The export window runs seven calendar days [5] and holds 3,931 requests, which averages roughly 205,000 tokens per request [2]. Code does not come out of a model at that rate. That volume is context going in: the same source files, the same test output, re-read turn after turn while the harness walks the tree. What the published totals support is a blended rate and nothing finer. $28.35 across 612.9 million tokens is about 4.6 cents per million [1].
The bookkeeping is the part that earns trust. LovelaceSharp stopped on March 17 at commit 9b26f05 and had no activity until August 31 [6], which gives a clean baseline to diff against rather than a repository someone claims an AI wrote. The author also trims his own bill: the DeepSeek export covers August 30 through September 5, but only August 31 through September 3 is attributed to the burst, discarding 192.5 million tokens and $8.33 [4]. His headline reads 800 million tokens for $36 [15]. His method says 612.9 million and $28.35 [5]. Quote the second one.
The representation swap is the dangerous change, because every numeric type in the library sits on top of `Natural` [9]. Knowing Knuth Algorithm D is table stakes for a model of this class; it shipped alongside schoolbook and Karatsuba multiplication, short division, native carry/borrow and divide-and-conquer radix conversion [10]. The move that matters came after: tests aimed specifically at limb boundaries, plus a randomized differential suite run against `System.Numerics.BigInteger` with adversarial operands around 2^64 and 2^128 [11]. The author describes the useful behaviour as a loop, from changing the representation to identifying the new failure boundaries to constructing an independent oracle and cross-checking against it [12]. A differential suite is something a reviewer can rerun on a clean checkout, which is more than a token total offers. An exact two-prime NTT path for very large operands landed on top of the limb work [13].
For 4.6 cents per million to transfer, your per-request context has to look like this one: a working set small enough that re-reading it dominates the bill, priced on the same vendor's sheet during the same week. The oracle was free here, since `System.Numerics.BigInteger` already ships in the framework [11], which is a luxury the next REST endpoint does not have [14].
The library also carries a Lean project for formally verifying the underlying arithmetic model [3], and the author notes that formal proofs can prove something subtly different from what the optimized production implementation actually does [14]. On this run, the artifact that would catch a carry bug at 2^64 is the randomized differential suite, not the proof.
