Skip to content

Manaka structural critique

This page keeps the diagnosis, not the full proposal history. The original 2026-08-24 analysis asked why Manaka's policy was weak without search and proposed five repairs. Those repairs were later implemented and measured; the final outcomes are in network history.

Problem observed

The early v2/v3 Manaka line learned the teacher distribution, but policy-only play remained weak. The important evidence was:

  • policy top1 around the mid-30% range on the supervised records;
  • strong improvement once search was added;
  • little gain from simply making another pass over the same corpus.

The old text described this as “bias-limited”. That remains a useful hypothesis, but it is not a standalone measurement: later architecture experiments are the stronger evidence.

Structural constraints in the prototype

1. Pair capacity was compressed too early

The baseline pair table has millions of rows and most of the model's parameters, but a position sums all active rows into only pair_dim channels before the head sees them. Large table capacity therefore does not imply equally large per-position bandwidth.

This concern was partly supported: removing the pair table hurt strength, so the table was useful, but later experiments also explored less-destructive aggregation (lse, ks) rather than only widening it.

2. Symmetric-polynomial interactions are restricted

The original power-sum block is elementwise. Higher orders combine a channel with itself; they do not create arbitrary cross-channel products. Second-moment features were added to test this directly.

Measured result: the tested moment variant did not beat the baseline. Cross-channel interaction was therefore not the first bottleneck at that configuration.

3. Summation discards spatial layout

Native tokens know which square a piece occupies, but after global aggregation the head no longer has a square-indexed representation. The original policy also lacked direct destination context.

This was the strongest diagnosis. The square-field proposal was the only A–E modification that clearly beat the prototype in the 2026-08-25 policy-only round robin, and widening field_dim from 2 to 4 improved further.

4. The prototype head was shallow

The original evaluator had one learned hidden layer after aggregation. A second hidden layer was added to test whether depth and multiplicative conditioning were the missing ingredient.

Measured result: the tested hidden2=256 variant finished below the baseline. More head depth alone did not solve the problem and also costs inference time.

5. The early policy head lost move identity

The v2 policy score depended on a move label and one global hidden vector, so labels could conflate different moving pieces. policy_from was added in v3 to include the moving piece/origin token directly.

This flaw is already fixed in the current architecture. It should not be cited as a current limitation.

What the A–E measurements changed

Policy-only round robin, one-epoch weights, 2026-08-25:

variantapproximate Elo vs setreading
field4+50spatial field helped most
field2+31same direction
baseline+12reference
order2~0third order contributed little in this test
pairtie~−1no clear gain
moment16−8no gain
pair0−35pair table is useful
hidden2=256−49extra depth hurt

These Elo values belong to that exact round robin and those weights. Later validation-only runs use different weights and must not be called contradictions when their ordering differs.

Current interpretation

The strongest supported statement is narrower than the original critique:

The baseline throws away useful spatially addressed information too early; preserving some square-level context helps policy quality.

The data does not support “make the head deeper” as the primary fix, and it does not show that the pair table is useless. The next architecture questions are aggregation shape, feature space, side-separated aggregation, and whether extra policy quality repays its search-speed cost.

For current structure, see architecture. For experiment history and later F–M ideas, see network history.