Rank-conditioned empirical study · 14

What 5d, 9d, and professional policies see in the same Go position

Across the same 50 professional games and 500 positions, a 9d Human SL policy covered 82.7% of the AI one-point good set at the median, versus 77.5% for 5d. The difference is real, but it is not a ladder that rises in every position.

1. Give four human-policy observers the identical frozen positions

   The sample fixes 50 professional games and ten middlegame positions per game, for 500 decisions. A strong b11 Transformer teacher supplies candidates at 128 and 512 visits. Human SL then returns rank_5d, rank_9d, era-matched professional, and 2023 professional policies for every position. Because the positions are identical, the comparison measures policy-distribution shifts rather than giving each rank a different test.

2. Do not begin by asking who guessed one exact coordinate

   Candidates retained by the 512-visit teacher and lying within one point of its leader form the good-move set. The set respects a continuous value surface: if six moves are nearly tied, entering five of them is not a failure to see the idea. Rank comparisons therefore use policy mass on the good set while retaining exact-top probability as a secondary measure, instead of turning search noise into one required answer.

3. The 9d policy sees more on average, by about five points

   Median good-set coverage is 77.5% for the 5d policy and 82.7% for 9d. Paired by position, 9d is higher by 5.1 percentage points on average, with a whole-game cluster-bootstrap 95% interval of 4.3–5.9 points. The 9d policy covers more in 76.2% of positions. This is a stable movement of probability mass, not a claim that the higher profile wins every item.

4. A higher rank is not more AI-like in every position

   Nearly one quarter of positions do not give the 9d policy higher coverage, and a small tail gives 5d materially more mass on the current AI-good region. Frequencies of familiar techniques, position type, sparse training strata, or an unstable teacher can all contribute. The defensible claim is that the overall distribution shifts with rank—not that one Human SL probability reads a particular player's mind or certifies a rank.

5. A teaching boundary is more useful than a rank label

   If a move is already common under 5d, review may only need to confirm the problem it solves. A sharp rise from 5d to 9d suggests a near-term learning target. If even 9d and professional policies assign very little mass, first audit budget, model, and symmetry stability before collecting it as a deep move. A product should show this reachability band instead of declaring that every 5d player must know one coordinate.

6. Personalized review needs receipts, not a hidden total score

   Each proposed lesson should retain played-move loss, the one-point candidate set, target Human SL profile, set probability, search budget, budget agreement, and model identity. If only a final label such as “for 5d” survives, a model upgrade becomes indistinguishable from human improvement. Keeping the components visible lets a player or coach explain exactly why a ranking is unhelpful and allows the system to recompute it later.

7. The next experiment must test transfer, not another average

   This study finds a stable rank-conditioned distribution difference; it does not show that reading an explanation improves play. The causal next step randomizes review ordered by maximum point loss against review ordered jointly by loss, reachability, and stability. Days later, structurally similar positions from different games should test checking order and candidate sets. Rank-aware teaching becomes a learning fact only when judgment changes in unfamiliar positions.