id
gotchas/meta-llamacpp-negative-holds-but-mention-count-is-growing
Concepts
llama.cpparxivprovenancecitationggmlcitation-drift
Symptom
The existing gotcha meta-llama-cpp-has-no-paper-do-not-invent-one was written against a search that returned 5 results for all:"llama.cpp". It now returns 70. That looks like the negative is decaying - as though a llama.cpp paper might now exist, or as though the original finding was simply incomplete. Neither is true, and the growth is the actual finding.
Root cause
The llama.cpp mention population grew 14x while the underlying fact did not change. Inspected entries on 2026-10-04 confirm every one is a third party USING, BENCHMARKING or ATTACKING the project rather than describing it: 2601.14277 evaluates llama.cpp quantization formats on Llama-3.1-8B-Instruct; 2406.10816 is an Armv9 inference-optimization study; 2511.05502 compares MLX, MLC-LLM, Ollama and llama.cpp on Apple Silicon; 2603.23996 is a forensic artifact analysis of Ollama, LM Studio and llama.cpp; 2608.09225 (BaseRT) lists llama.cpp as a baseline to beat. What changed is that llama.cpp became a default BASELINE, which generates third-party papers about it faster than anyone writes one about it.
Workaround
Keep citing the repository. The existing gotcha's conclusion is correct and needs no change; this record exists so a future reader who re-runs the search and sees 70 results understands why that is not evidence of a paper. Reject any candidate by asking whether the paper's SUBJECT is llama.cpp or whether llama.cpp is merely its experimental setup.
notes
The generalisable lesson is about how negatives decay. An honest negative recorded once is not self-maintaining: the evidence behind it changes underneath it while the conclusion stays true. A future pass that re-verified this gotcha and, seeing a much larger result count, concluded the project must now have published would be committing the same error in reverse - treating absence of current evidence as evidence of absence of the fact. This is the mirror image of the repo's numeric-conflict rule, where two sources disagreeing means keeping both and marking contested; here, one verified search and a much later verified search disagree about the SIZE of the mention population while agreeing about the FACT, and the right response is to record the second reading with its date rather than to overwrite the first or to distrust the conclusion. Note also the asymmetry with the two sibling negatives: llama.cpp and TensorRT-LLM fail quietly (empty or irrelevant results) whereas DeepGEMM fails loudly with a same-name paper, so the check that catches one will not catch the others.