Compression Research
Compression Experiments
Overview
Systematic compression testing on a single 41K-character financial document across six compressor versions. Each run measures input characters, packet count, and compression ratio. Results revealed kernel prepend overhead, atomic splitting regressions, clause packing failures, and recovery paths through entity aliasing.
13 Compression Runs
Run 1
v0.6.x
2.42x
Compression Ratio
Input: 41,034 chars
Packets: 207
Run 2
v0.7.0
1.92x
Compression Ratio
Input: 40,200 chars
Packets: 208
Run 3
v0.7.0
1.92x
Compression Ratio
Input: 40,200 chars
Packets: 208
Run 4
v0.8.0
1.38x
Compression Ratio
Input: 40,200 chars
Packets: 368
Run 5
v0.8.0
1.36x
Compression Ratio
Input: 40,200 chars
Packets: 370
Run 6
v0.8.1
1.34x
Compression Ratio
Input: 40,200 chars
Packets: 380
Run 7
v0.8.1
1.35x
Compression Ratio
Input: 40,200 chars
Packets: 378
Run 8
v0.8.2
1.39x
Compression Ratio
Input: 40,200 chars
Packets: 375
Run 9
v0.8.2
1.40x
Compression Ratio
Input: 40,200 chars
Packets: 372
Run 10
v0.9.0
1.83x
Compression Ratio
Input: 4,315 chars
Packets: 37
Note: no kernel
Run 11
v0.9.0
1.57x
Compression Ratio
Input: 4,315 chars
Packets: 37
Note: mini kernel
Run 12
v0.9.0
~1.6x
Compression Ratio
Input: 40,200 chars
Packets: ~220
Note: projected
Run 13
v0.9.0
~1.8x
Compression Ratio
Input: 40,200 chars
Packets: ~220
Note: no kernel
Key Regressions and Recoveries
- Kernel Prepend (v0.7.0): Added kernel metadata to each packet, reducing ratio from 2.42x to 1.92x. 21% performance loss.
- Atomic Splitting (v0.8.0): Clause-level packet splitting caused explosion from 207 to 368 packets. Ratio dropped to 1.38x.
- Clause Packing Failure (v0.8.1): Attempted clause re-packing made it worse (380 packets, 1.34x ratio).
- Incremental Fix (v0.8.2): Modest recovery to 1.39-1.40x through packet merging adjustments.
- Entity Aliasing Breakthrough (v0.9.0): Removed kernel prepend, added entity aliasing. 1.83x on mini-corpus, projected 1.8x on full memo without kernel.
- Mini Kernel Trade (v0.9.0): Optional mini kernel drops ratio to 1.57x but maintains decompression fidelity.
Cross-Model Consensus Tests
Grok (xAI)
Tested decompression of full CloudKitchen memo from AXL packet stream.
- Accepted packet stream as valid AXL
- Spoke AXL natively, parsed genesis packet
- Decompressed full memo end-to-end
- Reasoning time: 17 seconds
- Result: PASS
- Spoke AXL natively, parsed genesis packet
- Decompressed full memo end-to-end
- Reasoning time: 17 seconds
- Result: PASS
GPT-4 (OpenAI)
Tested decompression with self-issued loss contract.
- Accepted packet stream
- Issued own loss contract (safety mode)
- Performed deep semantic audit
- Decompressed with contract guidance
- Result: PASS
- Issued own loss contract (safety mode)
- Performed deep semantic audit
- Decompressed with contract guidance
- Result: PASS
Claude (Anthropic)
Tested decompression of encrypted packet stream.
- First attempt: flagged as potential prompt injection
- Second attempt: complied after framing validation
- Successfully decompressed full memo
- Added safety notes to decompressed text
- Result: PASS (with safety guardrails)
- Second attempt: complied after framing validation
- Successfully decompressed full memo
- Added safety notes to decompressed text
- Result: PASS (with safety guardrails)
Decompressor Receipt Test
Receipt-Mode Performance
Single-pass receipt decompression of 207-packet stream (v0.6.x baseline). No dynamic model calls, deterministic packet reconstruction and merkle verification.
Packet Receipt Reconstruction
Full CloudKitchen memo (41,034 chars) decompress from packet stream.
- Input packets: 207
- Execution time: 5.5 milliseconds
- LLM model calls: 0
- Merkle verification: PASS
- Output fidelity: 100%
- Result: PASS
- Execution time: 5.5 milliseconds
- LLM model calls: 0
- Merkle verification: PASS
- Output fidelity: 100%
- Result: PASS