The script behind runs/jax-ppo-match/altair_vs_borealis.json: duplicate,
seat-swapped play between the two models across single-deal and three-round
conditions, reporting win rate with a Wilson interval and mean margin. Kept so the
2x2 result is reproducible.
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01XBQKgvBbxbheiTF1AVy1Sh
Three-round match is now the default. The bigger fix is that round transitions
were invisible: the engine advances rounds atomically, so a round-ending move
flashed straight to the next round's board and its scores were never shown.
The client now holds at each round end (matchStep sets roundComplete instead of
rolling over; matchAdvanceRound continues on a button). While held, the finished
board stays up and an overlay shows that round's per-expedition breakdown and the
running match total -- no legal move, the rival waits, and nothing advances until
you choose to. The last round ends the match rather than pausing.
The end panel now summarises the whole match in one table -- each round's total
and the sum -- built from a new roundHistory on the match state. A single deal
still shows its per-expedition breakdown as before.
All UI text is Korean now: 승리 / 패배 / 무승부, the prompts, the menu, the score
plaques, both score tables. Removed the explanatory noise that was cluttering the
menu (the "seed shuffles the deck" and "match is decided on the total" blurbs, the
policy/provider line). Verified end to end in a browser: default is three rounds,
each round end pauses on its colour scores, and the final panel shows the
three-round summary.
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01XBQKgvBbxbheiTF1AVy1Sh
Three changes to how much the client helps you:
- The hint feature is gone -- button, board highlights, the whole path that asked
the policy for your best move. Playing against the model shouldn't come with the
model telling you what to do.
- Undo/redo is off by default and enabled from a menu toggle. Taking moves back is
a training aid, not how the game is played, so the honest game is the default.
The switch is reactive: flipping it on mid-game shows the controls immediately,
no restart.
- When on, undo/redo is limited to the current turn. The floor is the most recent
committed move -- a rival move or your own draw -- so you can revise a card
selection or placement before you draw, but you can't rewind into the rival's
move or an earlier turn. Verified in a browser: undo is disabled at turn start,
enabled after selecting a card, and disabled again once you've drawn and the
rival has answered.
Board and Card lose their now-dead hint props and the hint CSS goes with them.
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01XBQKgvBbxbheiTF1AVy1Sh
The table is a control surface, but the browser treats it as a document, so
dragging across it painted card numbers and labels blue and popped a copy menu on
a mobile long-press. Selection is now off across .app-shell and turned back on
only for input/textarea -- in practice the seed field, the one place you would
actually want to select and copy text.
user-select touches neither click nor focus, so every button and card still
responds. Verified in a browser: dragging across the hand and the score plaque
selects nothing, while the seed field still selects and copies.
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01XBQKgvBbxbheiTF1AVy1Sh
Both hands overlap, and the centered value is buried on every card but the
topmost one, so a tightly packed hand became unreadable. Cards overlap
left-over-right, so the one strip that stays visible is each card's left edge --
which is exactly where the top corner already sat. The markup had the corner index
all along; it just wasn't carrying its weight.
So the top-left corner gets a small ivory backing chip and a size bump, which lets
it read even when it falls over the card behind it -- the same reason a fanned
hand of playing cards stays readable. The rival's revealed cards had their corner
index hidden entirely (mini cards set it to display:none); it's re-enabled, scaled
to the 44px card, and their overlap eased from -17 to -14px to give it room.
Verified in a browser at 900/1180/1400px: an eight-card hand overlapping at 900px
now reads every card off its corner, and the rival's revealed cards read off
theirs.
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01XBQKgvBbxbheiTF1AVy1Sh
The trainer was learning a game nobody plays. Lost Cities is decided on three
rounds summed; the stack optimised a single deal. This branch builds the real
game, trains an agent on it, and ships it.
borealis beats altair -- the league policy that was on the web -- 0.6094 at the
real three-round game over 8192 duplicate matches (+22.0 points), at a matched
compute budget. It is also harder to farm: a from-scratch exploiter funded to the
same 131M reaches 0.6295 against altair and only 0.4657 against borealis.
What actually earned that, measured by ablation at a matched budget rather than
asserted:
both-seat training +17.1 points. Self-play already played the opponent's
plies with the same network and the old trainer
stop_gradiented them away. Free doubling of the data.
in-scan auto-reset 5.8x the learner actions per update at identical compute:
87% of every rollout was spent stepping already-done envs.
Exposed a latent zero-bootstrap in compute_gae.
linear total reward Beats tanh(total/scale) 0.5859. The two share an objective;
the linear one just pays it out every ply instead of once
per ~160, and risk attitude turns out to be worth under a
win-rate point in this game -- a policy told to gamble when
behind *loses* to a greedy clone.
match observation ~2.5 points. Mostly single-round defects: the observation
never said whose turn it was, and divided the score
difference by 780.
privileged critic +3.7 points, but only at scale -- ablated at 39.3M it
measured *negative*. An ablation run at a budget you do not
intend to ship can invert.
Also: rounds are independent (corr 0.004), so carry earns its place in the
observation through the start-player rule, not the objective.
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01XBQKgvBbxbheiTF1AVy1Sh
The web client now plays borealis (data/models.json), trained on the three-round
match and taking the 501-dim match view rather than a bare round. Only the actor
trunk is exported -- the critic exists to grade moves in training and never plays,
so the graph physically cannot leak the opponent's hand or the deck, which beats
promising not to call it.
The TypeScript match layer and observation mirror match.py and match_obs.py. They
have to agree to the bit: a mismatch throws nowhere, the ONNX policy just consumes
a wrong vector and plays worse for reasons nobody can see. So the port is not
trusted -- generate_match_parity_fixture.py emits 282 positions from real JAX play
(mid-round, both seats, past a roll-over, with a live carry) and the TS output is
checked against them to float32 round-off.
Match mode is a menu toggle. A seed fixes all three deals and the coin flips, so a
match stays a pure function of it. One-deal mode is unchanged from the player's
side; borealis simply sees it as round one at a carry of zero, a position it has
seen a great many times.
Two bugs found by driving the built app in a browser, both silent:
- The result card totalled the round, not the match. It read "-11 : 3" while the
match stood at -96 : 66 -- it would have named the wrong winner. It now headlines
the match total and breaks the round out beneath it.
- Game records were being rejected. The client's schema went to v2 (it now records
which model played; the old records stored the on-screen label, which stops
identifying anything once there are two models) while serve_web_with_logs.py
still only accepted v1, so every record would have 400'd into a console warning.
v1 stays accepted -- the 111 existing games are altair.
npm test and tsc were green through both. Hence web/.claude/skills/verify, which
records the recipe and the selectors so the next session drives the app instead of
re-deriving how.
.gitignore excluded the new model, which would have shipped a 404: the deploy
builds straight from the repo.
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01XBQKgvBbxbheiTF1AVy1Sh
The 111 games are the user's own strategy record, not ML data -- so they stay as
they are, and hint/undo usage deliberately goes unlogged. What they were missing
is which model they were played against: the v1 schema stores `policy: "WASM ·
FINAL PPO"`, which is the string shown in the corner of the screen, not an
identity. It stops meaning anything the moment there is a second final model,
which there now is.
data/models.json fixes that going forward. Codenames are astronomical and
alphabetical, so the first letter is the generation and you can always tell which
came later; a new letter means the observation space broke, not that the model got
better. A codename never encodes quality -- that is the trap "FINAL PPO" fell into.
The hash stays the truth, because a hash cannot lie about what actually played;
the codename is assigned here, not derived.
altair (e8241e305c01) -- single round, 454-dim, the league policy on the web.
All 111 v1 games were played against this.
borealis (4ae613b010ca) -- three-round match, 501-dim, 131M. Beats altair
0.6094 at the real game and holds up better under a
funded exploiter.
The records were also sitting in runs/tmp/, which is gitignored and documented as
disposable. They are in data/ now, and committed.
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01XBQKgvBbxbheiTF1AVy1Sh
The earlier exploiters ran a quarter of their targets' budget. Rerun at a matched
~131M learner actions (measured: 130.4M and 131.1M):
ours exploiter reaches 0.4657 [0.455, 0.477], mean lead -5.5
league exploiter reaches 0.6295 [0.619, 0.640], mean lead +28.8
The CIs are nowhere near each other and the gap widened under funding, 0.094 ->
0.164. A qualitative line gets crossed too: a properly funded attacker trained
from scratch *beats* league outright, and still loses to ours.
Two things a reader has to carry with that, both of which cut against us:
- Neither exploiter plateaued -- both climb through the final update, so these
remain lower bounds. And ours' attacker has the steeper tail (+0.031 vs +0.016
per 100 updates) while league's curve is already bending, so a much larger
attacker could narrow the gap. Cannot rule out that it closes.
- League is a single-round policy blind to carry and match score, and both frozen
policies are played greedily, which is maximally exploitable by construction.
So this is not a clean "training method A vs B".
The claim that survives is the matched-budget one, and only that one.
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01XBQKgvBbxbheiTF1AVy1Sh
The exploiters ran 32.5M learner actions, not the 65.5M I recorded -- they train
one seat, so a 250x1024 run yields half of what the same shape gives the
both-seat self-play trainer. Against targets trained on 131M and 122.6M, that
makes the attacker roughly 4x underfunded.
Which means the absolute number does not support "ours is only 22.8%
exploitable". It supports exactly one claim: at a matched budget, league gives up
more. Whether the ordering survives a properly funded attacker is now the open
question, so the script takes --updates and --batch-games to run it.
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01XBQKgvBbxbheiTF1AVy1Sh
Winning the head-to-head says a policy is strong on average, not that it is hard
to beat. So freeze each policy, train a fresh one from scratch whose only job is
to beat that policy, and see how far it gets.
Same exploiter budget (250 updates x batch 1024, 65.5M learner actions):
ours (match stack, 131M) exploiter reaches 0.2278 [0.219, 0.237]
league (web-deployed) exploiter reaches 0.3213 [0.311, 0.331]
League gives up 9.4 more points to a dedicated attacker. Both sit far from 0.5, so
neither is a pushover -- but the caveat left open by the head-to-head is now closed
in our favour on both axes: stronger on average (0.6094) and harder to exploit.
Worth noting against expectation: league was trained *with* an exploiter structure
and we ran pure self-play, and we still come out less exploitable. Whatever the
league machinery buys, it did not buy that here.
The number is a lower bound -- a bigger exploiter would find more -- so it only
means anything as a like-for-like comparison, which is how it is used.
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01XBQKgvBbxbheiTF1AVy1Sh
The 39.3M ablation was the critic's worst ground. At a matched 131M budget:
both-seat training is worth +17.1 points of win rate, the privileged critic
+3.7 -- a 4.6x gap. And they behave differently: both-seat is flat across scale
(+16.8 at 39.3M, +17.1 at 131M) while the critic swings (-1.6 to +3.7).
The mechanism explains it. Both-seat doubles the data at the same compute --
self-play already played those plies with the same network. The critic adds no
data, only accuracy, and a privileged value trunk has more to fit, so it grades
badly until there is enough to fit it with.
Fable ranked the critic first among missing ideas and both-seat fourth. That was
exactly backwards.
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01XBQKgvBbxbheiTF1AVy1Sh
The 39.3M ablation said the privileged critic hurt (switching it off won 0.5160
[0.505, 0.527]). Head to head at 131M, both sides trained identically, it says the
opposite: off *loses*, 0.4633 [0.453, 0.474]. Against league the critic-on model
scores 0.6094 and the critic-off one 0.5526.
The critic earns its keep once there is enough data to fit it -- at 39.3M the
privileged value trunk is underfit and only adds advantage noise. Defaulted back
on, with the small-scale number kept in the docstring as a warning: an ablation at
a budget you do not intend to ship can invert.
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01XBQKgvBbxbheiTF1AVy1Sh
Each piece switched off in turn, trained at identical compute, then played against
the full stack over 8192 duplicate matches. Below 0.5 means the removed piece was
doing work.
- both seats: 0.3317 [0.322, 0.342]. The biggest single contributor. Half of it is
simply sample count -- dropping the opponent seat halves the learner actions per
update -- but that is the point: self-play already produced those plies with the
same network, and the old trainer stop_gradiented them away.
- match observation: 0.4751 [0.464, 0.486]. Small but real. Since carry itself
contributes almost nothing (rounds decompose), most of this is likely the
single-round observation defects being fixed: to_move, the deck clock, and the
score_diff scale.
- privileged critic: 0.5160 [0.505, 0.527] -- turning it OFF makes the agent
significantly STRONGER. Fable called this the biggest missing idea; it is wrong.
A critic that knows the deck fits V(full state), which is not
E[return | masked obs], so the advantage picks up a component the actor cannot
act on. From the actor's side that is noise, not variance reduction. Asymmetric
critics hurting under partial observability is a known failure mode.
Defaulted off accordingly. (Reusing it as a PIMC leaf evaluator may still stand --
that is a separate claim from using it to train the policy.)
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01XBQKgvBbxbheiTF1AVy1Sh
The question was what any of this actually improved over the training method
that existed. Duplicate matches, 8192 a piece, same three deals and coins from
both seats:
- At 39.3M learner actions the match stack beats the Phase 0a gate agent
(0.5842) which had 411M -- 10.5x the data. Sample efficiency is the headline.
- At 39.3M it *loses* to the league policy (0.3142). That is a budget gap, not a
strength gap: league had 122.6M plus a league/exploiter structure.
- Scaled to a matched budget (131M vs league's 122.6M) it wins: 0.6094
(CI 0.599-0.620), +22.0 points.
So: same compute, stronger agent, measured on the actual game.
Caveat kept honest in the plan -- league was trained with exploiters, and we have
measured average strength, not exploitability. "Harder to exploit" is not shown.
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01XBQKgvBbxbheiTF1AVy1Sh
The user asked why we were not simply maximising the three-round total, and they
were right. Each ply now pays the points by which it moved the running match
difference; at gamma=1 that telescopes to the final total, so the objective is
exactly the rulebook's, handed out every ply instead of once 150 plies later.
Three measurements killed the tanh design:
- Rounds are independent (corr(m1,m2)=0.004, corr(m1+m2,m3)=0.05), so a reward
linear in the total decomposes the match into three independent rounds and
carry enters the objective nowhere. The only coupling, the start-player rule,
is worth +0.73 +/- 0.84 points -- indistinguishable from zero.
- Risk attitude, the one thing tanh buys, is worthless here. A policy made to
gamble when it trails by 20 entering round three *loses* to a greedy clone over
6144 duplicate matches (0.482); gambling only at -40 breaks even (0.498). A
marginal wager buys about +1.7 sigma for -2 to -3 expected points. Ceiling on
the whole carry-conditioning idea: under one win-rate point.
- Head to head over 10,000 duplicate matches at equal compute, the linear reward
*beats* tanh(total/12): 0.5859 (CI 0.576-0.596), +20.3 points. Dropping it is
not merely free, it is better -- not because of risk, but because tanh hands a
~150-ply match one saturated +/-1 and leaves all credit assignment to the critic.
The flat carry probe was not exploration collapse: sampled play still opens 5.00
expeditions, entropy settles at 1.36 nats (3.9 effective actions), and the critic
reads carry cleanly (round-three values run -0.87 to +0.86, monotone). The signal
was there; there was nothing to buy with it.
Criterion 1 (a monotone carry response) comes off the gate accordingly -- the
optimal response barely exists in this game. carry stays in the observation: it
costs nothing and the start-player rule keys off it.
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01XBQKgvBbxbheiTF1AVy1Sh
Self-play works: matches converge to 146.8 plies (~49 a round) with a 91% deck-
race rate, so the stalling that static opponents induced is gone. Duplicate match
eval scores 0.4968 with a mean lead of exactly 0.0 -- same deals, same coins, both
seats, deal luck cancelling exactly.
Success criterion 1 does not pass. The carry probe is close to flat: expeditions
opened sit at 5.00 whether the policy is 60 points down or 60 points up. Wager use
does move monotonically across all six carry levels, and in the right direction
(behind -> more multipliers), but the spread is 0.31 wagers.
Two diagnoses, one of which was mine and wrong:
- Residual potential shaping was NOT the cause. Annealing it fully to zero left
the probe just as flat.
- terminal_scale is. At carry -60, tanh((margin - 60)/50) is close to linear over
any realistic round margin, and maximising E[tanh] on a linear stretch is just
maximising E[margin] -- there is no reason to gamble. Risk-seeking only appears
where tanh is sharply convex, which needs a smaller scale. Dropping 50 -> 12
widens the wager spread 0.19 -> 0.31, which is the mechanism showing up.
The probe itself is also mis-scaled: at scale 12, tanh(60/12) is 1.0, so +/-60 is
a saturated dead zone with no gradient and the policy has learned nothing there.
The measurable band is |carry| <~ 2 * terminal_scale, and the probe levels have to
be set from the scale rather than fixed.
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01XBQKgvBbxbheiTF1AVy1Sh
Phases 2-4 land together because they all rewrite the same rollout, and doing
them in sequence would mean writing it three times.
Observation adds the four things a match policy cannot play without: carry as a
scalar *and* a binned one-hot (round three is a threshold problem -- "win by 41
or lose" plays nothing like "win by 39" -- and the old score_diff divided by
MAX_ABS_SCORE=780, squashing a decisive 50-point lead to 0.06); the round index;
whose turn it is, which the single-round observation never carried even though
the critic is trained on opponent-turn states; and the deck clock, since a round
ends on the last deck draw and players bend that parity by drawing from discard
piles. Live points per colour are split by hand / discard pile / unseen, because
a discard pile is public and recoverable.
The critic is asymmetric: it gets the opponent's hand and the deck in order, on
a separate trunk so none of it can reach the logits. A test pins that down --
perturbing the privileged input leaves the policy logits bit-identical while
moving the value. Deal luck is what makes a match-terminal reward hard to learn
from, and a state-value baseline may condition on anything action-independent.
Both seats now train. Self-play ran one network on both sides and stop_gradiented
the opponent, throwing away half of every game; each ply now emits a transition
per seat, folded into the batch so each seat keeps an independent GAE chain.
Reward is the match: rounds one and two only bank into carry, and round three
pays tanh(total / terminal_scale). Potential shaping on the running total covers
the early sparsity and anneals out.
match_eval adds the two measurements the plan turns on: duplicate match play
(same deals and coins from both seats -- self-play scores exactly 0.500 with zero
mean lead, so the mirroring cancels deal luck exactly) and the carry probe. On an
untrained net the probe is flat: 4.97 expeditions opened at a 60-point deficit
and at a 60-point lead alike. Breaking that flat line is success criterion 1.
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01XBQKgvBbxbheiTF1AVy1Sh
The cap item existed so a three-round match would fit inside a 400-step scan.
Phase 0a's truncation bootstrap removed that constraint -- scan length is a free
parameter now -- and the same phase showed the cap is the thing that turns
stalled games into losses: 20.6% of cap-hit games lost, 99.4% of all losses. A
120-ply cap would put *more* games into the wall and make "freeze the round
while ahead" easier, so the round cap stays at 400 and max_steps_rate stays a
watched metric.
Also records what the plan never listed: ppo.py, gates.py and league.py all take
a single-round State and still have to be wired to MatchState. That is the
largest remaining piece of Phase 1.
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01XBQKgvBbxbheiTF1AVy1Sh
Wraps the single-round engine rather than changing it: engine.py is the rules
oracle the TypeScript client is differential-tested against, and the round
itself is unchanged by the match. Only two rules live up here, both from the
Kosmos rulebook -- three rounds decided on the summed total, and "the player who
has more points begins" the next one. That is not alternating, and it is not
what reset_from_order hardcoded, so it takes a first_player argument now.
The rulebook says nothing about an exact tie, so the starter falls back to a
coin flip. A deterministic tie-break would give one seat a standing edge in
symmetric self-play and the agent would learn to steer for it. Round one needs
no special case: carry is (0, 0) there, so the tie branch already yields the
coin, which is exactly the rulebook's arbitrary "oldest player begins".
All randomness -- three deals and three coins -- is drawn in match_reset and
stored in the state, so match_step stays deterministic and needs no PRNG key
threaded through every rollout, eval, and gate body. It also makes a mirrored
match (same deals, seats swapped, same coins) a pure seat relabel, which the
antithetic pairing later depends on.
Tests cover the deck clock (44 deck draws a round, discard draws extend it),
carry banking each round exactly once, the start-player rule across all three
branches, a fair round-one coin, mirror symmetry, and that the running total
does not jump across a round boundary -- the last one matters because potential
shaping will be built on it.
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01XBQKgvBbxbheiTF1AVy1Sh
At equal compute (250 updates either way, 5.8x the samples) the new loop beats
the baseline against heuristic_expert: 0.4525 -> 0.5071 win rate, -4.28 ->
-0.56 mean score diff. Expert ends rounds at their natural length, so that is
a clean read on card play.
It appears to *lose* ground against heuristic_balanced (0.9866 -> 0.9292), but
that is entirely the MAX_STEPS=400 cap. Drawing from a discard pile does not
deplete the deck, so a round can be stalled indefinitely; against a weak
opponent the extra turns are worth points, and with 5.8x the samples the agent
learns the exploit harder (greedy rounds run 225 plies). Splitting 2048 games:
naturally-ended games lose 0.08%, cap-hit games lose 20.62%, and 99.4% of all
losses land in cap-hit games. The board freezes mid-expedition and the -20s
stand.
Self-play does not have this failure mode -- league runs converge to 53.6-ply
rounds and never hit the cap -- which is the regime the 3-round work targets.
Noted in the plan: static heuristic anchors are unusable as gates, and the
per-round cap needs care in Phase 1.
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01XBQKgvBbxbheiTF1AVy1Sh
rollout_steps was pinned to MAX_STEPS (400) while a round actually runs
~50 plies, and finished envs were only reset between updates. 83% of every
rollout was spent stepping already-done envs to produce masked-out zeros.
Measured at rollout_steps=400: active steps go 17.3% -> 100%, i.e. 5.8x the
learner actions per update for the same compute.
Resetting in-scan exposes three things that were previously benign:
- compute_gae bootstrapped truncated episodes from zero. That was safe only
because every episode used to terminate inside the scan; now episodes cross
the boundary, so thread V(s_T) through.
- rollout_metrics read final_env and summed rewards along the scan axis, both
of which assume one episode per slot. With several episodes per slot that
silently produces garbage, so aggregate at done boundaries instead.
- league assignments were redrawn only between updates, which would pin a slot
to one seat/opponent across every episode in a scan. Redraw them on reset.
Also anneal potential shaping against learner actions rather than padded scan
steps: the old accounting counted the dead steps, so a 5M-step anneal expired
within two updates of 250. Any earlier evidence that shaping does not help was
gathered with it effectively off.
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01XBQKgvBbxbheiTF1AVy1Sh
The rival's policy now exposes a ranking, which drives both its own move and a
new HINT button: it highlights the card to play, where to put it, and where to
draw, with the model's confidence.
Undo and redo work per action rather than per turn — picking a card, choosing
its destination, and drawing are separate steps, as are the rival's moves — so a
finished game can be stepped back through from the result screen. The rival is
suspended while undone moves are pending, and PLAY FROM HERE resumes from the
reviewed position.
Cards now travel between zones instead of teleporting: a motion layer measures
each card's old and new position and animates the difference, flying cards out
of the deck face-down and flipping them over, and back into it on undo.
Also: the result screen gets a per-expedition score breakdown mirroring
human_play.py, and a rival card revealed by a discard-pile draw no longer
renders at full size in the card-back-sized rival hand row.
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01LmyprzuzanXRhpomc3Ga1i
Every game is now dealt from a seed carried in the URL as `?seed=`, shown in
the menu, and re-dealable by typing it in, so a deal can be shared or replayed.
Seeds are hashed into a mulberry32 stream, independent of the Python shuffle
bank.
Cards previously teleported between zones: the only motion in the client was
the hover lift and the legal-target pulse. Cards are keyed by card id, so a
card that just moved into a zone mounts there and now animates in, with the
motion disabled under prefers-reduced-motion.
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01LmyprzuzanXRhpomc3Ga1i
Placing a card locked the turn in: the chosen card left the hand and no
affordance reverted the placement, forcing the move through. Clicking the
chosen destination again or pressing Escape now steps the selection back.
A failing ONNX inference left the AI's turn unadvanced, permanently
stalling the game. The AI turn now falls back to the heuristic policy.
Other fixes: the score plaque no longer covers hand cards (plaques become
compact chips at narrow widths and hand spacing tracks the viewport),
opponent cards drawn from a discard pile render face up, long expedition
stacks stay inside their lane, undo no longer bumps the generation counter
on empty history, and small viewports scroll instead of clipping.
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01LmyprzuzanXRhpomc3Ga1i
ismcts-bc-ceiling-2026-05-11 was filed under "Other"; promote it to a
proper SO-ISMCTS section between Deep CFR and Engine/Performance so the
catalog actually reflects the project's two algorithm families. As more
ISMCTS notes accrue they have an obvious home.
coolguyandClaude Opus 4.8