The Hollow Opponent — Where Game Depth Actually Lives

Decoding NA1’s daimyo AI to ground truth exposed it as one-ply greedy: every turn it finds the weakest reachable fief and invades if it clears a threshold (plus a ~1-in-100 poke at a sole neighbor even above threshold). No search, no coalition, no feint, no positional sacrifice — no dominance frontier on the AI’s side at all, because it isn’t choosing among trade-offs, it’s comparing one scalar to a line. The disappointment (“you don’t meet your heroes”) is itself the finding: the depth that made NA1 feel strategic for 30 years lived in the player’s ignorance of the rule, not in the opponent. Once the policy is decoded, the depth doesn’t vanish — it migrates to the board.

Links: NA1 — A Game-Design Crucible (the hub this promotes into), Battleship — Best-Response vs Minimax (this is the flat-opponent end of that gradient), The Dominance-Frontier Lens (NA1 supplies its computable random-start-viability instance), Randomness as Termination (N≥3), Capability Without Leverage

A thesis page (portable, source-independent), promoted from a session discussion. The decoded AI lives in the na1-decompiler repo; this page states the general claim.

The decode, stated plainly

province_ai_state → pick the weakest reachable prey → invade if your strength clears the threshold. One scalar, one comparison, a small vestigial random term. It’s the flat end of the best-response-vs-minimax gradient: a flat opponent makes no structure of its own, so it throws all the structure-making back onto the player’s model. For 30 years the human supplied that structure and credited the game for it.

The veil was the game (and that was a legitimate design)

The AI’s policy is hidden-trackable, not hidden-random — deterministic and learnable. Perceived depth held only as long as it stayed untracked. A game that wants to survive its own decode needs either genuine depth (a real search/trade-off surface) or genuine randomness (hidden-random, which can’t collapse). NA1 chose neither — it leaned on player ignorance, which in 1989 was a free and reliable resource. You stopped being a valid customer the day you opened the bank. This is the hidden-trackable vs hidden-random information-design distinction read as a design-lifetime property: ignorance-supplied depth has a shelf life; the other two don’t.

The depth migrated to the board

Kill the opponent’s contribution and the residual game is not empty — it’s a graph-scheduling puzzle wearing a strategy costume:

The adjacency graph. “Don’t be the weakest” is a per-border constraint. Satisfying it across all N neighbors simultaneously — while staying strong enough to pick off fiefs you can then also hold above threshold — is a real combinatorial problem.
The turn-economy. 4 action-slots/year is the binding constraint. Keeping every border safe and spending slots to expand is a scheduling tension, not a given.

Crucially this is internally consistent with the prior decode: the 17-fief tier analysis already found geographic isolation beats raw stats for low-tier survival. That’s literally “fewer neighbors → easier to never be the weakest.” The tier verdict predicted this AI before the AI was read. The decode didn’t contradict the play-experience; it relocated and explained it.

How to actually win (the manual survives the disillusion)

The shallowness makes the strategy crisper, not absent:

Never present as the weakest reachable neighbor to any border. This is the whole defensive game.
Watch the second-weakest flip: if the current weakest spends its turn buying men, you may become the weakest for the rest of the round — usually one-ply-avoidable, but it’s the one lookahead the player must run.
Pick off fiefs only when you can hold the result above threshold too — conquest that drops you below the line just relocates the target onto your new border.
Isolation is an asset (fewer borders to keep safe), tempered by the sole-neighbor random poke.

What this opens