8Z Lossless Audio Compressor

Evolutionary Roadmap & Architectural Timeline

Creator & Lead Architect
Bojan Dobrecevic
AIM³ Institute
Project Timeline
Conception: Feb 19, 2026
v1.0 Build: Feb 20, 2026 (Claude)
v1.7H Final: Feb 23, 2026
avFLAC v1.2: Mar 9, 2026 (beats lax_t6)
Core AI Coders
Claude · Gemini · ChatGPT
AI Review Board
Grok · Qwen · Kimi · GLM · DeepSeek
Ver. & Era Exec. Arch. Leap Features Size Time Spent
v1.0
"First Light"		 Per-Frame General-Purpose BackendsRelied on external algorithms (LZMA, zlib, zstd) to do the heavy lifting for residuals.
  • Basic LPC (1, 4, 8, 12, 16, 24, 32)
  • RAW & DELTA (1-3) predictors
  • Channel Decorrelation modes
 10,523,097 ~2h 53m
v1.1
"Rice Engine"		 Threaded Native Entropy CodingDropped generic coders. Implemented partitioned Rice coding natively for audio residuals.
  • Partitioned Rice + 5-bit escape
  • Exhaustive LPC search (0-32)
  • Wasted bits & constant detection
 8,266,279 5h 18m
v1.2
"Partition Fix"		 Threaded FLAC-Style PartitionsFixed partition sizing to properly accommodate LPC warmup samples.
  • Native bit-depth warmup storage
  • Nibble-packed Rice parameters
  • Varint subframe length fields
 ~8,200,000 ~13h (est.)
v1.3.1
"Unleashed"		 Threaded State Saving (Fault Tolerance)Introduced real-time checkpointing to survive crashes during massive batch encodes.
  • Multi-windowing (Tukey/Hann/Blackman)
  • Partition-aware LPC screening
  • Exhaustive Rice K-search
 7,982,522 ~10h+ (2w)
v1.4 & v1.5
"The DCC Era"		 Sequential Dynamic Codec ControlTried to speed up encoding by learning patterns, but forced the engine into a slow sequential bottleneck.
  • DCCSMeter (Adaptive budgets)
  • CodecLearner (Pruning)
  • Frame blocksize expanded to 16384
 8,031,058 ~3h
v1.6.1
"Two-Pass"		 Decoupled Scanner-Guided ArchitectureSplit the engine: fast preemptive scanner allocates budgets, allowing the packer to run in massive parallel.
  • Adaptive blocksize by sample rate
  • 0xFFFE Extensible WAV support
  • Checkpoint v4
 8,071,580 37.5 min
v1.7 / 1.7H
"The Hybrids"		 Decoupled Forensic HeuristicsThe ultimate synthesis. Eliminates wasted CPU cycles and hardens the memory pipeline.
  • PERIODIC predictor & Cascaded LPC²
  • Fair MDL Arena (LZMA on Top 4 only)
  • L/R/M Stereo Scanning
  • Adaptive LZMA CPU presets
 8,067,034 77 min 52s
v1.9
"MDL Probes"		 Decoupled MDL Frame Size ProbeCompression cost decides frame size. Pass 0 probes 4 candidates (2048-16384). Matches FLAC-12 on Ethereal Arc.
  • Pass 0 MDL frame-size probe
  • Compression levels 0-10
  • 28-column audit CSV
  • Checkpoint/resume
 ~7,993,366 ~550 min
avFLAC v1.2
"Arena Combiner"		 Decoupled Multi-Encoder Arena + MDL Combiner4 encoders compete per segment. MDL picks best of arena stitch / AF_whole / VF_whole. Beats lax_t6 on Abyssal. Beats OFR on Radiohead.
  • vFLAC v1.5 MDL block probe
  • aFLAC v1.3 segment arena (43 candidates)
  • avFLAC MDL final comparison
  • 8Z_Encode v1.7 orchestrator
 7,958,631 ~1600 min

Table Columns

Ver. & Era - The script version number and its architectural codename.
Exec. - Execution Paradigm (How the script handles system CPU threads).
Arch. Leap - The fundamental change in software engineering for that version.
Features - The mathematical or compression tools added to the engine.
Size - Final .8za file size in bytes.
Time Spent - Total wall-clock encoding time.

Execution Tags

Threaded Standard multiprocessing pool. Fast, but flies "blind".
Sequential Uses 1 core. Slowed down by AI learning feedback loops.
Decoupled Best of both worlds. Sequential fast-scan followed by highly optimized parallel encoding.

Terminology

MDL - Minimum Description Length. The arena where candidates fight for smallest byte size.
DCC - Dynamic Codec Control. A machine-learning approach to prune bad predictors.
L/R/M - Left, Right, and Mid. Scanning all 3 prevents missing hard-panned transients.
LPC - Linear Predictive Coding.
Note: Size and time for v1.2 are estimates. Final values will be updated when encoding completes.