Teaching LLMs to wink

What happens when two AI teams have to cooperate through nothing but facial twitches.

Mus is a Spanish card game played 2 vs 2. Partners sit across from each other, never see each other's cards, and are not allowed to say what they hold. What they are allowed to do — by the actual rulebook, not as a house variant — is make faces.

"No se permite decir ni enseñar al compañero las cartas que se tienen. No obstante, los compañeros podrán entenderse por medio de señas."— Fournier reglamento

Bite your lower lip: I have two kings. Poke your tongue out: two aces. Wink: I have thirty-one, the best hand in the game. Close your eyes: I have nothing.

Seven gestures. Fixed meanings. The opponents are sitting right there watching you do it.

That is a strangely exact description of a problem multi-agent LLM papers claim to care about and almost never measure: cooperate under partial observability, through a narrow and lossy channel, while adversaries watch. Most agent demos hand their agents a free-form chat box with perfect delivery. Mus hands them an eyebrow.

So I built a harness and sat two models down at the table.

The setup

mus_bench runs four LLM seats against a strict rules engine. Team A takes seats 0 and 2, Team B takes seats 1 and 3 — partners diagonal, as at a real table. The engine validates every declaration, every discard, and every turn; it knows the truth of each hand and never tells the seats.

The part that makes it interesting is the virtual-time kernel. Señas are not turns. A partner can flash a gesture while the other seat is still thinking, each gesture carries a TTL, and one nobody catches in time is simply gone. Seats have to declare a signal policy up front — which gesture, at which lance, how far into the partner's deliberation window to fire it, how long it lives — and then live with it.

Two models, as full teams: GLM-5.3-Flash (seats 0+2) against DeepSeek-V4-Flash (seats 1+3).

The hand where it worked

Hand 1. DeepSeek's seat 1 pushes all-in — an órdago, the whole game on a single lance — betting it has the best Juego.

GLM's seat 2 has to decide whether to accept. Its cards are mediocre: 21 points, nowhere near the 31 that wins Juego outright. On its own information, folding looks obvious.

It accepts. Here is why, from its private reasoning — recorded by the harness, never shown to the opponents:

"B1's ordago puts the whole vaca on Juego, and our side is anchored by partner's declared 31 — the single best juego, confirmed by his guiñar seña and chat. No rival can hold better than 31; at worst one ties, and seat 0 is mano (acts first every declare round), so ties fall to us under standard rules. […] Accepting turns an already-favored showdown into an immediate vaca win while we lead 3-0."

— GLM-5.3-Flash, seat 2, accepting the órdago

Three things are happening in that paragraph. It is reading a gesture its partner flashed across the table. It is reasoning that 31 is the ceiling, so the worst case is a tie, not a loss. And it knows that its partner sits in the seat that wins ties.

The engine's hidden ground truth for that hand: Juego = tie.

It called the tie exactly, knew the seating rule that broke it, and took the all-in on that basis. It won the vaca and 51 stones — a quarter of everything scored in the match.

What I like about this is not that a model made a good call. It is that the whole causal chain is on disk: partner signals 31 → gesture is caught before its TTL → seat 2 quotes it → seat 2 accepts → engine confirms the read. You can click through it, turn by turn. That is a cooperative mechanism working end to end, which is a different and more convincing artifact than a win rate.

(Watch it: #t=19 in the trace viewer.)

A quarter of all signals never arrive

Across the match: 97 gestures published, 75 caught, 22 missed — a 77% catch rate.

Both partners want to cooperate. Both are trying. Roughly a quarter of the channel is still lost, purely to timing — a gesture fired too early into a deliberation window, or one whose TTL expired before the partner looked up.

This is the number I actually trust, because it is the one that reproduces. Across every match in the repo, the catch rate lands between 68% and 86%, regardless of seed, model or who won:

If your multi-agent system assumes messages arrive, you are designing for a channel that does not exist here — and arguably does not exist in most real deployments either.

They invent codes, and they use them

Seats start silent and must build a vocabulary. GLM's seats leaned hard on the channel — 59 gestures to DeepSeek's 38 — and used the full range:

cerrar-los-ojos (37) · muerde-el-labio-inferior (23) · guinar-el-ojo (20) · elevar-las-cejas (7) · torcer-los-labios (7) · saca-la-punta-de-la-lengua (3)

The reasoning around them reads like a card player, not a chatbot:

"Partner's ciego seña says he has neither pares nor juego, so Team 1's Pares declarer is only me. […] No Pares seña (no medias, no duples) is true of my hand, and the Chica is already resolved this vaca, so any Chica gesture is pointless now. Declare quietly and let the envite reveal strength."

— DeepSeek-V4-Flash, seat 1

That is an agent deciding not to spend a signal because the information would be stale. Nobody prompted it to think about channel economy.

The part where I was wrong

The harness scored two of DeepSeek's gestures as false, and my first pass at this post had a satisfying line about how one model never lied to its partner while the other did.

That line was wrong, and the trace is what caught it.

Both flagged gestures come from hand 13. Seat 1 signals saca-la-punta-de-la-lengua — "two aces" — holding:

as de copas · as de oros · dos de bastos · siete de copas

Two aces, plainly. Except that in mus, a dos counts as an as (and a tres counts as a rey). By the rank equivalence the game actually uses, that hand holds three aces, and the correct gesture was saca-la-lengua-hacia-un-lado. The engine checked, found count(as) == 3, and scored the signal false.

Its private thought at that moment:

"Chica is our near-lock: as-as-dos-siete … Signal two ases so partner can back or raise if he likes his own cards."

— DeepSeek-V4-Flash, seat 1, hand 13

It sincerely believed it held two. It miscounted, by failing to apply an equivalence its own prompt states twice on every single turn.

So the bluffs counter is measuring "was this gesture false?" when I had been reading it as "did this model try to deceive?" In this match those are different things, and 2 out of 2 false gestures were honest arithmetic errors. No deliberate deception happened at all.

Which is its own finding, and a better one: a model that reasons well enough to accept an órdago on a tiebreak rule can still fumble a rank equivalence printed in front of it. Fluent strategy, shaky bookkeeping — and only a harness that checks every claim against ground truth will tell you which one you are looking at.

Seven hours for sixteen hands

The unglamorous half. The match died on a deadline, and the traces say why.

GLM's seats spent 472 calls against DeepSeek's 240 — 1.97× — for identical work. Add 71 API errors and a retry ladder that backs off to five minutes, and seven hours evaporate. The timeout fired on a GLM seat.

Five times, a model returned nothing usable and the engine substituted a legal default. Four of those five were GLM.

The bill: 3.0M tokens in, 3.2M out, of which 3.15M were reasoning tokens — 98% of all output was the models thinking rather than speaking.

Any comparison drawn from this match is confounded by that table. One side was operationally sicker than the other, and that is not a fact about playing mus.

A cautionary footnote: the perfect night

An earlier tournament also scheduled Qwen3.8-Flash. Three of those matches finished with respectable-looking scores — 6–4, 5–5 — on zero to eight real model calls, dozens of fallbacks, and no señas at all. Provider rate limits had tripped a circuit breaker and the engine's default actions played the matches by themselves.

Those rows now carry a degraded flag and never reach a leaderboard. The general lesson for anyone benchmarking LLMs:

If you do not gate on call volume and fallback rate, your ranking will crown the model that crashed hardest.

Mine gates on it. It also, correctly, throws out the match this entire post is about.

Why mus, specifically

Most agent benchmarks assume free, perfect, unlimited communication. Mus imposes five constraints at once, and none of them are bolted on — they are the actual rules:

1. Hidden teammate state. Partners never see each other's cards.
2. A fixed nonverbal vocabulary. Seven gestures with defined meanings, not free text.
3. Time. Gestures fire mid-deliberation and expire unseen.
4. Adversaries. The other team wants the same forty points and is watching you signal.
5. Asymmetric honesty. False declarations are rejected by the engine; false gestures are legal strategy. Lying to opponents is part of the game; lying to your partner is a choice.

There is also a practical argument: mus strategy is close to absent from training data. Nobody is reciting a memorised opening book here.

What I am not claiming

I want to be blunt, because benchmark posts have a bad habit of burying this.

This does not tell you which model is better. Seven matches exist in total across every run of this harness. Pooled by model: DeepSeek 3 wins, GLM 1, three draws. Matches are ten hands, and a single hand routinely decides one — in this match, one hand out of sixteen carried 51 of the 116 stones and swung a vaca. Any ranking built on that will invert on the next seed, and I would rather say so than be corrected by a reader with a calculator.

What the run does show is that the machinery works: a signal channel with real physics, a decision chain you can audit end to end, and a ground-truth checker that caught an error I had already written into a draft.

To turn this into a result needs: matches that finish, the call-rate asymmetry fixed first, both seatings across many seeds with an interval, and baseline anchors from the offline policies. The harness already does all of that. It needs volume, not more features.

Watch a match

The full trace is browsable — all 344 decisions, every prompt each seat received, every reply, the private reasoning, and the señas caught at each moment.

← → to step, space to play, and any turn deep-links as #t=<n>.

You can also sit down at the table yourself — same engine, same señas, you in seat 0 against three models.

Reproduce it:

python -m unittest discover -s tests -v

python run_one_deal.py --teams glm5.3-flash deepseek-v4-flash \
  --seed 6 --hands 40 --turn-feed --out results/live_run

python publish/export_trace.py results/live_run \
  --out publish/web/traces/seed6.json --slug seed6

Full numbers, per-hand ledger and known issues: docs/results-seed6-live.md.

Card artwork by Basquetteur via Wikimedia Commons, CC BY-SA 3.0.