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.
Summarizes the 13-cycle trap-exploration session: BC pretrain (heuristic
clone) is the self-play ceiling under our compute budget (1 GPU + 50
sims + 768x4 MLP). All variants (naive finetune, KL anchor, mirror
descent, mixed-opponent + opponent-aware search) either preserved BC
(~17-21/100 vs heuristic-cautious) or regressed to catastrophic
forgetting. The single largest improvement of the session — 4× win rate
on the same checkpoint — came from PUCT Q-value normalization at search
time, not from any learning change.
Records the mechanism (negative training signal from BC-vs-heuristic
games; search too shallow to find heuristic-beating moves), the
hypotheses we negated, and the dials left in code for future runs with
more compute.
C13/C14 cycles: heuristic-balanced bot plays a configurable fraction of
self-play games (training.mixed_opponent_fraction). Trainee turns are
stored as policy samples; opponent turns are taken by the bot directly
and not stored. When mcts.opponent_aware_search is set, the MCTS tree
also treats the opponent seat as that bot — opponent moves are applied
without expanding into the search tree, and all values are taken from
the traverser's perspective. This was Codex's top recommendation for
breaking the symmetric self-play weak fixed point.
Empirical: opponent-aware mixed self-play does NOT lift win-rate above
BC pretrain (vs heuristic-cautious 100-game eval):
C13 (mixed=0.5, no KL): 0/100 — catastrophic forgetting
C14 (mixed=0.2, KL beta=1): 17/100 — preserved BC, no improvement
BC pretrain baseline: 21/100
Combined with C10-C12 results, BC remains the ceiling under our
compute budget (1 GPU + 50 sims + 768x4 net). Code is left in place as
configurable dials for future runs with more compute.
Codex follow-up diagnostics identified two MCTS+training-loop issues that
together cap finetune-from-BC at the heuristic ceiling:
1. PUCT Q is in raw score units (~±100 for value_scale=100), but the
exploration bonus c_puct * prior * sqrt(N) / (1+n) is on order of 1-10
for our parameter ranges. Result: a single bad backup pushes q_eff
well below the bonus floor and that action is effectively never
visited again. With only 50 sims/move this is catastrophic for the
policy-improvement operator. Fix: divide q_eff by config.q_scale
(default 100, configurable) inside _select_action. Backups and value
targets remain in raw score units; only the selection signal is
normalized. AlphaZero canonical convention.
2. The current kl_anchor_beta path adds KL(current || ref) directly to
the loss. That preserves BC but prevents improvement (gradient
actively pulls policy back to reference). The standard regularized
policy improvement operator is to mix the target instead:
pi_target = softmax(alpha * log(pi_mcts) + (1-alpha) * log(pi_ref))
Anneal alpha from low (rely on BC) to high (rely on MCTS) over
training. Network learns to follow the regularized target, which
stays near BC early but lets MCTS-discovered improvements through
later.
Config additions:
- mcts.q_scale (default 100.0): PUCT Q divisor
- training.md_target_ref_ckpt: reference policy path (alternative to kl_anchor)
- training.md_target_alpha_start / _end / _iters: linear alpha schedule
Both Python mcts.py and Cython mcts.pyx updated; parity test passes.
Tests: 19/19.
Hypothesis: with normalized PUCT the network can actually explore and
exploit prior knowledge competently at 50 sims, and the mirror-descent
target lets self-play improvement happen while BC anchors the trajectory.
This is the operator-side fix that c9 (no anchor, collapsed) and c10/c11
(loss-side KL anchor, preserved-but-stuck) both missed.
Lets c6+ layer new exploration hyperparams on top of c5's learned
value head instead of restarting from random init. Saves ~60min per
cycle while preserving VPE-down trajectory observed in c5.
C4 (768x4 + rollout=True, 150 iter) result: 0/300 natural wins, score
avg -85 to -97 vs three heuristic opponents. Bigger network alone did
not produce wins; PA shifted up to 0.23-0.25 (similar to c3 without
rollout) but agent still loses every natural-end game.
Capacity hypothesis rejected: 2.5x more params (~2M vs ~800k) did not
break the loss pattern. Score average actually slightly worse than
512x3 baseline. So the bottleneck is not network capacity.
Going to the long-run experiment: AlphaZero-correct setup with the
network value loop closed. use_rollout_value=false means leaf Q comes
from network value head. Training signal: network value learns from
actual game outcomes; MCTS uses those values to pick actions; better
actions produce better outcomes; cycle closes.
100 iter previously gave essentially the same result as rollout=true
(comparing c3 to trapfix baseline). Both are too early in the AlphaZero
training curve. Standard AlphaZero papers train 1000s of iterations.
Going long: 1000 iter with the current config. Self-play ~9s/iter
without rollout, total wall ~150 min for the train phase.
Plotting strategy: at iter 200, 500, 1000, run 100-game standalone eval
and generate analyze.py plots to visualize trajectory.
Network kept at 768x4 since bigger capacity does not actively hurt.
coolguyandClaude Opus 4.8