Module format 

Aprender Model Format (.apr)

Binary format for ML model serialization with built-in quality (Jidoka):

• CRC32 checksum (integrity)
• Ed25519 signatures (provenance)
• AES-256-GCM encryption (confidentiality)
• Zstd compression (efficiency)
• Quantization (Q8_0, Q4_0, Q4_1 - GGUF compatible)
• Streaming/mmap (JIT loading)

Format Structure

┌─────────────────────────────────────────┐
│ Header (32 bytes, fixed)                │
├─────────────────────────────────────────┤
│ Metadata (variable, MessagePack)        │
├─────────────────────────────────────────┤
│ Chunk Index (if STREAMING flag)         │
├─────────────────────────────────────────┤
│ Salt + Nonce (if ENCRYPTED flag)        │
├─────────────────────────────────────────┤
│ Payload (variable, compressed)          │
├─────────────────────────────────────────┤
│ Signature Block (if SIGNED flag)        │
├─────────────────────────────────────────┤
│ Checksum (4 bytes, CRC32)               │
└─────────────────────────────────────────┘

Example

use aprender::format::{save, load, ModelType, SaveOptions};
use aprender::linear_model::LinearRegression;

let model = LinearRegression::new();
// ... train model ...

// Save with compression
save(&model, ModelType::LinearRegression, "model.apr", SaveOptions::default())?;

// Load with verification
let loaded: LinearRegression = load("model.apr", ModelType::LinearRegression)?;

Re-exports

pub use golden::verify_logits;

pub use golden::GoldenTrace;

pub use golden::GoldenTraceSet;

pub use golden::GoldenVerifyReport;

pub use golden::LogitStats;

pub use golden::TraceVerifyResult;

pub use model_card::ModelCard;

pub use model_card::TrainingDataInfo;

pub use validation::AprHeader;

pub use validation::AprValidator;

pub use validation::Category;

pub use validation::CheckStatus;

pub use validation::TensorStats;

pub use validation::ValidationCheck;

pub use validation::ValidationReport;

pub use validation::no_validation_result;Deprecated

pub use validation::fail_no_validation_rules;

pub use validation::Gate;

pub use validation::PokaYoke;

pub use validation::PokaYokeResult;

pub use converter::apr_convert;

pub use converter::apr_export;

pub use converter::apr_import;

pub use converter::apr_merge;

pub use converter::AprConverter;

pub use converter::Architecture;

pub use converter::ConvertOptions;

pub use converter::ConvertReport;

pub use converter::ExportFormat;

pub use converter::ExportOptions;

pub use converter::ExportReport;

pub use converter::ImportError;

pub use converter::ImportOptions;

pub use converter::MergeOptions;

pub use converter::MergeReport;

pub use converter::MergeStrategy;

pub use converter::QuantizationType;

pub use converter::Source;

pub use converter::TensorExpectation;

pub use converter::ValidationConfig;

pub use lint::lint_apr_file;

pub use lint::lint_model;

pub use lint::LintCategory;

pub use lint::LintIssue;

pub use lint::LintLevel;

pub use lint::LintReport;

pub use lint::ModelLintInfo;

pub use lint::TensorLintInfo;

pub use sharded::estimate_shard_memory;

pub use sharded::get_shard_files;

pub use sharded::is_sharded_model;

pub use sharded::CacheStats;

pub use sharded::CachedShard;

pub use sharded::ImportPhase;

pub use sharded::ImportProgress;

pub use sharded::ImportReport;

pub use sharded::ShardCache;

pub use sharded::ShardIndex;

pub use sharded::ShardedImportConfig;

pub use sharded::ShardedImporter;

pub use rosetta::ConversionOptions;

pub use rosetta::ConversionPath;

pub use rosetta::ConversionReport;

pub use rosetta::FormatType;

pub use rosetta::InspectionReport;

pub use rosetta::RosettaStone;

pub use rosetta::TensorInfo;

pub use rosetta::VerificationReport;

pub use rosetta_ml::AndonLevel;

pub use rosetta_ml::AnomalyDetector;

pub use rosetta_ml::CanaryFile;

pub use rosetta_ml::CategorySummary;

pub use rosetta_ml::ConversionCategory;

pub use rosetta_ml::ConversionDecision;

pub use rosetta_ml::ConversionIssue;

pub use rosetta_ml::ErrorPattern;

pub use rosetta_ml::ErrorPatternLibrary;

pub use rosetta_ml::FixAction;

pub use rosetta_ml::HanseiReport;

pub use rosetta_ml::JidokaViolation;

pub use rosetta_ml::PatternSource;

pub use rosetta_ml::Priority;

pub use rosetta_ml::Regression;

pub use rosetta_ml::Severity;

pub use rosetta_ml::TarantulaTracker;

pub use rosetta_ml::TensorCanary;

pub use rosetta_ml::TensorFeatures;

pub use rosetta_ml::Trend;

pub use rosetta_ml::WilsonScore;

Modules

compare

HuggingFace weight comparison module (GH-121).

converter

APR Converter Module - Import Pipeline

converter_types

APR Converter Types

gguf

GGUF Import/Export (spec §7.2)

golden

Golden Trace Verification

hexdump

Hex dump and data flow visualization for tensor inspection (GH-122).

lint

APR Format Lint Module - Best Practices & Conventions

model_card

Model Card for .apr format (spec §11)

rosetta

Rosetta Stone - Universal Model Format Converter

rosetta_ml

ML-Powered Rosetta Diagnostics (ROSETTA-ML-001)

sharded

Sharded model import module (GH-127).

v2

APR Format Module (GH-119)

validation

APR Format Validation Module - 100-Point QA Checklist

Structs

DistillationInfo

Complete distillation provenance (spec §6.3.2)

DistillationParams

Distillation hyperparameters (spec §6.3.2)

Flags

Feature flags (bitmask) - spec §3.2

Header

File header (32 bytes)

LayerMapping

Layer mapping for progressive distillation (spec §6.3.2)

LicenseInfo

Commercial license information (spec §9.1)

Metadata

Model metadata (MessagePack-encoded)

ModelInfo

Model information (from inspection)

SaveOptions

Options for saving models

TeacherProvenance

Teacher model provenance for audit trails (spec §6.3.2)

TrainingInfo

Training information

Enums

Compression

Compression algorithm

DistillMethod

Distillation method used (spec §6.3.1)

LicenseTier

License tier levels (spec §9.1)

ModelType

Model type identifiers

Constants

FORMAT_VERSION

Current format version (1.0)

HEADER_SIZE

Header size in bytes

MAGIC

Magic number: “APRN” in ASCII (0x4150524E)

MAX_UNCOMPRESSED_SIZE

Maximum uncompressed size (1GB safety limit)

MMAP_THRESHOLD

Threshold for switching to mmap loading (1MB)

Functions

inspect

Inspect a model file without loading the payload

inspect_bytes

Inspect model data without loading the payload (spec §1.1)

load

Load a model from .apr format

load_auto

Load a model with automatic strategy selection based on file size

load_from_bytes

Load a model from a byte slice (spec §1.1 - Single Binary Deployment)

load_mmap

Load a model using memory-mapped I/O (zero-copy where possible)

save

Save a model to .apr format