definition at line 511 of file decision_log.hpp.

◆ decode_varint()

uint64_t signet::forge::decode_varint	(	const uint8_t *	data,
		size_t &	pos,
		size_t	size
	)

inline

Decode an unsigned varint (LEB128) from a byte buffer.

Reads a variable-length encoded unsigned integer starting at data[pos]. On success, pos is advanced past the consumed bytes. Returns 0 with pos unchanged if the buffer is exhausted before a terminating byte is found. Includes overflow protection: decoding stops if the shift exceeds 63 bits.

Parameters

data	Pointer to the encoded byte stream.
pos	Current read position (updated on return).
size	Total size of the byte stream.

Returns: The decoded unsigned integer, or 0 on failure.

See also: encode_varint

Definition at line 84 of file rle.hpp.

◆ decompress()

expected< std::vector< uint8_t > > signet::forge::decompress	(	Compression	codec,
		const uint8_t *	data,
		size_t	size,
		size_t	uncompressed_size
	)

inline

Decompress data using the specified codec via the global CodecRegistry.

For Compression::UNCOMPRESSED, returns a verbatim copy of the input without consulting the registry.

Parameters

codec	The compression type that was used to compress the data.
data	Pointer to the compressed input bytes.
size	Number of compressed bytes.
uncompressed_size	Expected size of the decompressed output (from the Parquet page header).

Returns: The decompressed byte buffer, or an Error if the codec is not registered or decompression fails.

See also: compress(), auto_select_compression()

Definition at line 213 of file codec.hpp.

◆ deserialize_and_verify_chain()

expected< std::vector< HashChainEntry > > signet::forge::deserialize_and_verify_chain	(	const uint8_t *	chain_data,
		size_t	chain_size
	)

inline

Deserialize and verify a chain from serialized bytes in one call.

Combines deserialization with AuditChainVerifier::verify(), returning HASH_CHAIN_BROKEN when verification fails.

Parameters

chain_data	Pointer to the serialized chain bytes.
chain_size	Size of the serialized chain in bytes.

Returns: The verified entries, or an error if deserialization or verification fails.

See also: AuditChainVerifier::verify

Definition at line 998 of file audit_chain.hpp.

◆ encode_varint()

size_t signet::forge::encode_varint	(	std::vector< uint8_t > &	buf,
		uint64_t	value
	)

inline

Encode an unsigned varint (LEB128) into a byte buffer.

Appends the variable-length encoding of value to buf. Each output byte uses 7 data bits and 1 continuation bit (MSB), following the unsigned LEB128 convention used by the Parquet wire format.

Parameters

buf	Output byte buffer to append to.
value	The unsigned integer to encode.

Returns: Number of bytes written (1–10).

See also: decode_varint

Definition at line 62 of file rle.hpp.

◆ f16_to_f32()

float signet::forge::f16_to_f32 ( uint16_t h )

inlinenoexcept

Convert a 16-bit IEEE 754 half-precision value to a 32-bit float.

Handles normals, subnormals, infinities, NaNs, and signed zero.

Definition at line 59 of file vector_type.hpp.

◆ f32_to_f16()

uint16_t signet::forge::f32_to_f16 ( float val )

inlinenoexcept

Convert a 32-bit float to a 16-bit IEEE 754 half-precision value.

Rounds to nearest even. Handles overflow to infinity and subnormals.

Definition at line 100 of file vector_type.hpp.

◆ from_dlpack_dtype()

expected< TensorDataType > signet::forge::from_dlpack_dtype ( DLDataType dl_dtype )

inline

Convert a DLPack DLDataType back to a Signet TensorDataType.

Returns an error for types that have no Signet equivalent (e.g. uint16, uint32, uint64, bfloat16, or multi-lane SIMD types).

Parameters

dl_dtype The DLPack data type descriptor.

Returns: The corresponding TensorDataType, or UNSUPPORTED_TYPE for multi-lane types, bfloat16, or unsupported bit widths.

See also: to_dlpack_dtype

Definition at line 191 of file numpy_bridge.hpp.

◆ from_le_bytes()

template<typename T >

T signet::forge::from_le_bytes ( const std::vector< uint8_t > & bytes )

inline

Reconstruct an arithmetic value from its little-endian byte representation.

If bytes contains fewer than sizeof(T) bytes the result is value-initialized (zero).

Template Parameters

T	An arithmetic type.

Parameters

bytes At least sizeof(T) bytes in little-endian order.

Returns: The reconstructed value.

Definition at line 66 of file statistics.hpp.

◆ from_onnx_type()

expected< TensorDataType > signet::forge::from_onnx_type ( OnnxTensorType ort_type )

inline

Convert an OnnxTensorType back to a Signet TensorDataType.

Parameters

ort_type The ONNX tensor element type.

Returns: The corresponding TensorDataType, or UNSUPPORTED_TYPE for types that have no Signet equivalent (STRING, UINT16, UINT32, UINT64, BFLOAT16, UNDEFINED).

See also: to_onnx_type

Definition at line 90 of file onnx_bridge.hpp.

◆ generate_chain_id()

std::string signet::forge::generate_chain_id ( )

inline

Generate a simple chain identifier based on the current timestamp.

Format: "chain-<hex_timestamp_ns>" (e.g. "chain-1a2b3c4d5e6f7890"). This is NOT cryptographically random – it is a human-readable identifier for correlating chain segments across files.

Definition at line 203 of file audit_chain.hpp.

◆ has_encrypted_page_header_prefix()

bool signet::forge::has_encrypted_page_header_prefix	(	const uint8_t *	data,
		size_t	size
	)

inlinenoexcept

Definition at line 122 of file reader.hpp.

◆ hash_to_hex()

std::string signet::forge::hash_to_hex ( const std::array< uint8_t, 32 > & hash )

inline

Convert a 32-byte SHA-256 hash to a lowercase hexadecimal string (64 chars).

Definition at line 150 of file audit_chain.hpp.

◆ hex_to_hash()

expected< std::array< uint8_t, 32 > > signet::forge::hex_to_hash ( const std::string & hex )

inline

Convert a 64-character lowercase hex string back to a 32-byte hash.

Returns an error if the string is not exactly 64 hex characters.

Definition at line 164 of file audit_chain.hpp.

◆ human_override_log_schema()

Schema signet::forge::human_override_log_schema ( )

inline

Build the Parquet schema for human override log files.

Columns mirror HumanOverrideRecord fields plus hash chain + row lineage.

Definition at line 259 of file human_oversight.hpp.

◆ inference_log_schema()

Schema signet::forge::inference_log_schema ( )

inline

Build the Parquet schema for ML inference log files.

Columns: timestamp_ns INT64 (TIMESTAMP_NS) – Inference timestamp model_id BYTE_ARRAY (STRING) – Model identifier model_version BYTE_ARRAY (STRING) – Model version hash inference_type INT32 – InferenceType enum value input_embedding BYTE_ARRAY (STRING) – JSON array of floats input_hash BYTE_ARRAY (STRING) – SHA-256 of raw input output_hash BYTE_ARRAY (STRING) – SHA-256 of raw output output_score DOUBLE – Primary output score latency_ns INT64 – Inference latency (ns) batch_size INT32 – Batch size input_tokens INT32 – Input token count output_tokens INT32 – Output token count user_id_hash BYTE_ARRAY (STRING) – Hashed user ID session_id BYTE_ARRAY (STRING) – Session identifier metadata_json BYTE_ARRAY (STRING) – Additional JSON metadata chain_seq INT64 – Hash chain sequence number chain_hash BYTE_ARRAY (STRING) – Hex entry hash prev_hash BYTE_ARRAY (STRING) – Hex previous hash

Definition at line 441 of file inference_log.hpp.

◆ load_le32()

uint32_t signet::forge::load_le32 ( const uint8_t * data )

inlinenoexcept

Definition at line 129 of file reader.hpp.

◆ make_column_batch()

SharedColumnBatch signet::forge::make_column_batch	(	std::vector< ColumnDesc >	schema,
		size_t	reserve_rows = `64`
	)

inline

Convenience factory: create a shared batch with a given schema.

Definition at line 403 of file column_batch.hpp.

◆ now_ns()

int64_t signet::forge::now_ns ( )

inline

Return the current time as nanoseconds since the Unix epoch (UTC).

Gap R-5 (MiFID II RTS 25 Art.2-3): Uses system_clock for UTC traceability. The previous implementation used steady_clock, which has no relationship to UTC (arbitrary epoch, typically boot time) — every timestamp it produced was regulatory-invalid under RTS 25.

system_clock::now() returns UTC wall-clock time suitable for regulatory timestamp fields. On POSIX systems this is clock_gettime(CLOCK_REALTIME).

Guarantees monotonically increasing timestamps across concurrent callers (MiFID II RTS 24 Art.2 timestamp ordering, CWE-362 race guard). If system_clock returns a value <= the last observed (e.g. due to NTP step adjustment), the result is bumped to last_ns + 1 to preserve hash chain ordering invariants.

Definition at line 110 of file audit_chain.hpp.

◆ onnx_type_name()

const char * signet::forge::onnx_type_name ( OnnxTensorType t )

inline

Return a human-readable string for an OnnxTensorType value.

Useful for diagnostics, logging, and error messages. Returns "UNKNOWN" for values not in the OnnxTensorType enumeration.

Parameters

t	The ONNX tensor type.

Returns: A static string literal (e.g. "FLOAT", "INT64", "UNDEFINED"). Never returns nullptr.

Definition at line 298 of file onnx_bridge.hpp.

◆ parquet_to_arrow_format()

const char * signet::forge::parquet_to_arrow_format ( PhysicalType pt )

inline

Map a Parquet PhysicalType to an Arrow format string.

Arrow format strings for primitive types: "b" = bool, "c" = int8, "C" = uint8, "s" = int16, "S" = uint16, "i" = int32, "I" = uint32, "l" = int64, "L" = uint64, "f" = float32, "g" = float64, "e" = float16

Parameters

pt	The Parquet physical type to convert.

Returns: Single-character Arrow format string, or nullptr for types that have no direct Arrow primitive mapping (BYTE_ARRAY, FIXED_LEN_BYTE_ARRAY, INT96).

See also: tensor_dtype_to_arrow_format

Definition at line 176 of file arrow_bridge.hpp.

◆ physical_to_tensor_dtype()

expected< TensorDataType > signet::forge::physical_to_tensor_dtype ( PhysicalType pt )

inline

Map a PhysicalType to a TensorDataType (for column export).

Parameters

pt	The Parquet physical type.

Returns: The corresponding TensorDataType, or UNSUPPORTED_TYPE for variable-length types (BYTE_ARRAY, FIXED_LEN_BYTE_ARRAY, INT96).

Definition at line 252 of file arrow_bridge.hpp.

◆ physical_type_byte_size()

size_t signet::forge::physical_type_byte_size ( PhysicalType pt )

inline

Return the byte size for a PhysicalType (primitive types only).

Parameters

pt	The Parquet physical type.

Returns: Byte size of one element (1 for BOOLEAN, 4 for INT32/FLOAT, 8 for INT64/DOUBLE), or 0 for variable-length types.

Note: BOOLEAN returns 1 (byte-aligned storage), even though Arrow uses 1-bit packing in validity bitmaps.

Definition at line 272 of file arrow_bridge.hpp.

◆ prepare_for_onnx() [1/2]

expected< OnnxTensorInfo > signet::forge::prepare_for_onnx ( const OwnedTensor & tensor )

inline

Prepare an OwnedTensor for ONNX Runtime consumption (zero-copy).

Delegates to the TensorView overload via the OwnedTensor's view(). The OwnedTensor must remain valid for the lifetime of the returned info.

Parameters

tensor The OwnedTensor to export (must be valid and contiguous).

Returns: OnnxTensorInfo (zero-copy, is_owner = false), or an error.

See also: prepare_for_onnx(const TensorView&)

Definition at line 222 of file onnx_bridge.hpp.

◆ prepare_for_onnx() [2/2]

expected< OnnxTensorInfo > signet::forge::prepare_for_onnx ( const TensorView & tensor )

inline

Prepare a TensorView for ONNX Runtime consumption (zero-copy).

For all supported numeric types (FLOAT32, FLOAT64, INT32, INT64, INT8, UINT8, INT16, FLOAT16, BOOL), this is zero-copy: the returned OnnxTensorInfo.data points directly into the TensorView's memory.

The TensorView must remain valid for the lifetime of the returned info (is_owner will be false). The tensor must be contiguous; non-contiguous tensors are rejected – call clone() first to produce a contiguous copy.

Parameters

tensor The TensorView to export (must be valid and contiguous).

Returns: OnnxTensorInfo ready for OrtApi::CreateTensorWithDataAsOrtValue, or INTERNAL_ERROR for invalid tensors, UNSUPPORTED_TYPE for non-contiguous tensors or unmappable dtypes.

See also: OnnxTensorInfo, prepare_inputs_for_onnx

Definition at line 176 of file onnx_bridge.hpp.

◆ prepare_inputs_for_onnx()

expected< OnnxInputSet > signet::forge::prepare_inputs_for_onnx ( const std::vector< std::pair< std::string, TensorView > > & inputs )

inline

Prepare a batch of named TensorViews for ONNX Runtime inference.

Each pair is (input_name, tensor_view). All tensors must be valid and contiguous. If any tensor fails preparation, the entire call fails with an error message identifying the failing input by name.

Parameters

inputs Non-empty vector of (name, TensorView) pairs. Names should match the model's input node names.

Returns: OnnxInputSet with all tensors prepared (zero-copy), or INTERNAL_ERROR for empty inputs, or any error that prepare_for_onnx() would return (prefixed with the input name).

See also: OnnxInputSet, prepare_for_onnx

Definition at line 264 of file onnx_bridge.hpp.

◆ register_snappy_codec()

void signet::forge::register_snappy_codec ( )

inline

Register the bundled Snappy codec with the global CodecRegistry.

Call this once at startup (e.g. from a top-level initializer or a codec_init function) to make Compression::SNAPPY available through compress() and decompress().

See also: SnappyCodec, CodecRegistry::register_codec()

Definition at line 608 of file snappy.hpp.

◆ tensor_dtype_name()

const char * signet::forge::tensor_dtype_name ( TensorDataType dtype )

inlinenoexcept

Returns a human-readable name for a TensorDataType.

Definition at line 181 of file tensor_bridge.hpp.

◆ tensor_dtype_to_arrow_format()

const char * signet::forge::tensor_dtype_to_arrow_format ( TensorDataType dtype )

inline

Map a TensorDataType to an Arrow format string.

Parameters

dtype The tensor data type to convert.

Returns: Single-character Arrow format string, or nullptr if no mapping exists (should not occur for valid TensorDataType values).

See also: parquet_to_arrow_format, arrow_format_to_tensor_dtype

Definition at line 193 of file arrow_bridge.hpp.

◆ tensor_element_size()

constexpr size_t signet::forge::tensor_element_size ( TensorDataType dtype )

inlineconstexprnoexcept

Returns the byte size of a single element of the given tensor data type.

Definition at line 165 of file tensor_bridge.hpp.

◆ to_buffer_info()

expected< BufferInfo > signet::forge::to_buffer_info ( const TensorView & tensor )

inline

Create a BufferInfo from a TensorView for Python buffer protocol export.

The tensor must be valid and contiguous. The returned BufferInfo's data pointer points directly into the TensorView's memory (zero-copy). The TensorView must remain valid for the lifetime of the BufferInfo.

Strides are computed as C-contiguous byte strides (innermost dimension has stride = itemsize, outer dimensions are products of inner shapes).

Parameters

tensor A valid, contiguous TensorView.

Returns: BufferInfo describing the tensor layout, or INTERNAL_ERROR for invalid tensors, UNSUPPORTED_TYPE for non-contiguous tensors or unmappable dtypes.

auto info = *to_buffer_info(view);
// Use with pybind11:
//   return py::buffer_info(info.data, info.itemsize, info.format,
//                          info.ndim, info.shape, info.strides);

See also: BufferInfo, NumpyBridge

Definition at line 720 of file numpy_bridge.hpp.

◆ to_dlpack_dtype()

DLDataType signet::forge::to_dlpack_dtype ( TensorDataType dtype )

inline

Convert a Signet TensorDataType to a DLPack DLDataType.

This function is total – all TensorDataType values have a DLPack mapping. Lanes is always set to 1 (scalar).

Mapping:

TensorDataType	DLDataTypeCode	bits
FLOAT32	kDLFloat	32
FLOAT64	kDLFloat	64
FLOAT16	kDLFloat	16
INT32	kDLInt	32
INT64	kDLInt	64
INT16	kDLInt	16
INT8	kDLInt	8
UINT8	kDLUInt	8
BOOL	kDLUInt	8

Parameters

dtype The Signet tensor data type to convert.

Returns: The corresponding DLDataType descriptor.

Note: BOOL is mapped to UInt/8 per DLPack convention (no native bool type).

See also: from_dlpack_dtype

Definition at line 135 of file numpy_bridge.hpp.

◆ to_le_bytes() [1/2]

std::vector< uint8_t > signet::forge::to_le_bytes ( const std::string & value )

inline

Overload for std::string – returns raw bytes (no endian conversion needed).

Parameters

value The string whose bytes are copied verbatim.

Returns: A vector containing the string's byte content.

Definition at line 53 of file statistics.hpp.

◆ to_le_bytes() [2/2]

template<typename T >

std::vector< uint8_t > signet::forge::to_le_bytes ( T value )

inline

Convert an arithmetic value to its little-endian byte representation.

On little-endian platforms (x86, ARM) this is a straight memcpy. On big-endian platforms the bytes are reversed.

Template Parameters

T	An arithmetic type (int32_t, double, etc.).

Parameters

value The value to convert.

Returns: A vector of sizeof(T) bytes in little-endian order.

Definition at line 34 of file statistics.hpp.

◆ to_onnx_type()

OnnxTensorType signet::forge::to_onnx_type ( TensorDataType dtype )

inline

Convert a Signet TensorDataType to the corresponding OnnxTensorType.

All TensorDataType values have a direct ONNX mapping; this function is total (never returns UNDEFINED for valid inputs).

Parameters

dtype The Signet tensor data type.

Returns: The corresponding OnnxTensorType.

See also: from_onnx_type

Definition at line 68 of file onnx_bridge.hpp.

◆ validate_mmap_page_value_count()

expected< size_t > signet::forge::validate_mmap_page_value_count	(	int64_t	num_values,
		const char *	context
	)

inline

Definition at line 256 of file mmap_reader.hpp.

◆ validate_page_value_count()

expected< size_t > signet::forge::validate_page_value_count	(	int64_t	num_values,
		const char *	context
	)

inline

Definition at line 110 of file reader.hpp.

Variable Documentation

◆ HASH_CHAIN_ENTRY_SIZE

constexpr size_t signet::forge::HASH_CHAIN_ENTRY_SIZE = 112

inlineconstexpr

Chain summary stored in Parquet key-value metadata.

Size of a single serialized HashChainEntry in bytes.

Definition at line 89 of file audit_chain.hpp.

◆ kEncryptedPageHeaderMagic

constexpr uint8_t signet::forge::kEncryptedPageHeaderMagic[4] = {'S', 'P', 'H', '1'}

inlineconstexpr

Definition at line 120 of file reader.hpp.

◆ PARQUET_MAGIC

constexpr uint32_t signet::forge::PARQUET_MAGIC = 0x31524150

inlineconstexpr

"PAR1" magic bytes (little-endian uint32) — marks a standard Parquet file.

Definition at line 205 of file types.hpp.

◆ PARQUET_MAGIC_ENCRYPTED

constexpr uint32_t signet::forge::PARQUET_MAGIC_ENCRYPTED = 0x45524150

inlineconstexpr

"PARE" magic bytes (little-endian uint32) — marks a Parquet file with an encrypted footer.

Definition at line 207 of file types.hpp.

◆ parquet_type_of_v

template<typename T >

constexpr PhysicalType signet::forge::parquet_type_of_v = parquet_type_of<T>::value

inlineconstexpr

Convenience variable template: parquet_type_of_v<double> == PhysicalType::DOUBLE.

Definition at line 180 of file types.hpp.

◆ PARQUET_VERSION

constexpr int32_t signet::forge::PARQUET_VERSION = 2

inlineconstexpr

Parquet format version written to the file footer.

Definition at line 201 of file types.hpp.

◆ SIGNET_CREATED_BY

constexpr const char* signet::forge::SIGNET_CREATED_BY = "SignetStack signet-forge version 0.1.0"

inlineconstexpr

Default "created_by" string embedded in every Parquet footer.

Definition at line 203 of file types.hpp.

Namespaces

Classes

Typedefs

Enumerations

Functions

Variables

Typedef Documentation

◆ native_type_of_t

◆ SharedColumnBatch

◆ TDT

◆ WalRecord

Enumeration Type Documentation

◆ BuySellIndicator

◆ ChangeComplianceStatus

◆ ComplianceStandard

◆ Compression

◆ ConvertedType

◆ DataClassification

◆ DataSensitivity

◆ DecisionType

◆ DLDataTypeCode

◆ DLDeviceType

◆ Encoding

◆ ErrorCode

◆ EscalationLevel

◆ HaltReason

◆ IncidentPhase

◆ IncidentSeverity

◆ InferenceType

◆ LogicalType

◆ MitigationStatus

◆ NotificationChannel

◆ OnnxTensorType

◆ OrderType

◆ OverrideAction

◆ OverrideSource

◆ PageType

◆ PhysicalType

◆ QuantizationScheme

◆ RegulatoryChangeType

◆ RegulatoryImpact

◆ RegulatoryRegime

◆ Repetition

◆ ReportFormat

◆ RiskGateResult

◆ StrideCategory

◆ TensorDataType

◆ ThreatSeverity

◆ TimeInForce

◆ TimestampGranularity

◆ VectorElementType

◆ WalLifecycleMode

Function Documentation

◆ add_vector_column()

◆ append_le32()

◆ append_le64()

◆ arrow_format_to_tensor_dtype()

◆ auto_select_compression()

◆ bit_pack_8()

◆ bit_unpack_8()

◆ build_audit_metadata()

◆ compress()

◆ decision_log_schema()

◆ decode_varint()

◆ decompress()

◆ deserialize_and_verify_chain()

◆ encode_varint()

◆ f16_to_f32()

◆ f32_to_f16()

◆ from_dlpack_dtype()

◆ from_le_bytes()

◆ from_onnx_type()

◆ generate_chain_id()

◆ has_encrypted_page_header_prefix()

◆ hash_to_hex()

◆ hex_to_hash()

◆ human_override_log_schema()

◆ inference_log_schema()

◆ load_le32()

◆ make_column_batch()