AI FIELD NOTES · ENDER
Behind the AIEnder had an opponent model. It just never got to use it.
Ender was supposed to be the one who notices your plan, finds the dependency underneath it, and removes that dependency with the smallest possible move.
It was also, for a while, a very diligent way to choose the shared evaluator’s move almost every time.
The diagnosis came from an Ender-only research reset. The old aggregate metric diluted Ender changes across a gauntlet, omitted several named matchups, treated many capped games through a score-margin lens, and used a zero-regret personality rail so tight that the specialized policy almost never changed the shared choice. The right conclusion was not “Ender is nearly optimal.” It was “we have barely exercised the thing we are trying to evaluate.”
The non-promotion that made the next design possible
The first responsible move was to freeze the other AIs and give Ender a proper duel contract: complete official maps, crossed homes and starters, exact paired schedules, replay and fingerprint checks, and match points only for rules-resolved games.
That immediately produced bad news, which was excellent news. The initial 504-duel incumbent baseline had 242 turn caps: 48.0% of games. Those were not wins. They were unresolved diagnostics. Ender’s finished-game macro match-point rate against the mandatory opponents was 0.481, and a shallow screen also exposed a grotesque home-conditioned bias.
Once the winner-safe and rotation-fair boundary existed, several plausible Ender mechanisms had to accept their verdicts. Hand-authored command tools diverged in 10.7–39.3% of games and still lost strength. A public trajectory model diverged in 58.9% of games, lost 2.68pp, and collapsed on Simple. A larger personality story is not compensation for a negative paired result.
That is the first Ender insight: a behaviorally active policy can still be wrong. Activation is a prerequisite for measurement, not a substitute for it.
Alternatives inspected
7Passed inherited verifier
0Changed the move
0Changed the move
2This is an activation check, not an outcome study. It showed that “more careful” had become “never act.”
Research ledger
A visible policy is not yet a stronger policy.
- 01V21: every gate at once0 changes
Its public-command hypothesis selected no move in the sanity trace.
- 02V22: two bounded proof lanes2 changes
The same bounded policy selected a different move twice in the 60-turn trace.
The second problem: a great evaluator cannot choose an absent move
The next breakthrough was architectural rather than mystical. We had been asking a shared evaluator to rank a bounded candidate list while quietly assuming that the list was representative. Playtest-driven Petra work had already made that assumption suspect.
For Ender, the fix was to search different public futures, not just score the same enumeration prefix harder. One deduplicated census interleaves several deterministic ordering lenses:
- immediate height conversion;
- height setup;
- capture denial;
- neutral-tempo closure;
- protected breach;
- flat reach;
- compact leverage; and
- legacy coverage.
The live duel experiment adaptively uses one to eight lenses. Quiet or safely winning positions use fewer. Contested height, deficits, late closure, and decisive endings use more. This is not eight JavaScript threads, eight hidden opponent models, or eight separate games of search. It is one public-state, deterministic candidate portfolio with up to 6,400 candidate evaluations before deduplication.
What the first result says—and refuses to say
The 336-pair discovery screen measured +0.30 percentage points in match points, with a clustered 95% interval of −1.79pp to +2.68pp. Mean margin rose by 3.125. The Bean slice rose +4.17pp. Bigger regressed −7.14pp.
That is not a hillclimb promotion. It is a promisingly shaped question.
| Reading | What the result supports | What it does not support |
|---|---|---|
| overall paired effect | the estimate is compatible with a small improvement | “Ender is stronger” |
| Bean slice | an adversarial matchup worth retesting | reliable domination of Bean |
| Bigger regression | a concrete map warning | averaging the warning away |
| adaptive breadth | a compute-scaling hypothesis | a free performance gain |
The honest visual for this result has a bar, raw cluster dots, and a whisker that crosses zero. A triumphant arrow would be a chart crime.
The next policy: a small opening in the competence rail
The non-promotion led directly to a narrower live experiment: keep the shared evaluator as Ender’s competence rail, then permit a deviation only when an already-enumerated public successor has either a complete reply-backed comparison or a compact public threat proof. The player-facing account of that two-lane checkpoint is The best move is not always the clever move.
This research framing still gives us a sharper failure taxonomy:
- zero opportunities means the board class may be rare or the observation is wrong;
- opportunity but no activation means the gate is too strict or the comparator is inert;
- activation but no reply survival means the policy is telling a one-turn story;
- survival but no conversion means it mispriced medium-horizon opportunity cost;
- positive discovery, negative held-out means the mechanism needs a boundary, not a press release.
What would change our mind
For this to become an actual Ender strength claim, the next screen must use untouched seeds across official duel maps, both homes, both starters, and the frozen opponent roster. It needs rules-resolved endings, matched fingerprints, replay verification, a meaningful match-point lift with a lower confidence bound above zero, no unacceptable map or opponent regression, and a compute cost that a human game can tolerate.
Until then, the correct story is much better than “Ender got smarter.” It is this:
We learned why the prior Ender policy had no agency, built a bounded way for it to notice distinct public futures, and found a small, uncertain signal with one very visible warning label.
That is how a game AI earns the right to become clever.
The extracted table routine is The best move is the move you can prove.