Player advice against the search

The tunable search (expectimax-v2) plans its own turn as a small tree, plays the other seats with a fixed builder, scores leaves with hand-written terms, places setup with a planner, and prices trades by each partner's threat. Nearly all of its evidence comes from paired games against builders that never race, block, or answer a message, so it rewards ore and grain and cannot see the value of denial or negotiation. The status column is a reading of the code and the server's docs/expectimax.md, not a measurement: captured means the mechanism exists, rediscoverable that the search or a weight sweep would find it, gap that the policy cannot express it, and doubtful that an arena test could falsify the folk wisdom.

Tips

Tip	Source	Status for the AI	What would test it
Pips on 6, 8, 5 and 9 beat five-resource coverage; Bo Peng's "69,3" pair takes 6 and 9 numbers with three resources.	Bo Peng, Colonist	Rediscoverable: the planner trades `production` against `coverage` and `balance`, but the defaults were never swept.	Direction 5.
Cover ore and grain across the pair, let the second settlement fund the first build, and point setup roads at open ground or a port, not a site an opponent takes first.	Colonist, Keven Luan	Captured by `balance`, `tempo`, and roads chosen after the remaining picks are predicted; whether `tempo` prices the second settlement's starting grant is unverified.	Ablate `tempo`; count setup roads that lead nowhere.
Take a 2:1 port only beside a strong matching hex; a two-hex port corner has to beat an inland pair.	aka pastor guy, Hexagamers	Captured by the planner's `ports` term and the leaf's `port_match`.	`opening_weights.ports` at 0, 1 and 2.
Spread the pair over several numbers and prefer 5 and 9 so the robber has a less tempting target; never let one number carry a city.	UltraBoardGames, Hexagamers	Doubtful: nothing prices robber exposure at setup; the modelled builders rob by pips times points, so concentration is punished only at depth.	Direction 5.
Block the resource the leader needs to finish, not the hex they produce most from.	King Panda Games, Colonist	Gap: `robber_candidates` in `search.rs` ranks hexes by blocked pips times public points; need never enters.	Direction 1.
Never rob a hex you produce on; aim to be a resource's sole producer.	aka pastor guy	Over-captured: the candidate filter drops every hex touching an own building, so a shared 8 under the leader's city is never considered.	Direction 1.
Rob the biggest hand or whoever holds the card you need; card tracking decides.	Catan Vanguard, Colonist	Captured in part: a hand-size bonus picks victims and thefts branch over hands sampled within the knowledge bounds.	Victim by chance of holding the plan's missing card versus hand size.
Use the robber as leverage: no-block and no-steal deals before placing.	Colonist, Fjord Labs	Gap: no conditional offer exists, and builders would not answer one.	Direction 7.
Play a knight before rolling when blocked; keep one knight in hand.	aka pastor guy, Keven Luan	Rediscoverable: pre-roll plays are root candidates and the roll branches exactly; a held knight is worth only the flat `development` term.	Direction 8.
Almost never trade with the leader or anyone one card from winning; ask what the partner builds next.	Colonist, strategygame.org, Everything Is A Game	Captured: threat premium, the 0.75 cutoff, priced partner gain; builders neither race nor retaliate, so the ablations measured nothing.	Cutoff 0.75 versus none in tables with two `v2` seats.
Trade before you build, and stop naming what you need near the finish.	Hexagamers, strategygame.org, Everything Is A Game	Gap: every `v2` offer names the build it completes, and `v2` counterparties price explicit requests as eagerness.	Direction 6.
Race two sites at once so a rival can win only one, and build roads that cut off opponents' expansion.	Treeckosaurus, UltraBoardGames	Gap: `expansion` values only own sites and `opponent_production` only current income; nothing prices the site an opponent loses.	Direction 4.
Build a city before buying cards; buy a card when nothing else is affordable to shrink the hand.	Keven Luan, aka pastor guy	Rediscoverable: `plan` compares a city with a card and `hand_risk` rewards spending; the `development` slider has not been swept.	Direction 8.
Claim awards late, count opponents' knights, save Road Building for the final snipe; awards are worth more when nobody else contests them.	Alex Cates, aka pastor guy	Gap for the army: `road_contention` exists but no army race term, and `knight` pays for every knight played. The 61% and 50% finishing rates are correlational.	Direction 3.
Count every seat's points including hidden potential, hide your own, keep two routes to the last point, break a road with a settlement.	aka pastor guy, strategygame.org, Minotaur Illusionist	Partly captured: hidden cards are sampled, a road break lowers `opponent_max_point`, wins inside the horizon are terminal; depth 2 sees two own turns and the modelled opponents never race back.	Direction 2.
Keep the hand at seven or fewer; spend before ending the turn.	Fjord Labs, strategygame.org	Captured by `hand_risk` and discard enumeration.	`hand_risk` at 0 versus default should lose.
Five settlements plus Longest Road stalls at seven points; winners hold cities and usually an award (2 of 251 logged games were won without a city).	Bo Peng, Alex Cates, Colonist	Captured by `plan` and `points`; `road_length` could still over-invest in roads.	Builds per seat from arena output.

Three claims seen only in search snippets from settlersboard.com, which refused the fetch, could not be checked: opening pairs should total at least 10 pips, brick is the bottleneck resource in tournament data, and development cards should wait until 7 points. The same snippet recommended a Year of Plenty plus Road Building finish in one turn, which the base rules forbid. Noah Miller's Medium list, the Catan wiki robber page, the BoardGameGeek port thread, and King of Catan were also inaccessible; nothing above rests on them.

Directions for autonomous agents

Each direction names one hypothesis, the mechanism in crates/expectimax/src/v2, the paired contrast, and the result that counts as a gap closed. Unless stated, the candidate or control seat plays with eta, fast, eta on 64 deterministic boards in four rotations, and a gap is closed at +0.04 wins per game with a 95% interval excluding zero, confirmed on fresh seeds. Directions 4, 6 and 7 need two or more v2 seats, since builders cannot respond to what they test. Keys marked new do not exist yet.

Rob what the leader needs. Hypothesis: choosing the robber hex by the most threatening seat's build deficit, including hexes this seat also touches, wins more than blocked pips times points. Mechanism: replace the robber_candidates prefilter in search.rs with a score weighting each hex's resource by that seat's need under its ResourceKnowledge bounds, keep own-shared hexes so the leaf pays the lost production, and pick the victim by the sampled chance of holding the plan's missing card (new robber: "need"). Contrast: v2:{"depth":2,"robber":"need"} versus v2:{"depth":2}, then with two v2:{"depth":2} opponents. Closed: the win contrast plus a longer leader time to next build.
Search deeper when someone can win. Hypothesis: the search at 8 or 9 points loses races because two own turns is too short and its modelled opponents never race back; deepening once any seat holds 7 public points, or a term for the chance an opponent wins first, reduces those losses. Mechanism: new endgame_depth in mod.rs and the deepening loop of search.rs; leaf term opponent_win_risk from sampled hidden cards and affordable builds. Contrast: v2:{"depth":2,"endgame_depth":4} versus v2:{"depth":2}, with explain: true for the losing decisions. Closed: fewer games lost from 8 or more points and the win contrast; rerun a null with two v2 seats before dismissing it.
Model the army race. Hypothesis: an army contention term (knights needed to take or hold the award against the largest opponent count, discounted by cards left in the deck) beats a flat bonus per knight played. Mechanism: army_contention in weights.rs and leaf.rs, mirroring road_contention. Contrast: v2:{"depth":2,"weights":{"army_contention":8}} versus v2:{"depth":2}, then a variant that discounts an award until a set turn. Closed: more games ending with an award held and the win contrast, which also tests the Colonist finishing rates causally.
Price denial. Hypothesis: valuing the site or lane an opponent loses when this seat settles, roads, or races two sites at once improves wins against seats that race. Mechanism: leaf term denial equal to the drop in the most threatening seat's best reachable site, reusing the space threat from bargain.rs; road candidates add roads toward two contested sites. Contrast: v2:{"depth":2,"weights":{"denial":12}} versus v2:{"depth":2} with two v2 seats and two builders. Closed: a positive contrast on mixed tables and no loss on builder tables; a gain only against builders would mean the term is a production proxy.
Sweep the opening. Hypothesis: pips on 6, 8, 5 and 9 over five-resource coverage, and a penalty for pairs whose production sits on one number, beat the default weights. Mechanism: opening_weights exists; add exposure (the largest share of the pair's pips on one hex) to opening.rs. Contrast: v2:{"depth":2,"opening_weights":{"coverage":0,"balance":0,"specialize":1}} and v2:{"depth":2,"opening_weights":{"exposure":0.3}} versus v2:{"depth":2}, so only setup differs. Closed: a setting that wins on the confirmation cohort; since the 47% against 38% planner result shows the opening currency matters, a null means the defaults already match the advice.
Say less when trading. Hypothesis: offers that omit the build they complete, and no explicit requests from 7 points on, lower the counterparty's price and win more against seats that read messages. Mechanism: new speech: "terse" in bargain.rs that drops the reason from offers and declines; counter pricing already reads requests as eagerness. Contrast: v2:{"depth":2,"speech":"terse"} versus v2:{"depth":2} with two v2:{"depth":2} opponents and one builder. Closed: the contrast, with counters accepted at a better ratio; on builder tables the contrast must be zero.
Conditional robber offers. Hypothesis: an offer of the form "one ore and the robber goes elsewhere" extracts cards that plain swaps cannot once a counterparty can price the avoided loss. Mechanism: a proposal type in bargain.rs carrying a robber commitment the proposer honours at its next robber decision, priced by the partner's expected loss from blocked production and theft. Contrast: v2:{"depth":2,"leverage":true} versus v2:{"depth":2} in tables of four v2 seats. Closed: more cards per robber event and the win contrast. This is the largest build, so it comes last.
Time the cards. Hypothesis: the development slider at 0.5 buys cards too early; cities first and one knight held for pre-roll unblocking win more. Mechanism: none for the sweep; then a hold_knight bonus in leaf.rs for one unplayed knight while the robber can reach an own hex. Contrast: v2:{"depth":2,"inclinations":{"development":0.2}} and 0.8 versus v2:{"depth":2}; then v2:{"depth":2,"weights":{"hold_knight":6}} versus default. Closed: a slider setting that wins on the confirmation cohort, and more pre-roll knight plays when blocked without fewer awards.