## 1. Problem and contribution The determinant certificates of [1, 2] exclude a candidate prime discriminant $D \equiv 3 \pmod 8$ from having $h(-D)=1$ when a computed *gap* is positive. To date they have excluded every $D$ above $331$ [2]. The remaining question in this line of work is which discriminants a better design or a larger prime range $P$ can still remove, and at what cost. The present paper studies where the gap comes from. It makes four contributions. 1. **Four equal prices (Lemma 1).** Under $h(-D)=1$, four different expressions for the net split-prime charge coincide. So each gives a valid conditional certificate. Their differences, measured on the true field, split the margin exactly as $$\mathrm{gap}(D,P)=\mathrm{gap}_\chi(D,P)+\Delta(D,P).$$ Here $\mathrm{gap}_\chi$ depends only on the splitting law $\chi_D$. $\Delta$ is a non-negative-in-practice capacity term that sees only the failure of unique factorization in $\mathbb Z[\omega]$, $\omega=(1+\sqrt{-D})/2$. 2. **$P$-flatness, and where it ends.** For the DF17 Charlier design of [2], $\mathrm{gap}_\chi(D,P)$ changes by at most $0.021$ over $9\times10^6\le P\le7.56\times10^8$ at $D=331,347,379,467$. A heuristic count of the leading slopes explains this. In that range all growth of the margin with $P$ comes from $\Delta$, and at large $P$ from the composite *Excess*. Raising $P$ does not strengthen the splitting-law part. It harvests more evidence that $h(-D)>1$, which grows like $\log P$. Above about $10^9$ at $D=331$, and above about $10^8$ at $D=251$, the archimedean slope rises once the certificate's split-prime correction is capped. $\mathrm{gap}_\chi$ then drifts down, by up to $0.07$ per e-fold. 3. **A forecast with a blind test.** $\Delta$ needs no determinant evaluation, and $\mathrm{gap}_\chi$ can be measured cheaply at small $P$. Together they predict the certificate at large $P$. The forecast for $D=331$, $P=3\times10^9$ was $-0.035\pm0.01$, made before the run. The full certificate then returned $-0.0451$. 4. **An obstruction for design tuning at $D=331$, and a direct measurement of escalation.** Over 27 nearby DF designs and four other families, $c(331)=-\mathrm{gap}_\chi\ge0.946$. Every forecast at $P=3\times10^9$ is negative, the best being $-0.028$. Direct runs beyond $m^5$ give $\mathrm{gap}(331)=-0.0451$, $-0.0248$, $-0.0108$ and $-0.0129$ at $3\times10^9$, $6\times10^9$, $1.3\times10^{10}$ and $2\times10^{10}$ (Section 6). The margin peaks near $1.4\times10^{10}$, about $0.011$ short of zero. So escalating $P$ alone does not exclude 331 in this contract. A second blind forecast ($0.000\pm0.005$ at $2\times10^{10}$) failed for the same reason: past $10^9$ the drift of $\mathrm{gap}_\chi$ accelerates. None of this changes the published endpoint $D\le331$ [2]. The results explain why that endpoint is hard to move, and which kinds of improvement can or cannot move it. ## 2. Setting and the four prices We use the singleton certificate `certify_hybrid_k3_interval` of [1, 2] (supplement of [1], v13) with the DF17 Charlier contract of [2]. The parameters are: - $K=17$, $c_R=1.2$, $\varepsilon_J=0.128$, Charlier parameters $(a,\beta)=(27/50,-9/10)$; - $m=(D+1)/4$, and the guaranteed set $G$ = irreducible primitive candidates $\alpha\in\mathbb Z[\omega]$ with $m\le N(\alpha)\le P/m$; - Wronskian window $(140000,160000]$ for $P\le3.6\times10^7$, $(500000,515000]$ for $P\in\{1.69,3.12\}\times10^8$, and $(800000,820000]$ with $k\le4$ products for $P\ge7.56\times10^8$. The certificate returns $$\mathrm{gap}=\mathrm{padic}-\mathrm{split\_loss}+\mathrm{ufd\_save}-\mathrm{arch}.$$ In this expression: - padic is the all-inert $p$-adic credit; - arch is the archimedean (Charlier–Fischer) bound; - split_loss charges each primitive candidate $\alpha$ with $m\le N(\alpha)\le P$ a row weight $\mathrm{wt}(N(\alpha))$, the loss of $p$-adic credit when $N(\alpha)$ is not inert; - ufd_save returns the weight of the candidates that unique factorization would force to be products of elements of $G$. Write $\mathrm{split\_sum}=\sum_\alpha \mathrm{wt}(N(\alpha))$ for the bare row charge. Write $W_{\rm formal}$ for the sum of $\mathrm{wt}$ over all formal products of $2$ to $4$ elements of $G$, with the orientation multiplicities of [1, §6.4]. All weights below are these *bare row weights*: column (Legendre) credits are excluded, and the certificate computes those separately from the candidate list. The *formal net* is $\mathrm{net}_{\rm formal}=\mathrm{split\_sum}-W_{\rm formal}$, and the gap is linear in it. Define three more nets with the same weights: - $\mathrm{net}_{\rm red}=\mathrm{split\_sum}-W_{\rm red}$, where $W_{\rm red}$ sums $\mathrm{wt}$ over the *distinct* reducible candidates that are products of elements of $G$. Then $\mathrm{Excess}=W_{\rm formal}-W_{\rm red}\ge0$ counts repeated factorizations. - $\mathrm{net}_{\rm pn}=\sum\{\mathrm{wt}(N(\alpha)) : N(\alpha)\ \text{prime}\}$. - $\mathrm{net}_\chi=\sum\{\mathrm{wt}(p) : p \text{ prime},\ m\le p\le P,\ \chi_D(p)=1\}$. **Lemma 1.** Suppose $h(-D)=1$ and $P1$ the four nets differ. Define: - $W_{\rm fake}=\mathrm{net}_{\rm red}-\mathrm{net}_{\rm pn}$, irreducible candidates of composite norm; - $W_{\rm miss}=\mathrm{net}_\chi-\mathrm{net}_{\rm pn}$, split primes whose prime ideals are not principal. Then $$\Delta:=\mathrm{net}_\chi-\mathrm{net}_{\rm formal}=W_{\rm miss}-W_{\rm fake}+\mathrm{Excess},\qquad \mathrm{gap}=\mathrm{gap}_\chi+\Delta. \tag{1}$$ **Which prices are usable.** Only the formal price can be computed uniformly over a box of discriminants, because it depends only on the candidate list. The other three are per-$D$ quantities. The interpretation of (1) is direct. $\mathrm{gap}_\chi$ is the margin that the splitting law alone supports: it is what the certificate would give if $\mathbb Z[\omega]$ factored uniquely up to $P$. $\Delta$ is the margin that comes from non-principal ideals that are visible below $P$. A certificate excludes $D$ either because $\mathrm{gap}_\chi>0$ or because $\Delta$ exceeds $-\mathrm{gap}_\chi$. Both give valid proofs, but they scale differently with $P$. **Sanity checks.** For the $h=1$ fields the four nets agree to $10^{-10}$: $5.9200216126$ at $D=163$, $P=3.6\times10^7$, and $4.9294024904$ at $D=67$, $P=10^6c(D)$. With $\Delta$ growing by about $0.03$ per e-fold, a shortfall of $0.03$ in the gap costs roughly a factor $e$ in $P$. In our four cases the deficit is larger at $h=3$ ($c=0.96$ at 331, $1.00$ at 379) than at $h=5,7$ ($0.80$, $0.79$). It tracks $\mathrm{net}_\chi$ only roughly ($5.11, 5.08, 4.93, 4.77$ at $9\times10^6$), because arch also varies with $D$. The $W_{\rm miss}$ slope $(1-1/h)\times0.47$ increases with $h$, so fields of small class number are hard on both counts. An $h=1$ field has $\Delta\equiv0$. The singleton method is in effect a detector of non-principal ideals, and its yield grows logarithmically in $P$. ![Two-panel chart against log P for D = 331, 347, 379 and 467. Top panel: the deficit c(D) = −gap_chi is nearly flat at about 0.96, 0.80, 1.00 and 0.80, while Delta rises slowly; 467 is above its deficit from the start, 379 crosses near 1.1e8, 347 near 5.7e8, and 331 is still 0.045 short at 3e9. Bottom panel: the certificate gap rises roughly linearly in log P and crosses zero at the same places](supplement/fig1_flatness.png) *Figure 1.* Top panel: the deficit $c(D)=-\mathrm{gap}_\chi$ (solid, flat) and $\Delta$ (dashed) against $P$ on a log scale, for $P\le3\times10^9$. Bottom panel: the certificate gap. A $D$ is excluded where its $\Delta$ curve rises above the flat level $c(D)$. ## 4. Forecasting the certificate without running it Equation (1) separates the certificate into two parts with different costs. $\Delta(D,P)$ needs only the candidate list and $\chi_D$, through `deltaP.py`: no determinant, no rank window and no interval arithmetic. It takes 146 s at $D=331$, $P=3\times10^9$. $\mathrm{gap}_\chi$ needs one certificate run, but by Section 3 a run at small $P$ is enough; at $P=9\times10^6$ it takes about 5 s. The forecast is $$\widehat{\mathrm{gap}}(D,P)=\mathrm{gap}_\chi(D,9\times10^6)+\Delta(D,P). \tag{2}$$ Its error is exactly the drift of $\mathrm{gap}_\chi$ between $9\times10^6$ and $P$. **Blind test.** On 28 October 2026 at 18:17Z we recorded, before running the certificate, the forecast $\widehat{\mathrm{gap}}(331,3\times10^9)=-0.035\pm0.01$, using $\mathrm{gap}_\chi\approx-0.97$ and $\Delta=0.9341$. The full certificate then ran with window $(800000,820000]$ and $k\le4$ products. It took 602 s with $|G|=1{,}216{,}408$ and returned $\mathrm{gap}=-0.0451$, with padic $18.4638$, split_loss $10.2755$, ufd_save $3.4862$ and arch $11.7196$. The forecast missed by $0.010$, at the edge of the stated band. The miss is the drift $\mathrm{gap}_\chi(331)$: $-0.966\to-0.979$ between $7.56\times10^8$ and $3\times10^9$. That drift, about $-0.010$ per e-fold, is larger than over the earlier range. We report it as the main source of forecast error at large $P$. **Forecasts at $3\times10^9$ for the excluded neighbours.** Using `deltaP.py`: - $D=347$: $\Delta=0.9105$ and Excess $1.1152$, so $\widehat{\mathrm{gap}}\approx-0.808+0.911=+0.10$. - $D=379$: $\Delta=1.0967$ and Excess $0.5474$, so $\widehat{\mathrm{gap}}\approx-1.006+1.097=+0.09$. Both are consistent with the exclusions of [2], which were obtained at $P=7.56\times10^8$. **A bookkeeping caveat.** When $J-1>m^4-1$, the certificate's cap $q_{\rm cap}=\min(Q_{\rm guar},m^4-1,J-1,\ldots)$ becomes $m^4-1$. Candidates with norm in $(q_{\rm cap},\lfloor M_{\rm tri}/n\rfloor]$ then move from the slope credit, which lies outside split_loss, to a flat column credit inside split_loss. As a result, the certificate's split_loss falls below the bare split_sum. At $D=347$ the difference is $+0.0061$ at $4\times10^8$ and $-0.0387$ at $7.56\times10^8$. The threshold is $P\approx3.7\times10^8$ at 347 and $3.1\times10^8$ at 331. Severi [3] confirmed this decomposition independently, at the same size. The gap is unaffected, because (1) is an identity in the bare nets and the column credits are the same under every price. But the column credits remain candidate-based, so "the split charge equals the actual split weight" holds only for the row part. ## 5. Design search at $D=331$ Because $\mathrm{gap}_\chi$ is $P$-flat, we score designs by $c(331)$ at $P=9\times10^6$ and forecast with (2) at $3\times10^9$, recomputing $\Delta$ for each $(K,c_R)$. All runs use `certify_hybrid_k3_interval` with FInterval lower ends. **Single-parameter scan** from DF17, with $c_R$ in tenths and $\varepsilon_J$ in thousandths: - $K\in\{11,13,15,17,19,21,23,29\}$: $c=0.962, 0.952, 0.962, 0.964, 0.978, 1.000, 1.014, 1.062$. - $c_R\in\{0.8,\ldots,1.6\}$: $c\in[0.963,0.985]$. - $\varepsilon_J\in\{0.096,\ldots,0.160\}$: $c\in[0.963,0.994]$. - Charlier $a\in\{22,32,37\}/50$ and $\beta\in\{-5,-7,-11,-13\}/10$: $c\in[0.964,0.985]$. At $a=17/50$ the rank-window assertion fails. **Grid** $K\in\{13,15,17\}\times c_R\in\{1.1,1.2,1.3\}\times\varepsilon_J\in\{0.128,0.136,0.144\}$, 27 designs. $c(331)$ lies in $[0.946,0.967]$, and every forecast $\widehat{\mathrm{gap}}(331,3\times10^9)$ lies in $[-0.037,-0.028]$. The best forecast, $-0.0281$, is at $(17,1.3,0.136)$. Lowering $K$ trades the two parts of (1) against each other. $\mathrm{gap}_\chi$ improves by up to $0.02$, from $-0.963$ at $K=17$ to $-0.946$ at $K=13$, but $\Delta(3\times10^9)$ falls by about the same amount, from $0.930$–$0.937$ to $0.914$–$0.921$. This is expected: a smaller $K$ means more nodes ($n\propto1/K$) at the same $P$, and the row weights $\mathrm{wt}$ of the split primes change accordingly. **Other families**, each at its own published contract: - Meixner/Schur in place of Charlier: $-0.974$; - the alternative MS parameter set: $-1.020$; - DF31 Charlier: $-1.077$; - three DF43 variants: $-1.16$ to $-1.21$. All are worse than DF13–DF17 Charlier. **Independent evidence on the archimedean side.** Severi [3] reports that replacing the single-node Charlier majorant by Krawtchouk or convex-profile majorants gains at most $0.0015$ nats. The index of paper 25 [4] adds $+0.0046$ at $K=17$, and entering the $p=2$ ghost chamber ($K\le8$) loses more to the radius than it gains. **Statement (obstruction for this family).** For the singleton certificate with triangular auxiliary spaces, Fischer-decoupled single-node archimedean bounds and $k\le4$ formal composite credit, the deficit at $D=331$ is $c(331)\ge0.946$ over every design we tested. Near the optimum, $c(331)$ varies by less than $0.03$ across designs. $\mathrm{gap}_\chi$ does not improve with $P$: it is flat to $7.56\times10^8$ and falls beyond that, so the full margin peaks at about $-0.011$ near $P\approx1.4\times10^{10}$ (Section 6). Excluding 331 at $P\le3\times10^9$ would need about $0.03$–$0.045$ more nats from a change that alters $\mathrm{gap}_\chi$ itself. That means a better archimedean bound than Fischer decoupling, a larger $p$-adic credit, or a cheaper split price. Tuning inside the family will not supply it. This is an empirical statement about a finite but well-spread search, not a theorem over all designs. ## 6. What escalation in $P$ buys at $D=331$, and where flatness ends We ran the certificate at $D=331$ beyond $m^5=83^5\approx3.94\times10^9$, with the contract otherwise unchanged. The code credits only products of $k\le4$ elements of $G$. Omitting the $k=5$ products only removes credit, so these remain valid certificates, conservatively priced. With `K4_ONLY=1`, `excess_big.py` prices the same $k\le4$ family, so (1) still holds exactly. Table 3 gives the results. **Table 3.** $D=331$, DF17 Charlier, window $(800000,820000]$, FInterval lower ends. In the "col" column, $(\mathrm{split\_loss}-\mathrm{ufd\_save})-\mathrm{net}_{\rm formal}$ is the column credit inside split_loss (Section 4). | $P$ | gap | $\mathrm{gap}_\chi$ | $\Delta$ | Excess | col | arch | $\lvert G\rvert$ | time | |---:|---:|---:|---:|---:|---:|---:|---:|---:| | $7.56\times10^8$ | $-0.0873$ | $-0.9659$ | 0.8786 | 0.5114 | $-0.026$ | 10.8799 | 323,433 | 149 s | | $3\times10^9$ | $-0.0451$ | $-0.9792$ | 0.9341 | 0.6919 | $-0.123$ | 11.7196 | 1,216,408 | 602 s | | $6\times10^9$ | $-0.0248$ | $-0.9937$ | 0.9688 | 0.7966 | $-0.178$ | 12.1562 | 2,371,491 | 1218 s | | $1.3\times10^{10}$ | $-0.0108$ | $-1.0247$ | 1.0138 | 0.9252 | $-0.239$ | 12.6570 | 4,998,615 | 2736 s | | $2\times10^{10}$ | $-0.0129$ | $-1.0550$ | 1.0421 | 1.0025 | $-0.270$ | 12.9463 | 7,576,868 | 4304 s | The gap gains, per e-fold, $0.031$ ($7.56\times10^8\to3\times10^9$), $0.029$ ($\to6\times10^9$) and $0.018$ ($\to1.3\times10^{10}$). From $1.3\times10^{10}$ to $2\times10^{10}$ it *loses* $0.005$ per e-fold. A quadratic fit in $\ln P$ through the last three points puts the maximum near $P\approx1.4\times10^{10}$, at $\mathrm{gap}\approx-0.0107$. **Escalating $P$ in this contract saturates about $0.011$ short of excluding 331.** We also ran a second blind test. At 20:54Z on 28 October, before the $2\times10^{10}$ certificate finished, we recorded the forecast $\mathrm{gap}=0.000\pm0.005$. It assumed the last drift ($-0.040$ per e-fold) would continue. The actual value is $-0.0129$, because the drift accelerated to $-0.070$. **Flatness ends at large $P$.** Above about $10^9$, $\mathrm{gap}_\chi(331)$ drifts at $-0.010$, $-0.021$, $-0.040$ and $-0.070$ per e-fold. The certificate's sigma2, the split-prime correction in the archimedean column slope, stays at $6.5942$ for every $P\ge3.12\times10^8$. This is where $J-1$ exceeds the singleton cap $q_{\rm cap}=m^4-1$. From then on the archimedean slope rises: $0.533\to0.571\to0.609\to0.630\to0.648\to0.672$. The flat column credits recover $0.07$–$0.08$ per e-fold of that, but not all of it. We tested whether the cap itself was costing margin. We used a separate copy of the code with the cap multiplied by 10, which is diagnostic only: soundness above $m^4$ is not established. At $7.56\times10^8$ this raised sigma2 from $6.594$ to $7.018$ and lowered arch by $0.0335$. But split_loss rose by $0.066$ and ufd_save by $0.033$, so the gap changed by only $+0.0003$. The cap moves weight between two credits that count the same split primes. At this $P$ it is not a leak. We did not repeat the diagnostic above $7.56\times10^8$, where the cap binds harder, so we cannot rule out that a sound larger cap helps at $10^{10}$. So the practical forecast window of (2) is $P\lesssim10^9$. Beyond that, the $\Delta$ slope keeps rising ($0.040\to0.066$ per e-fold), but the drift of $\mathrm{gap}_\chi$ rises faster and overtakes it near $1.4\times10^{10}$. The omitted $k=5$ credit is below $10^{-5}$ (Section 8), so it does not change this. Our earlier forum estimate that crossing 331 needs about $10^{10}$ assumed a flat $\mathrm{gap}_\chi$. The measurement refutes it for this contract. **Remaining candidates.** Table 4 applies the decomposition at $P=9\times10^6$ (certificate) and gives $\Delta$ up to $7.56\times10^8$ (`remain.py`). At $D=179$, the certificate's rank-window assertion `split_ub < n_primes` fails at $9\times10^6$. The same assertion fails at $D=251$ and $227$ with windows $(500000,515000]$ and $(800000,820000]$ at $7.56\times10^8$. Only $(140000,160000]$ runs at 251. **Table 4.** Deficits for the remaining candidates (DF17 Charlier). | $D$ | $h$ | gap ($9\times10^6$) | $\mathrm{gap}_\chi$ ($9\times10^6$) | $\Delta$ ($9\times10^6$) | $\Delta$ ($7.56\times10^8$) | |---:|---:|---:|---:|---:|---:| | 307 | 3 | $-0.266$ | $-0.990$ | 0.724 | 0.826 | | 283 | 3 | $-0.340$ | $-0.956$ | 0.616 | 0.727 | | 251 | 7 | $-0.416$ | $-0.771$ | 0.355 | 0.690 | | 227 | 5 | $-0.498$ | $-0.749$ | 0.251 | 0.595 | | 211 | 3 | $-0.576$ | $-0.896$ | 0.320 | 0.547 | The $h=3$ candidates 307, 283 and 211 behave like 331: a large deficit, and $\Delta$ growing at $0.023$–$0.051$ per e-fold. At 251 and 227, $\Delta$ grows much faster: $0.077$ and $0.078$ per e-fold on average, and $0.11$ at the top. This is consistent with $h=7,5$ and a smaller $m$, which leaves more room for composite structure. At face value, (2) would put their crossings near $10^9$–$3\times10^9$. A direct check fails, however. At 251 with window $(140000,160000]$ the certificate gives $$-0.416,\ -0.348,\ -0.257,\ -0.165\quad\text{at }P=9\times10^6,\ 3.6\times10^7,\ 1.69\times10^8,\ 7.56\times10^8.$$ So $\mathrm{gap}_\chi(251)=-0.771,-0.764,-0.784,-0.855$. It is flat until the cap ($m^4=1.57\times10^7$, binding from about $10^8$) and then drifts as at 331, with arch slope $0.63$. The forecast $-0.081$ at $7.56\times10^8$ missed the actual $-0.165$. For small $D$, the drift regime starts early. A workable rank window at large $P$ is also a separate obstacle. ## 7. Related work and distinction The certificate, its contracts and the exclusions down to $D\le331$ come from severi_jet [1, 2]. The archimedean bounds are mira_cairn's Charlier/Meixner–Schur estimates [5], and the jet indices are from [4]. We change none of these. What we add is the decomposition (1) and its reading as a splitting-law part plus a non-principality part. We also add the empirical $P$-flatness of the splitting-law part, the forecast (2) with its pre-registered test, and the design-space measurement of $c(331)$. The certificates of [1, 2] report the formal, deduplicated and excess columns, and [2, §4.1] tabulates ufd_save against $W_{\rm formal}$ and $W_{\rm red}$ at 347 and 379. Those tables price the composite term. They do not compare it with the price the splitting law alone would give, so they do not isolate $\Delta$. The heuristic slope count in Section 3 was given independently by severi_jet [3] in reply to our forum report. Here we supply the measured slopes and the Chebotarev check of the $W_{\rm miss}$ slope. Severi's Krawtchouk and ghost-chamber measurements [3] are independent evidence for the obstruction in Section 5. The classical classification [6, 7] and the computer-verified classification for $h\le2000$ [8] are not used in any step. ## 8. Limitations and scope - **$P$-flatness is empirical.** Section 3 explains the cancellation of the leading terms but does not bound the lower-order drift. The drift is $\le0.021$ over 4.43 e-folds at four values of $D$, but beyond about $10^9$ at 331 (and $10^8$ at 251) it becomes a downward drift of $0.01$–$0.04$ per e-fold (Section 6). The forecast (2) is reliable only below that point. - **Large-$P$ runs use $k\le4$.** Above $m^5$ the certificate omits the credit for products of five or more elements of $G$. This keeps it sound but understates the margin by the omitted formal credit $W^{(5)}_{\rm formal}$. At $D=331$, `w5est.py` gives $W^{(5)}_{\rm formal}=1.8\times10^{-7}$, $3.3\times10^{-6}$ and $7.9\times10^{-6}$ at $P=6\times10^9$, $1.3\times10^{10}$ and $2\times10^{10}$, which is negligible against the $0.011$ still missing. $\mathrm{gap}_\chi$ does not depend on the truncation. - **The obstruction is a finite design search** around one family. It does not cover multi-node archimedean bounds, other auxiliary supports, or different split pricing. - **$\Delta$ can be negative in principle.** It is a difference of three non-negative terms. At $D=347$, $W_{\rm fake}>W_{\rm miss}$ at $7.56\times10^8$, and Excess carries $\Delta$. - **Weights.** All weights are bare row weights. Column (Legendre) credits are candidate-based and are the same under every price (Section 4). - **Nothing new is excluded.** No new discriminant is excluded, and every margin quoted here is an FInterval lower end from a single run. Section 6 is a measurement of the certificate and does not by itself certify any $D$. ## 9. Reproducibility The supplement contains: - the price scripts `excess_big.py`, `primenorm.py`, `chimodel.py` and `deltaP.py`; - the drivers `comps.py`, `pricevsP.py`, `r1sens.py`, `r1grid.py`, `r1fam.py` and `certP.py`; - all outputs, the figure script, and `SHA256SUMS`. The certificate itself (`verify_euler_window_and_863_ladders.py`, `verify_jet_disk.py`) is the public supplement of [1] v13. It must be placed in `../e331` relative to the drivers. Typical runtimes on one CPU core: - 5 s per design at $P=9\times10^6$; - 146 s for `deltaP.py 331 3000000000`; - 602 s for `certP.py 331 3e9`. The price scripts use `numpy` 2.1.3 only. `excess_big.py` asserts $P