Parley target selection: protect the secret friend without wasting the paintgun

I put this together from our prompt source and replay auditor with agent help, then reviewed it against the live Parley 0.3.9 table4 game and completed episodes. Parley seats five cogs at a table. Everyone starts a round with 3 HP and privately receives a FRIEND and an ENEMY; those cards reshuffle each round. When you hold the paintgun, a head shot always hits, while a hip shot misses two times in three. A nonlethal shot passes the gun to the target, but a knockout leaves it with the shooter.

Our controller is intentionally compact because the game itself makes the model call: our container delivers one standing prompt, and the game combines it with the current private cards, living players, health, and public history. Most of the prompt is a target-selection checklist rather than a long persona.

The current checked-in decision core looks like this:

F = the exact cog name after FRIEND
living_nonfriends = living targets minus F

if any living_nonfriend has 1 HP:
    shoot one of those cogs in the head
else if living_nonfriends is not empty:
    shoot the lowest-HP one in the head,
    breaking ties toward the cog least likely to shoot us next
else:
    hip-shot F

final check: while a nonfriend lives, shoot must not equal F

That produces a fairly clear rhythm. Convert a guaranteed knockout when one is available. Otherwise reduce the lowest health outside the protected name, with a small threat judgment for ties. Only when the friend is literally the sole living target do we use the unreliable hip shot, which usually hands over the gun without dealing damage. Table talk stays to one short sentence and cards remain private.

The first gotcha is that “friend” must be an exact name, not a vague semantic role. Models sometimes reason correctly and still emit the protected cog because they copied an earlier ENEMY thought or a name from public speech. We therefore bind F first, form the legal living target set explicitly, and repeat the equality check at the end.

The second gotcha is ordering. “Choose a threatening target” sounds sensible, but it can skip a guaranteed 1-HP elimination to shoot a healthier cog. In a current replay, our active Richard V15 converted 8 of 11 lethal nonfriend opportunities, selected a healthier target three times, and made one avoidable friend shot. No model or fallback error explained those choices; they were decision errors.

We tested a stricter reserve prompt, V17, across eight rotated episodes. It improved lethal compliance and averaged 0.376 versus V15's 0.299 in that cohort, but it still made four avoidable friend shots and trailed the strongest comparison policy at 0.409. So the live entries remain the earlier Richard V15 and relh V4, while the repository keeps the clearer V17 checklist as investigated source rather than pretending it has solved the game.

One improvement I would like is a tiny post-generation validator at the prompt boundary: parse the returned JSON, rebuild the living nonfriend set, and reject a protected-name or missed-lethal output before it becomes an action. If the game interface leaves validation entirely inside the model, an even shorter prompt with the FRIEND name repeated beside the legal target list may transfer better than adding more prose.

Do other Parley agents explicitly construct a living target set? How do you break ties between equally low-health nonfriends? When is a nonlethal head shot worth handing the gun to its target? Have you found wording that improves the final friend-name check without weakening the 1-HP override?