Outdated Drivers Are Slowing You Down
One free scan finds every outdated or missing driver and matches the right update for your exact hardware.Free scan · exact hardware matchWindows Errors? Fix Them Before They Spread
Repair common Windows errors and clear accumulated junk for a smoother, more stable PC - no reinstall needed.Free scan · no reinstalliTechGuides is reader-supported. When you buy through links on our site, we may earn an affiliate commission. As an Amazon Associate I earn from qualifying purchases. Learn more
The 100% decoder score means the reconstruction produced exact round trips in 19 of 19 reported cases. The encoder’s 8.3% means it matched the original lha program’s output byte for byte in 1 of 12 trained cases. Those figures measure different things, so they are not contradictory—and neither is a broad score for AI coding ability.
Jean-Luc Martel’s report describes a clean-room attempt to reconstruct LHA’s -lh5- archive method. Its useful lesson is about evaluation: tests can confirm behavior only for the cases they actually exercise.
Why the two percentages are not comparable
| Result | What was measured | Reported outcome |
|---|---|---|
| Decoder | Exact round trips on the reported test cases | 19 of 19 cases |
| Encoder, trained input classes | Byte-for-byte match with real lha output |
1 of 12 cases (8.3%) |
| Encoder, held-out input classes | Byte-for-byte match with real lha output |
5 of 7 cases (71.4%) |
All figures are reported by Martel in his 2026 article. The decoder test asks whether the reconstruction can recover the original data. The encoder test asks whether it makes exactly the same encoding choices as the original program. A decoder can be correct even when an encoder chooses a different valid representation; byte identity is a stricter requirement than successful decoding.
What’s actually slowing this PC down?
Pick the symptom - the matching free tool is one click away.
What the reconstruction attempted
The target was LHA’s -lh5- method, described in the report as LZSS using an 8 KB window followed by static Huffman coding. Martel says the original program served as a behavioral oracle, while its source was kept sealed as a grading key until the reconstruction was frozen. He also says a cold-recall file was committed beforehand to record what the model believed it already knew.
The work was divided among Gemini 3.1 Pro for the decoder, Codex/GPT-5 for the encoder and cold-recall baseline, and Claude for the design thread. These are the author’s descriptions of the experiment’s protocol and model roles; the repository and artifacts were not independently inspected.
What the decoder’s 100% establishes—and what it does not
In the reported test set, the decoder reproduced the original data in all 19 cases. That is strong evidence that it handled those examples correctly. It does not establish that every rule or mechanism in the original implementation was recovered.
Martel notes that important format details—including the 8 KB window and several coding parameters—were already present in the recorded prior. One inference he highlights was that -lh5- uses pbit = 4 with np = 14, despite a generic macro that suggested a different position-table field width. The result therefore mixes prior knowledge with deductions made during reconstruction; the round-trip score alone cannot separate them.
Why the encoder matched so few trained cases
The encoder was graded against the original program’s exact bytes, not merely on whether those bytes could be decoded to the same input. Encoders can differ in choices that still produce valid output. In this experiment, the report locates the main visible divergence in Huffman code-length assignment: the original assigned lengths in heap-extraction order, while the reconstruction’s canonical assignment broke ties differently. That changed the output bytes.
Rank #3
The trained cases exercised real compression heuristics, and the reconstruction matched just 1 of 12. The held-out set’s higher 5-of-7 rate is not evidence that the encoder generalized better to unfamiliar inputs. Martel says those cases were mostly incompressible or trivial examples that used stored mode or other simple paths, bypassing much of the behavior under scrutiny.
What the corpus did not trigger
A test can reveal a defect only when its inputs reach the relevant code path. The report identifies two omissions that the corpus did not expose:
- A 32 KB buffer threshold for splitting blocks. The tested corpus topped out at 8 KiB, so it never reached the threshold.
- A cap on the match-finder chain. The report says the test cases did not expose this behavior.
Martel summarizes the risk: “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.” A green result is bounded by the cases and behaviors represented in the tests; it cannot certify unexercised mechanisms.
Recommended Free Tools
How to read the experiment’s accounting
Beyond the headline percentages, Martel says six encoder decisions were recalled correctly, five were derived correctly, one was derived incorrectly but passed, two were never triggered, and one visibly diverged. These categories distinguish prior knowledge, successful inference, lucky agreement, blind spots, and an observed mismatch. They also show why one aggregate score cannot explain how a reconstruction arrived at its behavior.
Independent reader supportYour contribution helps us test, update, and keep practical guides available for everyone.What makes this a useful evaluation
The report’s practical point is not that one score proves understanding and another disproves it. It is that test design and scoring rules determine what a result can support.
- Separate functional correctness from exact implementation matching. Round trips test whether data survives; byte comparisons test whether encoding decisions match.
- Choose inputs that exercise the mechanisms being evaluated, including thresholds and less common paths.
- Keep trained and held-out sets comparable in difficulty and behavior. A held-out set dominated by trivial paths cannot demonstrate generalization on complex compression decisions.
- Record what was known beforehand and freeze the evaluation protocol before revealing the answer key. That makes it easier to distinguish recall from inference.
On the evidence Martel reports, the decoder passed its 19 examples, while the encoder rarely reproduced the original program’s bytes on compression-heavy trained cases. The 8.3% figure is a narrow byte-match result—not a general verdict on AI programming—and the 100% figure is a narrow round-trip result, not proof that the full mechanism was understood.
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.
