A euchre simulator with rule-based, parameterized bots, built to answer one question: where does competitive advantage actually come from in euchre?
Want the answer as something you can actually use at a table, not as
simulation code? See STRATEGY.md — a practical bidding
cheat sheet and the one card-play rule that matters most, distilled from the
findings below and validated to win ~98% of games against a typical
undisciplined opponent.
euchre/
cards.py Card model, bower-aware trump/led-suit comparisons
rules.py legal plays, trick winner, hand-strength heuristic, scoring
players.py HeuristicPlayer + Strategy (tunable thresholds/flags)
engine.py one hand / one full game (deal -> bid -> play -> score)
simulate.py batch runners (run_hands, run_games) used by the CLI/experiments
cli.py quick head-to-head simulator from the command line
experiments/ scripts that isolate one variable at a time and report the effect
tests/ rule-correctness unit tests (bowers, follow-suit, scoring, ...)
python3 -m pytest tests/ -q # verify the rules engine
python3 cli.py --games 2000 # Team A (seats 0,2) vs Team B (seats 1,3)
python3 cli.py --team-a-alone 18 --team-b-alone 24
python3 experiments/going_alone_threshold.py
cli.py --help lists every tunable knob (order/call/alone thresholds,
--no-signal-a/b to disable partner-aware play, --no-stick-the-dealer).
The strategy is no longer just invented and tuned in a vacuum — it's checked against a published euchre strategy guide (euchre.cards, specifically its bidding, going-alone, and common-mistakes pages), and every deviation from that source is documented in code comments and below.
hand_strength(hand, trump) (euchre/rules.py) sums a per-card point count
anchored to that guide's published trick-win-probability estimates (right
bower = 1.0 trick, left bower = 1.0, trump ace = 0.8, trump king = 0.5,
off-suit ace = 0.7, off-suit king = 0.3 — scaled x20 here to stay integer),
plus a trump-length bonus:
| 9 | 10 | Q | K | A | left bower | right bower | |
|---|---|---|---|---|---|---|---|
| trump | 2 | 4 | 8 | 10 | 16 | 18 | 20 |
| off-suit | 0 | 1 | 3 | 6 | 14 | — | — |
This corrected a real, previously-unsourced assumption in this repo. The
old model asserted "the weakest trump always outranks the strongest
off-suit card" and had a test enforcing it. The sourced numbers say
otherwise: a lone off-suit ace (14) is a genuinely better standalone card
than a bare trump king (10), queen (8), 10, or 9 — only trump ace and the
two bowers outrank it. Trump still dominates in aggregate, but through
TRUMP_LENGTH_BONUS (+6 for a 3rd trump card, +14 for a 4th, +24 for a 5th)
rather than every individual trump card outscoring every off-suit card —
matching the guide's own framing, which gates bidding on trump count
("three trump including a bower"), not a flat point sum. Two off-guide
adjustments, both commented in rules.py: the left bower is scored 18 vs
the right bower's 20 (the source ties them at 1.0 trick each, which is only
true if you hold both — the earlier per-tier monotonicity fix, 9 < 10 < J < Q < K < A with no collapsed ties, is preserved); and queen/10/9 values
(not given by the source) are interpolated to decay smoothly within each
tier.
Per the guide: ordering up hands the dealer — whoever that is — a free
upgrade (pick up the turned card, discard the worst). If you're on the
dealer's own team, that upgrade helps you; if you're an opponent, you're
strengthening the hand you have to beat. So the same hand should be bid
differently depending on where you sit relative to the dealer this hand,
regardless of which "team" you're on. Strategy.position_offset() adds an
offset to order_threshold/call_threshold by bidding position:
first seat (opponent, bids first) +6 (strictest), second seat (dealer's own
partner) −6 (loosest), third seat (opponent) +2. Verified empirically this
is actually worth playing: with everything else held equal, the
position-aware side wins 52.0% to 48.0% against a flat-threshold
opponent (experiments/ ad hoc check, 6,000 games).
The dealer's post-pickup discard (pickup_and_discard) now: (1) never
discards a trump card while any off-suit card is available; (2) never breaks
a guarded ace or king (an off-suit ace/king is exempt from the void-building
discount below, regardless of suit length); (3) among the remaining weak
filler cards, prefers trimming a suit you hold 2+ of down toward a singleton
over sacrificing a card that's already a singleton — sourced directly from
the guide's example: "discard a ten from a two-card suit rather than a nine
from a singleton, despite the ten's higher rank," because ending up with two
near-void suits creates more future ruffing opportunities than fully voiding
one suit while keeping an intact (but still weak) doubleton in another.
_choose_lead now takes is_maker_team into account. The guide is blunt
about this: "leading trump when the opponents called trump...almost always
hurts your team." So: the maker's team leads trump aggressively with 3+
trump (draws out defenders' trump, protects the maker's winners). The
defending team essentially never leads trump — the sole exception is leading
the right bower itself, which wins outright and needs no protection —
leading an off-suit ace instead to bank a trick before trump gets drawn.
A Strategy bundles all the tunable pieces — the three thresholds, the
three position offsets, and signal_aware (duck instead of overtaking your
own partner's winning card, validated separately in Finding 4 below).
Sweeping one knob while holding the rest fixed is what makes the experiments
below causally interpretable. The defaults (order_threshold=45, call_threshold=45, alone_threshold=60, offsets (6, -6, 2)) aren't
guessed — they're empirical win-rate-maximizing / self-consistent values
found by the same sweep methodology used in the findings below (see "Where
the thresholds come from").
hand_strength() requires exact arithmetic ("51") that nobody actually does
mid-hand. simple_hand_count() (Strategy(use_simple_count=True)) rounds
the same anchors to single digits — 3 per bower, 2 per ace, 1 per king or a
trump queen, 0 for everything else, plus a length bonus for extra trump
(+1/+3/+5 for a 3rd/4th/5th trump card) — so it's countable at the table in
a few seconds. Bidding decisions can use either evaluator; discard/lead
mechanics always use the precise one, since those are pattern-following
("don't break a guarded ace"), not arithmetic under time pressure. See
STRATEGY.md for the resulting cheat sheet and Finding 6
below for what the rounding costs.
All numbers below come from the scripts in experiments/ (re-run them to
reproduce; each prints its own sample size and settings).
1. Dealer-side advantage is the single biggest effect in the game — and correct play makes it bigger
experiments/position_advantage.py — 40,000 hands, identical strategy in
all four seats, so any gap is structural, not skill.
| points/hand | |
|---|---|
| Dealer's team | 0.950 |
| Non-dealer team | 0.518 |
The dealer's team scores ~1.8x the points of the other team, on identical skill. The surprising part: this used to be driven roughly evenly by "the dealer's team wins the bid somewhat more" plus "they make bigger when they do." Once bidding became position-aware (Finding-that-isn't-numbered above), the dealer's team's maker share jumped to 71.9% (was 56.8% under flat thresholds) — because a well-calibrated opponent now recognizes that ordering up mostly just hands the dealer a free upgrade, and should pass far more often. First and third seat (the two opponent positions) now only end up as maker 14.3% and 13.8% of hands combined 28.1%, vs. dealer 35.1% and dealer's partner 36.8%. Playing correctly as an opponent means bidding tighter, which means bidding less, which means ceding the hand to the dealer side even more often than a naively symmetric bot would. The one offsetting effect: opponent hands that do clear the higher bar are stronger on average, so their per-call success rate improves (points/hand ratio actually narrowed from ~2.2x to ~1.8x even as the frequency gap widened) — but not enough to close the gap. This is where most of the built-in, skill-independent edge in euchre lives, and no amount of good defensive-side bidding discipline removes it — it can only be partially compensated for.
experiments/going_alone_threshold.py — sweeping Team A's alone-threshold
against a fixed baseline opponent (alone_threshold=60), 4,000 games/point:
| alone_threshold | win rate | loner share | loner euchred |
|---|---|---|---|
| 34 (very loose) | 32.3% | 99.1% | 32.4% |
| 46 | 40.0% | 78.3% | 25.0% |
| 58–66 (optimum) | ~49.2–50.3% | 13–30% | 1.7–6.3% |
| 80 (very strict) | 46.4% | 1.9% | 0.0% |
| never | 45.6% | 0% | — |
In point-count terms, the tier the guide calls "near-guaranteed" (both bowers
- trump ace + two off-suit aces = 18+20+16+14+14 = 82) is far above our empirical optimum (~60) — our threshold sits closer to the guide's Tier 2 ("strong candidate") hands, e.g. right bower + trump ace/king + two off-suit aces (20+16+10+14+14=74, though that's still above 60 — a hand like both bowers + trump king + one off-suit ace, 18+20+10+14=62, lands right at it). The curve is still asymmetric exactly as before: calling alone too often is far more costly (32.3% win rate at the loose end) than never calling alone at all (45.6%). Err conservative, not aggressive.
experiments/trump_calling_threshold.py — same sweep design, 4,000 games/point,
against a fixed baseline opponent (order/call=45):
| threshold | win rate | maker share | euchred as maker |
|---|---|---|---|
| 18 (call almost anything) | 20.0% | 88.6% | 46.6% |
| 32 | 32.7% | 71.9% | 36.3% |
| 43–47 (optimum) | ~48.6–50.9% | 47–53% | 16–20% |
| 60 | 39.6% | 29.6% | 12.8% |
| 78 (call almost nothing) | 22.4% | 21.8% | 21.2% |
Calling trump too often is still the single fastest way to lose (worse than
never going alone, worse than sloppy card play) — being maker 88.6% of the
time sounds dominant but you're euchred nearly half of those hands. Being
too tight is also a loser, and interestingly for a different reason than
before: at the very tight end (78) the euchre-as-maker rate climbs back up
to 21.2%, because the only hands that still clear such a strict bar skew
toward long-but-low trump suits (the length bonus can push a mediocre hand
over a high bar) without the individual high cards needed to actually run
the suit — a reminder that hand_strength is a heuristic, not a perfect
oracle, and extreme settings expose its blind spots.
experiments/partner_signaling.py — 6,000 games, identical bidding
strategy for both teams (order/call=45, alone=60); the only difference is
whether a player ducks (plays low) when its own partner is already winning
the trick, vs. always playing the cheapest card that wins even if that means
beating its own partner's card:
| win rate | points/hand | euchred as maker | |
|---|---|---|---|
| Signal-aware (ducks for partner) | 54.9% | 0.758 | 17.0% |
| Naive (always takes the trick) | 45.1% | 0.705 | 20.3% |
A >9-point win-rate swing from one simple, cost-free rule ("don't beat your own partner's winning card") — the cheapest lever in the whole simulation: no extra information needed, just don't waste tricks your partner already won. This number has stayed essentially unchanged (54.9–55.2%) across every revision of the card-valuation and bidding model in this project, because it doesn't depend on hand_strength at all — only on trick-order logic.
experiments/discard_and_lead_rules.py — 16,000 games per variant, bidding
strategy identical for everyone (order/call=45, alone=60, signal_aware=True).
Team B is always the fully naive baseline (pre-source discard/lead
behavior); Team A cycles through all four combinations of smart_discard
(guarded-honor / void-building discard) and role_aware_lead (maker leads
trump, defender doesn't):
| variant | win rate | points/hand | euchred as maker |
|---|---|---|---|
| naive (baseline behavior) | 49.5% | 0.730 | 17.3% |
| smart_discard only | 50.4% | 0.735 | 17.4% |
| role_aware_lead only | 51.6% | 0.739 | 17.4% |
| both (fully sourced) | 51.5% | 0.743 | 17.4% |
Both rules help, but neither is large, and role-aware leading (~2.1 points) matters roughly twice as much as the discard rule (~0.9 points) on its own. The two together (~2.0 points) land below the sum of their individual effects (~3.0 points) — mild diminishing returns rather than synergy, likely because both rules mostly matter in overlapping situations (a hand strong enough to matter for the discard choice is often also the hand that determines whether you're leading trump). These are genuinely the smallest effects measured in this project — a reminder that not every piece of "correct" advice from a strategy guide moves the needle by the same amount; bidding discipline and partner-awareness dominate, and card-play micro-optimization is a distant third and fourth.
6. A human-computable strategy loses ~4 points to a perfect calculator — and still wins 98% against a casual opponent
experiments/simple_count_validation.py — 16,000 games each. Same overall
strategy (position-aware bidding, signal-aware play, smart discard, role-aware
leads) played two ways: with simple_hand_count() (the single-digit,
countable-at-the-table version) vs. with the precise hand_strength(), and
separately, simple_hand_count() vs. a "casual" opponent (loose flat
threshold, no seat awareness, no partner-awareness, no discard/lead
discipline — a stand-in for a typical unstudied player):
| matchup | simple-count win rate |
|---|---|
vs. precise hand_strength() (exact arithmetic) |
48.1% |
| vs. casual/undisciplined opponent | 97.5% |
Rounding to numbers a human can actually count still costs something —
about 4 win-rate points against an opponent computing exact values. That
gap used to be ~13 points; experiments/simple_count_upgrade.py (16,000
games) shows where most of it went: the original simple_hand_count()
scored a trump queen as 0, the same as a bare 9, and grew its trump-length
bonus linearly (+1/+2/+3 for a 3rd/4th/5th trump card) where the precise
model's bonus is convex. Counting the trump queen like a king and
steepening the length bonus to +1/+3/+5 — both still trivial to track at
the table, no new categories to memorize — took the isolated matchup from
43.7% to 47.4% on its own. A more "principled"-looking alternative (bump
every trump card a full tier above its off-suit counterpart: bower/A/K/Q ->
4/3/2/1) was also tested and performed worse (46.7%) — a reminder that
"looks more accurate" and "wins more" aren't the same thing; it added a
category to track without buying anything empirically.
That remaining ~4-point gap is almost never the comparison that matters in
practice: no one plays exact hand_strength() arithmetic at a real table.
Against the kind of opponent you're actually likely to face, applying
just the two dominant levers (a disciplined, seat-aware bidding count,
and never overtaking your own partner's winning card) — with everything
else rounded to single digits — wins essentially every game. See
STRATEGY.md for this as an actual usable cheat sheet.
Ranked by how much win-rate is on the table: dealer position (structural, you can't opt out of it — and correct opponent bidding discipline makes the frequency gap worse even as it narrows the per-hand quality gap) > bidding-threshold discipline (~30-point win-rate spread between good and bad calibration, steep penalty in both directions, for related but distinct reasons at each extreme) > going-alone calibration (~18-point spread, asymmetric — bad to overuse, only mildly bad to underuse) > partner-aware card play (~9-point spread, but "free" — no new information needed) > position-aware bidding (~4-point spread, worth doing but not dominant) > role-aware leading (~2 points) > guarded-honor/void-building discard (~1 point). The two rankings that might surprise: correct defensive bidding amplifies the dealer's edge rather than closing it, and the two card-play refinements sourced most directly from "common mistakes" advice (discard shape, lead selection) are the smallest levers in the whole simulation — most of euchre's skill gap lives in bidding discipline and not wasting your partner's winning tricks, not in fine card-play technique. Finding 6 is the practical payoff of that ranking: because the two dominant levers are cheap to apply exactly (a seat-aware bidding count, a single "don't overtake partner" rule) while the smaller levers are the ones that actually need precise arithmetic to squeeze out, you can round everything to single digits and still capture the overwhelming majority of the available edge against a real opponent.
The defaults in Strategy (order/call=45, alone=60, position offsets
(6, -6, 2)) are empirical best-response / self-consistent values: for each
knob, one side was swept against a fixed opponent until win rate peaked,
then the opponent was moved to that peak and re-swept to confirm the peak
didn't move (a rough fixed-point check, not a full game-theoretic solve).
The position-offset magnitudes were separately checked against smaller
(3,-3,1) and larger (10,-10,4 and 14,-14,6) variants — (6,-6,2) beat all of
them. See the sweeps above, or rerun with different starting baselines to
re-derive them from scratch.
- euchre.cards/strategy/bidding/ — seat-relative bidding minimums, trick-value point system
- euchre.cards/strategy/going-alone/ — loner hand tiers, score-based aggression
- euchre.cards/strategy/ — leading strategy (maker vs. defender), duck-for-partner
- euchre.cards/strategy/common-mistakes/ — guarded-honor discard, void-building discard example