Cell p3 s5 b4

grid order 3, 5 stages, cycle bucket 4

The rk run ended on 2026-11-10, and no more compute goes to it. Epoch 2 is frozen and these pages are its final snapshot: the site is no longer rebuilt. The overview site sums up the final findings, and the harness is public for anyone to download and run.

This page is the full archive record for this cell's elite. Every number here was produced by the pinned scoring code, and verifier_hash pins which version of it; see the explicit archive for where this cell sits in the grids. The tier and phase label are assigned mechanically.

tier
heldout_verified
phase label
search result
tableau_hash
5f506b0793cbe188c5b1f55d97210b48cbdd2898f94404ea044b2d80d3d0dec6
verifier_hash
2db0816c621c44aa5b4b8efbadfde3a37497780460be138cd8af5439ab176611
cycle_id
312
seed
31105
directive_id
D-0308
hypothesis_id
H-008
recorded
2026-09-23 14:22 CT (stored 2026-09-23T19:22:29Z UTC)

Tableau

A[.][0]A[.][1]A[.][2]A[.][3]A[.][4]
c[0] = 0/10/10/10/10/10/1
c[1] = 2763/81922763/81920/10/10/10/1
c[2] = 23623/1638410451/3276836795/327680/10/10/1
c[3] = 28281/32768-9587/32768759/40967949/81920/10/1
c[4] = -27773/32768-8373/32768-4269/32768-23543/327682103/81920/1
b112384864660023175569177103/67508629812447792062201856-34300118171452581016771705/6750862981244779206220185622834076463857529636431/61082726938515917537280-5664005650924659057253/30541363469257958768640-5653/16384

All values are exact fractions; the Q15 integrator applies each one as a dyadic m/2^s pair, listed below. The tableau layout is defined in the glossary.

Raw coefficient representation (m/2^s)

The Q15 integrator applies each nonzero A and b entry to a state value v as (v * m) >> s, an arithmetic right shift that floors. Zero entries are skipped. When exact is no, m/2^s only approximates the fraction and the largest such gap is the record's coeff_quant_error. CSD weight is the fewest signed power-of-two terms that write m; it prices the advisory avr_approx model, and the two M0+ models price this coefficient from its entry in the coefficient price table.

entryexact valuemsm/2^sexactcsd weight
A[1][0]2763/81922763132763/2^13yes7
A[2][0]10451/32768104511510451/2^15yes7
A[2][1]36795/327689199139199/2^13no4
A[3][0]-9587/32768-958715-9587/2^15yes7
A[3][1]759/409675912759/2^12yes4
A[3][2]7949/81927949137949/2^13yes5
A[4][0]-8373/32768-837315-8373/2^15yes6
A[4][1]-4269/32768-426915-4269/2^15yes6
A[4][2]-23543/32768-2354315-23543/2^15yes5
A[4][3]2103/81922103132103/2^13yes4
b[0]112384864660023175569177103/67508629812447792062201856272751427275/2^14no8
b[1]-34300118171452581016771705/67508629812447792062201856-1664915-16649/2^15no4
b[2]22834076463857529636431/61082726938515917537280244991624499/2^16no6
b[3]-5664005650924659057253/30541363469257958768640-607715-6077/2^15no5
b[4]-5653/16384-565314-5653/2^14yes6

Per problem

Final-state error of this tableau on each problem, integrated in Q15 under m0plus_fast at the fixed cycle budget. Bar length is error on a log scale; the printed value is exact. dahlquist, damped_osc and vanderpol_mild are the search set the optimizer sees; the other four are the held-out set that decides archive fitness, and on several problems floor rounding dominates the method choice, so bars can look similar across very different tableaus. The share column gives each problem's part of its own set's sum of squares, because both aggregates are a root mean square: the three search rows share one total and the four held-out rows share another, and a share from one set never adds to a share from the other. Each set adds to 100 before rounding, so a set printed to two decimals can read 99.99 or 100.01. A problem that finished with no error carries none of its set's total and reads 0. A set holding an overflowed run has no usable total, and neither has a set whose errors are all zero, so in both cases the shares read n/a.
1e-41e-31e-21e-1dahlquistdahlquist: 0.000167470.00016747damped_oscdamped_osc: 0.00210240.0021024dc_motordc_motor: 0.01575960.0157596pendulumpendulum: 0.0032720.003272quaternionquaternion: 0.01268030.0126803rc_thermalrc_thermal: 0.0459560.045956vanderpol_mildvanderpol_mild: 0.004753320.00475332
Every plotted bar, in one table (7 rows)
problemsetfinal-state errorshare of its set, percent
dahlquistsearch0.000167470.10
damped_oscsearch0.002102416.35
dc_motorheld-out0.01575969.81
pendulumheld-out0.0032720.42
quaternionheld-out0.01268036.35
rc_thermalheld-out0.04595683.42
vanderpol_mildsearch0.0047533283.55

Cycle counts (n_states = 1)

modelcyclesnote
m0plus_fast156
m0plus_slow311
avr_approx380avr_approx is an approximate model, shown for context only.

Score

metricvalue
measured_order3.02822
order_fit_points6
error_constant1.03662
stability_real-2.02822
stability_imag (bisection floor)1.53587
csd_weight_total84
coeff_quant_error3.05176e-05
search_error0.00300234
heldout_error0.0251585
overflow_margin1.7508

stability_imag is the smallest imaginary-axis interval the scorer's search can resolve, from 4000 samples and 200 bisection steps, so it is a floor rather than a measured extent: a method whose stability region meets the imaginary axis only at the origin reports that floor and not 0. Read a value at the floor as no resolvable interval. stability_real carries no such floor, because the real-axis interval of these methods is well inside the sampled range.

How to read this

measured_order is the slope of a log-log fit of float64 final-state error against step size on the dahlquist problem, over the longest usable run of points (order_fit_points of them); it is measured evidence, distinct from the algebraic order that keys the grid. error_constant is the L2 norm of the order-condition residuals one order past the achieved one, a size estimate for the leading truncation term. stability_real and stability_imag are read off the stability region by the bisection described under the table above, along the negative real axis and along the imaginary axis.

csd_weight_total sums CSD weights over the non-trivial coefficients, a proxy for coefficient-arithmetic cost. coeff_quant_error is the largest gap between an exact coefficient and its m/2^s form. search_error and heldout_error are root-mean-square errors over the two problem sets at the fixed cycle budget; heldout_error is the archive fitness. overflow_margin is 1 / max|state| observed at twice the nominal amplitude and must exceed 1.0, meaning a doubled signal still fits in Q15 range.

Every per-problem error, by cost model

problemsetm0plus_fastm0plus_fast share, percentm0plus_slowm0plus_slow share, percentavr_approxavr_approx share, percent
dahlquistsearch0.000167470.100.000167470.00380.000167470.00095
damped_oscsearch0.002102416.350.004756483.070.008372372.37
vanderpol_mildsearch0.0047533283.550.026723696.930.053768597.63
pendulumheld-out0.0032720.42infn/ainfn/a
dc_motorheld-out0.01575969.810.00853681n/a0.00731512n/a
rc_thermalheld-out0.04595683.420.0264906n/a0.0203956n/a
quaternionheld-out0.01268036.350.0135059n/a0.0180309n/a
search_errorsearch0.00300234set total0.0156717set total0.0314175set total
heldout_errorheld-out0.0251585set totalinfset totalinfset total

Each cost model is scored by its own integration, so it has its own per-problem errors and its own two sums, and one share column would describe one of these three while mislabeling the other two. The set column says which total a row belongs to: within one cost model the search rows divide one sum of squares and the held-out rows divide another, and a share is the row's error squared over the sum for its own set. Each set adds to 100 before rounding, so a set printed to two decimals can read 99.99 or 100.01, and a share below a tenth of a percent prints to two significant figures instead so that a small contribution does not read as none. The two rows in bold are those sets' own root mean square values rather than members of a set, which is why they carry the words set total in place of a share; under m0plus_fast they are the record's own search_error and heldout_error, and the other two models store theirs under their own prefix. A row whose error is 0 carries none of its set's total and reads 0. A set holding an overflowed run has an infinite error and so has no usable total, and a set whose errors are all zero has nothing to divide by, so in both cases every share reads n/a rather than being spread over the rows that finished.