What’s actually slowing this PC down?
Pick the symptom - the matching free tool is one click away.
The headline figures come from one reported reconstruction experiment, not a general test of AI coding: a decoder passed 19 of 19 exact round trips, while an encoder reproduced the original program’s bytes in only 1 of 12 trained cases. Those scores measure different outcomes. The decoder result shows that the reconstructed program could recover tested inputs; the encoder score tests whether it made the same byte-level choices as the original.
What was reconstructed—and how it was evaluated
In his 2026 report, Jean-Luc Martel describes a clean-room reconstruction of LHA’s -lh5- archiver method, which he characterizes as LZSS with an 8 KB window and static Huffman coding. The original program served as a behavioral oracle. Martel says its source was kept sealed as a grading key until the reconstruction was frozen, and that a cold-recall file recording the model’s prior beliefs was committed before the work began. These are details of the author’s account, not independently inspected experiment artifacts.
As an Amazon Associate I earn from qualifying purchases.
The work assigned different roles to three models: Gemini 3.1 Pro reconstructed the decoder, Codex/GPT-5 handled the encoder and cold-recall baseline, and Claude ran the design thread. The scores therefore describe this particular setup and protocol; they do not establish how AI systems generally perform at reverse engineering.
Why the decoder’s 100% does not mean the mechanism was fully recovered
The decoder produced exact round trips in all 19 reported cases. A round trip checks whether data encoded in the tested format can be decoded back to the expected original. It does not require the reconstruction to have discovered every internal rule used by the original implementation, especially when several implementations can decode the same stream.
#1 Best Overall
- The logic for each channel sampling rate of 24M/s. General applications around 10M, enough to cope with a variety ofoccasions; 8-channel
- Sampling rate up to: 24 MHz , can be 24MHz. 16MHz, 12MHz, 8MHz, 4MHz, 2MHz, 1MHz, 500KHz, 250KHz, 200KHz, 100KHz, 50KHz, 25KHz;
- The logic for each channel sampling rate of 24M/s. General applications around 10M, enough to cope with a variety ofoccasions;
- Input voltage range: -0.5V to 5.25V; Input Low Voltage: -0.5V to 0.8V; Input High Voltage: 2.0V to 5.25V
- Input Impedance: 1Mohm || 10pF (typical, approximate); Crystal: +/-20ppm, 24MHz
Martel notes that important format details—including the 8 KB window and several coding parameters—were already present in the recorded prior. One notable reported inference was that -lh5- used pbit = 4 with np = 14, even though a generic macro suggested a different position-table field width. The 19 successful cases demonstrate tested behavior, not a complete account of how that behavior was produced.
What the encoder scores measure
For the encoder, success meant matching the bytes emitted by real LHA, not merely producing an archive that decoded to the same content. An encoder can be correct in the round-trip sense while choosing different valid codes or making different compression decisions. Byte-for-byte comparison is a stricter test: it demands the same output choices as the reference implementation.
Rank #2
- 【High-Speed 8-Channel Analysis】Captures digital signals at up to 24MHz across 8 channels, enabling precise debugging of complex protocols like I2C, SPI, and UART—ideal for advanced STEM projects without the limitations of basic 4-channel models.
- 【User-Friendly Design】Base module and breakout board simplify connections to breadboards, microcontrollers, and other setups.
- 【Logic Level Expansion Board】Breaks out all 8 channels to 2.54mm male pins and pads for alligator clips, enabling flexible and secure connections in diverse projects.
- 【Logic Level Breadboard Adapter】 Easily connects the logic analyzer to breadboards, providing direct and convenient access to all 8 channels for prototyping and testing.
- 【Dual USB Connectivity】Comes with both USB-A and Type-C cables for universal compatibility with older PCs, modern laptops, and devices, ensuring hassle-free plug-and-play across Windows, Mac, Linux, and Ubuntu.
| Encoder set | Byte-identical results | What the author says the cases exercised |
|---|---|---|
| Trained input classes | 1 of 12 (8.3%) | Classes exercising actual compression heuristics |
| Held-out input classes | 5 of 7 (71.4%) | Mostly incompressible or trivial cases using stored mode or other trivial paths |
The higher held-out rate should not be read as proof that the encoder generalized better to unfamiliar inputs. Martel attributes much of that set’s success to cases that were easier or bypassed the active compression heuristics. Conversely, the 8.3% result is a byte-identity rate for 12 trained examples—not a general score for coding ability or even a measure of whether 11 outputs could still be decoded correctly.
The Tool Desk
Outbyte PC Repair FREERepair Windows errors before they cause bigger problemsFix Now →Outbyte Driver Updater FREEFix the driver behind crashes, sound loss and screen glitchesFind Drivers →Where the encoder diverged—and what the corpus missed
Observed mismatch: Huffman tie and ordering behavior
Martel identifies Huffman code-length assignment as the main visible source of divergence. In his account, the original assigned lengths in heap-extraction order, while the reconstruction’s canonical assignment broke ties differently. That difference changed the encoded bytes, even where the resulting stream could still represent the same content.
Rank #3
- 8 Digital/Analog inputs (multi-use)
- Decode SPI, I2C, and 23+ more analyzers
- Digital sample rate up to 500 MS/s, Analog sample rate up to 50 MS/s
- 10 Billion+ samples of digital, 500 Million+ samples of analog (uses PC memory, USB 3.0)
- Cross platform - Mac, Windows, & Linux
Untriggered behavior: block splitting and match finding
The report also identifies two omissions that the test corpus did not expose: a 32 KB buffer threshold for splitting blocks and a cap on the match-finder chain. The corpus topped out at 8 KiB, so it never reached the stated block-splitting threshold. Martel’s account says six encoder decisions were recalled correctly, five derived correctly, one derived incorrectly but passed, two were never triggered, and one visibly diverged. A test can reveal an error only when its inputs activate the relevant behavior.
Independent reader supportYour contribution helps us test, update, and keep practical guides available for everyone.What this experiment says about behavioral tests
The central lesson is about evaluation design, not a universal verdict on AI understanding. A narrow corpus can make a reconstruction look complete because it tests only a subset of the implementation’s decision paths. Exact results on those cases remain real evidence—but only about those cases and that metric.
Rank #4
- 16 channels dual-mode support: ①Stream mode captures and transfers data in real time for long sample duration; ②Buffer mode captures and stores data temporarily for high sample rate
- USB 2.0 Type-C interface with up to 16G sample depth in stream mode
- Support for adjustable threshold and shielded wires for a better, cleaner waveform
- 256Mbits on-board SDRAM memory with multiple buffer modes
- Compatibility with WinXP-Win10, macOS, and Linux, supporting nearly 100 protocol decoders, and being open-source on Github
- Separate correctness from identity. A successful decode tests recovery of content; byte matching tests whether the encoder made the same choices as the reference.
- Make the corpus exercise the mechanisms at issue. Include inputs that reach thresholds, trigger alternative paths, and stress heuristic decisions—not only easy or stored-mode cases.
- Keep prior knowledge visible. A documented cold-recall baseline helps distinguish details already known from rules inferred during reconstruction.
- Freeze the evaluation protocol before revealing the answer. Martel describes sealing the original source until the reconstruction was frozen, limiting the opportunity to tailor it to the grading key.
- Report untested behavior plainly. A passing test says nothing about a branch that its inputs never reached.
As Martel puts it: “A strong oracle over a narrow corpus hides exactly the mechanisms your corpus never triggers, and it hides them silently, because everything it can see is green.”
Quick Recap
Product prices and availability are accurate as of the date/time indicated and are subject to change. Any price and availability information displayed on Amazon at the time of purchase will apply.
