preview_play omits three game_state keys that score_hand reads — Loyalty Card's cycle freezes

[idIing]

Summary

playground/exact_score.py::preview_play builds a synthetic game_state for the engine's
score_hand, but never sets three keys that score_hand actually reads. Each defaults, so the
preview silently scores some boards differently from the engine that will actually run the hand.

GreedyTactical ranks every candidate play through preview_play
(src/jackhammer/playground/harness.py:195), so this reaches the reference agent's play
selection, not just a diagnostic path.

Verified against main @ c3a24a0, engine pin 4d6f19d.

The three keys

Set difference between what jackdaw/engine/scoring.py reads out of game_state and what
preview_play sets on synth (19 keys set, 19 read, three of them disjoint):

key	read at	consumer	effect when it defaults
hands_played	scoring.py:447 → hands_played_run	Loyalty Card, jokers.py:1214	joker evaluated at run counter 0
skips	scoring.py:458 → ctx.game.skips	Throwback, jokers.py:1868	x = 1 + skips * extra collapses to x1
chips	scoring.py:535,908	Mr. Bones rescue path	round chips treated as 0

Loyalty Card — confirmed empirically

Loyalty Card's window is recomputed from the run-wide hands counter
(Balatro/card.lua:3633, config {Xmult = 4, every = 5} at Balatro/game.lua:393), and Jackdaw
models that faithfully. With hands_played absent the preview always evaluates at counter 0, so
whether it fires depends only on hands_played_at_create % 6 — constant for a whole run, rather
than cycling.

Measured on one real position at the same engine pin: run counter 17, joker created at 7.

• engine: (5-1-(17-7)) % 6 = 0 → does not fire
• preview: (5-1-(0-7)) % 6 = 5 → fires

Preview reported total=440 (mult=4.0) where stepping the engine added 110. Ratio exactly
4.00, reproduced on three of four plays in one hand; the fourth agreed because the engine
genuinely fired there too.

Why skips is the one I would fix first

skips is currently 0 in practice only because agents were never given the blind-skip
decision. Protocol v2 removed the BLIND_SELECT auto-step, which is exactly the fix that makes
skips start incrementing. So this defect is dormant today and arms itself as a direct
consequence of the fidelity fix — on Throwback boards, silently, with no error.

Suggested check

A regression that asserts preview_play's total equals the engine's actual chip delta on a board
holding each of Loyalty Card, Throwback, and Mr. Bones. The existing bit-exactness gate over
recorded plays did not catch this, which suggests those three boards are simply absent from the
corpus — worth confirming rather than assuming.

What is not established here

I have not measured the magnitude of the effect on any published number. This report
establishes that the mechanism exists, that it reaches the reference agent's ranking, and that it
is live on main — not how far it moves the reported deltas. Treat the impact as unaudited.

Provenance

Found by Claude (Opus 5) during offline search experiments against this engine pin. The search
certifies each positive result by replaying its candidate line through the real engine; one line
previewed 5128 and replayed 3070, and tracing that disagreement produced this report. Worth noting
for the fix: a scorer that is never re-checked against the engine cannot surface this class of
bug, which is why a preview-vs-engine regression is the durable form of the check above.

[idIing]

Title corrected, and an independent replication of the fix

Title. It said "Loyalty Card scores x4 in every preview". That is one of two cases, and the
body above already had the general statement right — recording the correction here so the title and
the body agree.

With hands_played absent the preview evaluates at counter 0 permanently, so
remaining = (every-1-(hands_played - hands_played_at_create)) % (every+1) is constant for the
whole run: the cycle freezes rather than shifting. Whether that constant equals every depends
only on hands_played_at_create % 6, so the defect is all-or-nothing per joker instance —
1 residue in 6 fires on every hand, the other 5 fire on none. The instance measured above
(created at 7) is in the always-fires class and over-estimates; the majority class never fires and
under-estimates instead. Reading the title literally sends a reader looking for over-estimates
only.

Replication, engine-checked

Reproduced at engine pin de733eb on a different seed, and settled against the engine rather than
against another preview. Battery train seed GILEUEUU, hands_played_at_create = 3, holding
j_devious, j_loyalty_card, j_raised_fist, j_gift, j_wee:

• real cycle fires at hands_played = 8, 14, 20, …
• pre-fix preview sits at remaining = 1 at every decision of the run and never fires once

At hands_played = 14, the same subsets under both scorers, with ground truth obtained by
checkpointing and actually stepping the engine:

play	pre-fix preview	fixed preview	engine
Pair, 5 cards	50 × 20 = 1,000	50 × 80 = 4,000	4,000
High Card	47 × 7 = 329	47 × 28 = 1,316	—

Exactly ×4 both ways, so the ranking of candidate plays is unaffected — the fix moves only
GreedyTactical's absolute clinch test (best[0] >= need). Chips 0 against a 4,000 target, so
need was exactly 4,000: the fixed preview clinches and clears the blind with that hand, the
unfixed one declines and plays on.

Blast radius, measured

The issue notes this reaches the reference agent's play selection, so it is worth saying how far.
Holding the engine pin and the episode loop fixed and backing out only these three keys, over the
whole 240-seed train split: 1 of 240 trajectories changes. Eight of the 240 seeds hold Loyalty
Card at some blind deal; one of the eight moves. That is consistent with the published headline
moving 3.204 → 3.196 when the fix landed in v2.1 — a real change, and a small one.

Worth stating because the direction is counter-intuitive on that one seed: the fixed agent dies
two antes earlier. The unfixed preview under-valued a hand that would have cleared the blind, so
the agent declined to clinch, played on, and the run diverged into a different (and luckier) line.
Run depth is not a measure of the ranker's honesty.