The learning slopegraph — what changed as the sample grew

the three pattern pages · A inventory B topology C change

Neither an inventory nor a structure but a difference: what the model learned at 64, 1,000 and 10,000 bytes, as three columns with one line per source byte joining its argmax successor at each size. Vertical position is the successor's byte value, so a flat line is a rule that did not change. 175 of the 256 are flat and grey and carry no information; this page exists for the other 81.

64 bytes1,000 bytes10,000 bytesSP→xSP→SP0A→000A→SP>→00>→0Ae→de→SPn→sn→a<→m<→/r→gr→eo→ro→ni→ki→ns→=s→SPc→00c→et→-t→;,→00,→SPh→th→e[→00[→[]→00]→]a→wa→ty→00y→SPd→id→>g→/g→e'→00'→'f→00f→SPl→nl→SPv→00v→eq→00q→u"→SP"→>

The flips, with the counts behind them

Every rule that changed its mind, and what the data actually says at each size. The counts are counts; the stored weight beside them is a log count with a stochastic, uncorrelated error (LSA.md), so the two are not required to agree. Where the rule names a successor the counts do not, the estimate inverted the order of two small counts — the property LSA is designed around, seen from the data side.

source64 — the rulewhat the data says1,000 — the rulewhat the data says10,000 — the rulewhat the data says
0A00 w 0never occursSP w 4SP 19 SP w 7SP 128 0A 6 [ 2 T 2 …
SPx w 1x 2 SP w 6SP 83 < 19 k 12 x 4 …SP w 9SP 542 < 127 a 83 t 77 …
"SP w 1h 1 SP 1 rule ≠ dataSP w 3> 12 SP 5 h 3 0 2 … rule ≠ data> w 4> 27 1 9 p 7 SP 5 …
#00 w 0never occurs00 w 0never occurs9 w 59 14 R 6
&00 w 0never occurs00 w 0never occursa w 5q 24 l 20 g 20 a 14 rule ≠ data
'00 w 0never occurs00 w 0never occurs' w 4' 34 , 7 & 7 SP 6 …
(00 w 0never occurs00 w 0never occurs& w 1T 2 r 1 2 1 1 1 … rule ≠ data
)00 w 0never occurs00 w 0never occursSP w 2. 2 SP 2 | 1 , 1 … rule ≠ data
*00 w 0never occurs00 w 0never occursSP w 1/ 1 SP 1 rule ≠ data
,00 w 0never occurs00 w 0never occursSP w 6SP 62 & 1
.3 w 1o 1 m 1 3 1 rule ≠ datao w 3o 5 3 4 m 3 w 2 …SP w 3SP 15 & 11 o 8 h 8 …
/" w 1x 1 w 1 e 1 / 1 … rule ≠ datan w 4n 11 w 5 / 5 e 4 …n w 5t 22 i 22 n 21 r 17 … rule ≠ data
100 w 0never occurs" w 1" 2 / 1 . 1 5 w 55 8 8 7 9 6 0 6 …
200 w 0never occurs" w 2" 2 0 1 0 w 40 10 - 4 5 3 " 3 …
3/ w 1/ 1 " w 2/ 2 . 2 " 2 rule ≠ data" w 2< 3 . 3 " 3 T 2 … rule ≠ data
400 w 0never occurs" w 1" 1 5 w 45 9 3 5 6 2 2 2 …
500 w 0never occurs" w 1" 1 < w 3; 6 < 5 8 4 7 3 … rule ≠ data
600 w 0never occurs" w 1a 1 " 1 rule ≠ data7 w 21 4 < 2 : 2 8 2 … rule ≠ data
700 w 0never occurs" w 1" 1 ; w 3; 4 0 3 Z 2 6 2 …
800 w 0never occurs" w 1" 1 0 w 39 8 0 5 Z 2 : 2 … rule ≠ data
900 w 0never occurs" w 1" 1 4 w 34 13 8 4 6 4 < 3 …
;00 w 0never occurs00 w 0never occurs[ w 4[ 14 & 14 # 14 r 12 …
<m w 1m 1 / w 4/ 15 n 14 s 2 m 1 …/ w 8/ 96 t 21 n 21 i 21 …
>00 w 0never occurs0A w 50A 19 M 4 W 3 U 2 …0A w 60A 127 A 10 1 9 2 7 …
?00 w 0never occurs00 w 0never occursi w 1i 1
A00 w 0never occurs00 w 0never occursm w 3n 21 m 6 r 3 p 3 … rule ≠ data
B00 w 0never occurs00 w 0never occurse w 2e 3 r 2 a 2
C00 w 0never occurs00 w 0never occursT w 4a 10 o 6 T 6 h 4 … rule ≠ data
D00 w 0never occurs00 w 0never occursI w 2I 6 i 4 e 3 o 2 …
E00 w 0never occurs00 w 0never occursn w 4n 8 D 6 C 6 v 1 …
F00 w 0never occurs00 w 0never occursa w 3a 4 r 3 i 1
G00 w 0never occurs00 w 0never occursr w 2r 2 o 2
H00 w 0never occurs00 w 0never occursi w 3i 5 o 3 e 2 I 2
I00 w 0never occursm w 1m 2 R w 3R 6 n 5 m 3 d 1 …
J00 w 0never occurs00 w 0never occursa w 1s 1 o 1 a 1 rule ≠ data
K00 w 0never occurs00 w 0never occursr w 3r 9
L00 w 0never occursS w 1o 1 S 1 rule ≠ datao w 3o 4 a 4 i 2 e 2 …
M00 w 0never occurse w 2e 4 a 1 L 1 e w 3e 5 u 4 o 1 a 1 …
N00 w 0never occurs00 w 0never occursA w 1o 1 a 1 A 1 rule ≠ data
O00 w 0never occurs00 w 0never occursr w 1r 4
P00 w 0never occursa w 1a 1 o w 3e 5 o 3 r 2 h 2 … rule ≠ data
R00 w 0never occurs00 w 0never occursE w 4E 12 o 6 u 4 SP 4 …
S00 w 0never occursc w 1p 1 c 1 rule ≠ datao w 2t 3 o 3 p 2 a 2 … rule ≠ data
T00 w 0never occursa w 1a 1 SP w 3h 8 SP 6 1 5 r 3 … rule ≠ data
U00 w 0never occurss w 2s 2 s w 2s 2 n 1
V00 w 0never occurs00 w 0never occursa w 1a 1
W00 w 0never occursi w 4i 6 i w 6i 10 h 2 e 2 a 1
X00 w 0never occursM w 1M 1 M w 1M 1
Z00 w 0never occurs00 w 0never occurs< w 2< 7 e 2
[00 w 0never occurs00 w 0never occurs[ w 5[ 55 h 12 A 8 s 5 …
]00 w 0never occurs00 w 0never occurs] w 7] 55 SP 22 , 19 . 7 …
_00 w 0never occursP w 1P 1 2 w 2i 2 2 2 o 1 c 1 … rule ≠ data
aw w 1w 2 m w 5m 27 c 24 l 7 w 4 …t w 8m 95 n 88 r 72 c 59 … rule ≠ data
b00 w 0never occursa w 1a 2 e w 4e 17 u 16 a 11 y 5 …
c00 w 0never occurse w 4e 25 a 3 h 2 i 1 e w 6e 66 o 44 h 41 i 24 …
di w 1i 2 i w 4i 12 " 1 > w 7SP 57 > 40 i 32 e 27 … rule ≠ data
ed w 2d 2 x 1 SP w 5s 26 > 17 SP 13 y 12 … rule ≠ dataSP w 7SP 125 s 100 r 96 > 64 …
f00 w 0never occursi w 1o 1 i 1 rule ≠ dataSP w 7SP 34 & 20 r 13 i 9 …
g/ w 1/ 1 / w 2e 5 / 5 = 1 rule ≠ datae w 6e 35 t 20 SP 13 i 10 …
ht w 1t 1 e w 2t 5 e 2 a 1 rule ≠ datae w 7e 59 i 49 t 28 a 20 …
ik w 2k 2 a 2 . 1 SP 1 a w 5a 13 k 12 p 4 t 3 …n w 7s 80 n 80 d 54 t 52 … rule ≠ data
j00 w 0never occurs00 w 0never occurso w 2o 2 s 1 p 1 l 1
ki w 1i 2 e w 4i 12 e 12 < 5 rule ≠ datai w 6i 26 e 23 < 9 SP 2 …
ln w 2n 2 / 1 / w 2k 5 / 3 n 2 p 1 … rule ≠ dataSP w 5e 30 SP 30 i 24 t 23 … rule ≠ data
ns w 1s 1 a w 5a 27 s 3 e 2 = 2 …a w 6a 101 t 53 SP 47 d 31 …
or w 2r 2 r w 2r 10 n 2 c 1 > 1 n w 7n 81 r 73 t 48 m 44 …
p: w 1o 1 : 1 rule ≠ dataa w 4a 25 e 5 : 5 o 3 …> w 6a 65 o 23 r 20 > 16 … rule ≠ data
q00 w 0never occurs00 w 0never occursu w 3u 28
rg w 1t 1 g 1 rule ≠ data< w 2g 5 t 3 s 2 a 2 … rule ≠ datae w 8e 86 i 60 o 37 c 35 …
s= w 1= 1 p w 6p 25 i 6 e 6 t 2 …SP w 6SP 67 t 64 p 51 e 51 …
t- w 1t 1 p 1 - 1 rule ≠ data- w 3t 6 p 5 a 5 e 4 … rule ≠ data; w 7i 75 h 74 e 65 ; 64 … rule ≠ data
u00 w 0never occurs00 w 0never occurso w 5t 33 s 26 o 26 r 13 … rule ≠ data
v00 w 0never occurse w 1e 1 e w 5e 29 i 21 o 7 a 3 …
w. w 1w 2 i 2 . 1 rule ≠ data. w 2w 8 i 6 . 4 3 1 rule ≠ datai w 4w 24 i 14 . 12 h 6 … rule ≠ data
y00 w 0never occurs= w 3= 12 SP w 6SP 41 = 20 , 6 / 5 …
z00 w 0never occurs00 w 0never occursa w 2a 2 p 1 h 1 e 1 …
{00 w 0never occurs00 w 0never occursR w 2{ 5 R 4 A 1 rule ≠ data
|00 w 0never occurs00 w 0never occursK w 2R 3 K 3 t 1 s 1 … rule ≠ data
}00 w 0never occurs00 w 0never occurs} w 3} 5 < 4 0A 1
A900 w 0never occurs00 w 0never occursr w 1r 1
C300 w 0never occurs00 w 0never occursA9 w 1A9 1

Support against sample size

Each of the 256 k=1 rules' stored support at each size, on one shared axis in LSA units — which are already a log scale, so this is a log axis by construction and not by choice. Read underneath it the number the whole generation turns on: the argmax's share of its context, over every context that occurs.

64 bytesargmax share 0.6351,000 bytesargmax share 0.51510,000 bytesargmax share 0.2900123456789
The share falls: 0.635 at 64, 0.515 at 1,000, 0.290 at 10,000. That is the generation's central measured fact about the k=1 strength, and it decides the axis it is on. A stated constant delivery is right at the small samples, where the argmax carries about half its context's mass and 8 is the weight that says so; at 10,000 bytes the same contexts have recurred with different successors, the share is down to 0.290, and the stored matrix already knows better than any constant (#f-p8-forward). It is the same direction #viz_standard records for the trace rate: a model learned from the bytes it is measured on has memorised a small prefix, and every measure of its sharpness is optimistic there.
One page, not eight, and here is why that is honest. All eight generation-5 variants carry one learned model, byte for byte — axes B, G and H change how a model is read and settled, never how it is learned — and the builder refuses to build unless that identity holds. So there is one thing to draw here. The stored supports are read from wordsv2-v025, the only variant in the generation that serializes the count matrix; a B1 model would leave the second panel empty, which is what the B1 and B2 inventory pages are for.

Published under #viz_standard. Self-contained: no external stylesheet, no font, no image, nothing fetched. Built from the spec in p8v2-pattern-viz-specs.md by docs/pprog/build-p8v2-patterns, never hand-edited. Every pattern drawn was reconstructed through the query layer from (program text, model file); no page reads a de-optimized dump. Nothing here is a result and no page carries a chart point — the picks are the programmer's (#variant_protocol).