[aywrite]

Mobility was fitted once, over 1,812 games. Six of its eight weights rounded
to nothing, so SCORED_KINDS stopped counting three of the four kinds at the
leaf and the term shipped as a rook count. This refits it on 43,675 games with
every other term held, and prices all four.

knight  4 / 1        bishop  6 / 3
rook    4 / 4        queen   1 / 7

+76 ±25 elo, sprt [0, 10] passed at LLR 4.15 against a bound of 2.94,
settled in the first batch of 500. The baseline is the commit below the
refit, which is the same tree with the old weights, so the difference measured
is the eight weights and nothing else.

Three commits, each measurable on its own.

--hold-pawn. The holds were a ladder downward, one per term as it was
fitted. A refit of a term already in the tree has every other term older than
it rather than newer, and the terms above it had no flag. Without this the
sixteen pawn structure weights would have moved beside the eight and the match
could not say which it measured, which is what 1b0862a found the last time a
hold was missing.

A tuner fix. Corpus refused any named sealed pair that reached no row of
the extraction. That conflated a seal drawn against the wrong archive, which
nothing can be read under, with a pair whose positions all went to a
lower-keyed game or were filtered out of the extraction, which is ordinary
once the archive is large. The first is still refused, the second is counted
and reported. No engine code.

The refit. 259 artifacts over 54 runs, 5,355,792 positions, 2,461,322
quiet rows, with the tables, the material, the shelter and the pawn structure
all held.

What it costs, which is the other half of the arm

Every weight is non-zero now, so SCORED_KINDS has nothing to skip and all
four kinds are counted at every leaf again. Callgrind over the bench with
mobility_with forced out of line:

build	term	share of run	nodes
the rook alone, as it shipped	229,040,403	6.00%	4,089,373
four counted, old weights	423,840,322	10.56%	4,089,373
four priced, this vector	528,497,023	13.11%	3,868,987

The middle row is the counting alone, with the tree identical because weights
still at zero cannot move a score. The term is 1.85 times what it was to count
and 2.31 times to count and price. The run goes +5.52% in instructions while
the tree shrinks 5.39%, so a node costs +11.5%.

specs/mobility.md allows a leaf term 5%. Mobility has never met it and this
more than doubles the distance. The games say that is worth paying, which
makes the ceiling the thing under question rather than the term.

What the instruments say

	before	after
bench	4,089,373	3,868,987
strategic suite	96,532	99,938
tactical suite	231	224

Eleven of the strategic suite's fifteen themes rise. Open Files and Diagonals
is up 562, which is the theme this term is most nearly about and the one the
original mobility arm predicted would move.

The tactical suite falls seven. That is the tripwire and it had never moved by
more than one across any earlier evaluation arm. It is not a gate, and the
match was run knowing it; +76 buys it.

The fit

Selection reads -0.000798 ±0.000113 and the sealed group, 4,749 pairs from
seventeen runs nothing had read, reads -0.000510 ±0.000107. Both outside
their intervals, 1.85 standard errors apart, so selection overstates this fit
by about a third; king safety's two agreed to 0.68 and pawn structure's to
0.15.

The ridge is zero and that is not the trap it was for pawn structure: the
vector quantizes identically at zero, 1e-8 and 1e-7, the largest weight is 7,
and every slot carries a coefficient in 37% to 70% of the training rows.

The learning curve is flat from an eighth of the training pairs. The corpus
stopped being this term's constraint some time ago.

Four tests moved that are not node count pins

None was weakened to fit, and each says something about the refit.

A known mate in two is missed at depth four and found at three, where the old
weights miss it at three and find it at four. Whether a depth finds that mate
was never monotone; the hole moved rather than appeared. The test asserted it
at one depth and now asserts the invariant in its own name, that the distance
does not change, across four depths.

Adding late move pruning now costs nodes on three of the bench's eighteen
positions, where under the old weights it cost nodes on none. That test
asserted the skip searches less, which is not a property the skip has, and now
asserts what its comment says it is for: that the switch reaches the search.

The residual replay needs two depths whose answers disagree, and says so in
its own first assertion; five and two came to the same answer and five and
three disagree under both vectors. The reduction ledger needs a run holding
all three scout outcomes, and at depth five these weights leave no fail high
at any rate.

The uci session's node budget moves with the tree, which its comment says.

The finding worth keeping

The first mobility fit read 1,812 games, priced one piece and took +12 ±8 over
4,000 games. This one read 43,675, priced four and took +76 ±25 over 500.
Nothing about the term changed between them.

So SCORED_KINDS was never an optimisation. It was a measurement of how
little the old corpus knew, and it had three quarters of a term switched off
at the leaf for want of evidence rather than want of value.

Rebased onto 6a2c5c7 in a scratch copy before opening this: it applies
clean, the bench is the same 3,868,987 and both suites and both pin sets hold.

🤖 Generated with Claude Code

[github-actions]

Speed against 6a2c5c7e

Both sides built and run on this runner in this job, the way
scripts/speed.sh measures a perf commit. The change is between
medians, and the spread beside it is what the change has to be
read against. The node counts are the search's: they move when
the search does, and a speed change leaves them alone.

round     base nps  candidate nps
    1      8856997        8141379
    2      9237596        8112257
    3      9275078        8297170
    4      9023130        8226741
    5      9044804        8078178
    6      8939516        8169676
    7      8906903        8114792
    8      9049908        8053024
    9      9259789        8150915

             nodes    time  median nps  fastest nps
base       4089373  0.45 s     9044804      9275078
candidate  3868987  0.48 s     8141379      8297170
change       -5.4%   +5.1%      -10.0%       -10.5%

the node counts differ, so this is a search change as well as a speed
one. nps only says what a node costs. Time to depth is what the change is
worth at this depth, and whether the new tree is the right one is for
games to say.

Speed: -10.0% (bench nps, 9 interleaved rounds vs 6a2c5c7e, spread 4.6%)