Model integrity: building this required finding and fixing three errors in our own earlier models — a depositor keep-branch double-count (the escrow returned at keep is the depositor's own money, not income — the fee stream is what pays for the NFT; break-even NFT value is ≈1× backing under rational keeps, not 2×), the trade tax briefly miscounted as token value, and a VRF fee that silently read zero (it is a measured 1.04M-gas prepay priced off gas). The corrected models are the ones every number here comes from.
Aug 19 realism pass — a full review of every input against "does this target the real thing", sixteen findings, all addressed. The structural ones: every "per year" figure now integrates a pace path decaying at the fitted weekly trend of the daily pull counts (a flat ×365 is the zero-decay special case, kept as the paper's stated convention and shown as an explicit upper bound on the capitalization tile); pace seeds from an exponentially-weighted average (half-life 3 days) rather than the flat 7-day mean; the optional reflexive solver's non-believers realize at the realization slider instead of a hardcoded 1×, and realization itself seeds at spot NET of the local price impact of selling the distributed flow into measured pool liquidity. The measurement ones: gas seeds from the median effectiveGasPrice pulls actually paid, sampled evenly across the trailing 7 days (median ≈0.17 gwei vs spike-mean ≈0.5 — and ±1 gwei moves RTP ≈2 pts, which is most of any day-to-day gap vs the paper's 92.2 headline); the floor branch now counts unclaimed finalizations with the bid-takers (they forfeit to the same floor — previously they sat in the kept branch at par in one function and at the floor in another); the cold-pool share reads off the measured inter-pull gap distribution (pulls are bursty, median gap one block, so 0.01% of pulls followed ≥60 min idle where the uniform-spacing interpolation said 0.48%); acceptBidAsTokens was verified from source to pay the identical 90% through the same settlement path (the ~29% of sell-backs taken as $FWA just market-buy tokens with the payout, netting slightly less after the 1% trade tax); and a failed read now degrades to the bundled snapshot's measured values instead of inventing free gas, a 50% keep rate, or an untaxed token — flagged as stale in the stamp. The definitional ones: RTP is labeled and decomposed as a mean over a heavily skewed lottery, with median and tail draws from a lognormal fit of the live backing distribution tilted by the real 1/backing selection; the depositor keep chance co-moves with the NFT value set (a logistic anchored to reproduce the measured pool average); the crown NFT seeds at a typical listing value, not its 300 ETH backing; and break-even belief divides by the rebate at exactly spot. Measured but not yet promoted to seeds: trade-tax receipts at the OwnerSplitter (≈2.5 ETH/day vs ≈5 implied by DexScreener volume × 1% — the proxy may double the tax legs) and the crown transition history (39 holder-to-holder changes in the trailing 30 days vs the 35%/yr judgment — pending source verification of the event semantics); both in sim/measure_real.ts.