Outcome-based verification. Does NOT verify the computation at all. An inference node runs the generative model off-chain; k independent "quality assessors" each score the (query, response) pair with a lightweight BERT cross-encoder (nli-deberta-v3, 0.5-1.7 GB), commit-then-reveal their scores, and consensus aggregates them and pays out. The protocol is called PQML.

A DIFFERENT CLAIM, not a cheaper proof of the same claim -- and the paper's framing invites the confusion. zkML proves "THIS committed model produced THIS output on THIS input". PoQ establishes "SOME model produced an output that a BERT cross-encoder scored highly". Model substitution -- precisely the attack zkML exists to prevent -- is not prevented here, only made unprofitable for a RATIONAL node by parameter tuning. CORRECTED 2026-07-13 -- the previous note said "The authors concede that a cheaper model whose output is good enough is rewarded equally." They concede no such thing; they claim the OPPOSITE, and it is their central design goal. Sec 3.3: the reward function is built so that "the node owner would not make more earnings by replacing F with a less powerful F'", and Theorem 1 gives a threshold on alpha above which "there is no motivation for inference nodes to deploy less accurate models from an economical perspective." The guarantee is ECONOMIC and it rests on Hypothesis 1 (all adversaries are rational) and Hypothesis 2 ("There is no domination in the model market") -- which ASSUMES the gap away rather than conceding it. It is not soundness, but it is not a concession either. (Note also that Theorem 1's printed statement -- "the inference node would always choose the most affordable model in the list" -- contradicts its own proof and Sec 3.3; that looks like a typo in the paper.) THE NUMBER THAT MATTERS, and it is theirs: Table 3 gives the Pearson correlation between the cross-encoder score and GPT-4 ground-truth quality across five models as -0.06 to 0.35 (GPT3.5 0.12, GPT4 0.13, Mistral-7B 0.28, Mixtral-8x7b 0.35, Llama3-70b -0.06). On Llama3-70b the judge is ANTI-CORRELATED with the thing it is judging. The previously recorded range "0.12-0.35" silently dropped the negative value. The authors' own reading, verbatim (Sec 6.1) -- the cross-encoder "is capable of capturing the missing of information in the generative AI output. It is unable to identify the subtle difference on the accuracy of the information itself." (The earlier paraphrase "subtle differences on accuracy" was quoted as if verbatim and is not.) A verification scheme is only as sound as its verifier. Note the name: despite "Proof of", nothing is proven. No proof object, no soundness error, no extractor -- it is mechanism design plus a statistical judge. Filed under alternatives rather than inference for exactly that reason. Cost is the genuine contribution, and it is near-zero -- but READ TABLE 2 CAREFULLY. Its latencies (0.12-0.61 s on A100, 3.7-20.5 s on CPU) are BATCH latencies for 1,000 query-response pairs, not per-assessment costs: "The latency is the average waiting time in seconds when processing a batch containing 1,000 query-response pairs because batch processing is way more efficient than processing one pair at a time." PER ASSESSMENT that is ~0.12-0.6 MILLISECONDS on an A100 and ~4-21 ms on CPU (throughput 1,666-8,256/s GPU, 48-241/s CPU) -- the figures previously recorded here were off by ~1000x as per-assessment costs. Consensus settles at ~50 ms regardless of k = 1..30 assessors (Fig 5). Evaluated on 100 SQuAD questions against GPT-3.5/4-turbo, Mistral-7B, Mixtral-8x7B and Llama3-70B. Admitted limits: a single inference node only, NLP only (no image generation), and the cross-encoders are off-the-shelf document-search models not trained for this task. Useful to this SoK as the far endpoint of the trust spectrum: the point where you stop verifying the computation entirely and start verifying the OUTPUT. Everything else on the alternatives page still checks, samples, or re-executes the actual inference.