Corbenic AI Paper: Frozen 12B Scores 180/180 on Verified Work

A single-author preprint from Corbenic AI landed on arXiv this week with a claim big enough that it deserves a careful read rather than a shrug. In the paper, Sietse Schelpe reports that a frozen 12-billion parameter model, paired with a persistent memory of verified solutions, scored 180/180 across 180 fresh instances spanning nine problem families, tested on four architectures from four vendors including both dense and mixture-of-experts models.

The mechanism is the interesting part. Once a problem family is solved and passes an independent verification step that never consults the answer key, every new instance in that family is served by lookup rather than generation. Memory selection takes 1.4 microseconds, and a full reuse completes in 6-23 milliseconds at 36 milliwatt-hours. There are no inference tokens spent on a query the store already answers. The paper also reports 88/88 consistency-gated acceptances on open-ended reasoning tasks and 77/80 on reasoning-method transfer.

There is a systems claim buried in there too. The verified store doubles as working context at a scale the paper says no shipped engine matches, a 6,000,000-token movable window on a single 46 GB GPU. For comparison, the paper notes vLLM stops at 30,399 tokens and SGLang silently truncates past 32,000. If those numbers hold up in independent hands, the practical envelope for long-context work on a single card looks very different from what the current serving stacks offer.

The honest caveat is that this is one author, one preprint, and benchmarks self-reported on the paper's own testbench. The paper itself flags a limit worth taking seriously, approximate similarity retrieval picks the wrong item 94.3% of the time on a 4,500-item store, so the 100% number depends on exact addressing rather than fuzzy lookup. What the write-up does not settle is how the verification step behaves on problems whose correctness is not machine-checkable, or how the approach performs on families it has never seen.

If the results replicate, the interesting shift is not smaller models beating bigger ones but the idea that solved work becomes a cached asset with essentially zero marginal serving cost, which is a very different economic story from paying per token forever.