After ten turns, you should have a rough map of two of your opponent's melds in progress. That map is the difference between expert defense and guessing.
Modelling means building a mental picture of what an opponent's hand probably looks like. Not exact cards. Rough shape. Which suits they hold, which ranks, which melds they're building.
You don't need X-ray vision. You just combine the visible evidence. Every pickup is a fact. Every discard is a fact. Every meld on the table is a fact. Ten turns of facts adds up to a strong picture.
Expert players do this automatically. Their hand-shape read informs every decision they make. What to discard. Whether to declare or hold. Whether to break their own combos for defense.
On Play Rummy, all the public information is on screen. There's no memory limit forcing you to guess. Modelling here is about noticing, not remembering.
Any single pick or discard could be misleading. An opponent might pick up a random card. Might discard something that fits their build. Individual events are noisy.
A model averages the noise. Over ten turns, the true shape emerges. Random moves cancel. Real patterns compound.
A single read might say, "this opponent wants spades." A model says, "this opponent has a spade run in the middle, a low heart pair, and a set of fours forming." Big difference. The model gives you specific danger zones.
Track five things for each opponent through the round.
Fact one, pickups. Every card taken from the discard pile. Rank and suit.
Fact two, discards. Every card thrown. Rank and suit. Also, timing (early vs late).
Fact three, melds. Any group placed on the table.
Fact four, extensions. Cards added to existing melds.
Fact five, skipped pickups. Turns where the top of the discard pile looked useful but the opponent went to the stock instead.
Fold these facts together. After ten turns, you'll have a picture that's usually right about two of the opponent's three melds.
Two-fact reads are the base. Pick plus discard, or pick plus pick.
Opponent picks 6♠ on turn three, then discards K♥ on turn four. Reads. Building spades in the middle. Not building around Kings. High Kings unlikely.
Opponent picks 6♠ on turn three, then picks 4♥ on turn seven. Reads. Two melds in progress. A spade run and a low heart run or set.
Each new fact refines the model. By turn ten, you're confident about the two main directions.
Skipped pickups are underrated. When the top of the discard pile looks obviously useful and the opponent skips it, they're telling you something.
If the top is 7♠ and they don't take it, they either don't have 5♠-6♠ or 8♠-9♠ or 6♠-8♠. Some spade combos are ruled out.
If the top is a King and they don't take it, they don't have a K-K pair or a K-Q of the same suit. Kings are ruled out.
Ruling things out is as valuable as ruling them in. The model gets tighter every time an opponent skips.
The model is a living picture, not a static one. Every turn adds new facts. Every meld resolves some questions and opens others.
When an opponent melds, cross that group off the model. Their hand just shrank by three or four cards. What remains is a shorter hand with different unknowns.
When an opponent discards a card that would have completed a meld you'd guessed, revise the guess. That meld might not exist. Or they had it but decided to break.
Update every turn. A stale model is worse than no model, because you'll act on wrong information.
The model tells you which discards are safe and which are dangerous. Any card that fits either of your two best guesses is dangerous. Anything else is fine.
Example. You've modelled an opponent as building a middle spade run and a low heart set. Any middle spade (4-8) is dangerous. Any low heart (2-5) is dangerous. Everything else is safe (with normal one-rank-away caution).
Now your discard decisions are quick. Scan your hand for danger cards. Throw the safest thing you have. If the safe thing is a face card, so what. The point cost of holding one turn is smaller than the meld cost of gifting.
The model also tells you what the opponent needs. That's an opportunity.
If you know the opponent is building a spade run and you have the exact card they need (say 5♠), you can hold it as a block. The opponent can't get it from you. They can't draw it from the stock (fewer live outs). They may fail their meld.
Blocking a specific card has costs. You're carrying deadwood. But if the block prevents a full meld (worth 20+ points to them), it's worth it.
This is called dead-carding. It's an expert move because it only works when your model is right. If you dead-card the wrong rank, you carry deadwood for nothing.
Building a model from one fact. A single pickup is a hint, not a plan. Wait for two or three data points before committing to a model.
Never updating the model. Facts change every turn. If you built a model on turn five and haven't touched it since, it's probably wrong by turn ten. Fix, update every turn.
Trusting the model absolutely. Even a good model can be wrong. Some opponents bluff. Some make random choices. Treat the model as a strong guess, not a certainty. Adjust when new facts contradict.
Skipping the model in fast games. Speed games make players lazy. That's when the model matters most. Two facts a minute is enough to build one.
In your next five games, after turn ten, pause. Name your two best guesses about the opponent's hand.
When the round ends and the opponent's hand is revealed (in variants that show it), check your guesses. How many were right? How close were the near-misses?
Most players get one out of two right at first. After a few weeks, two out of two becomes common. The model gets sharper with practice.
Rummy at expert level isn't about your hand alone. It's about your hand and their hands, played against each other on a mental map.
The map is built one fact at a time. Every pickup, every discard, every skip contributes.
Ten turns of facts is enough. Build the model. Use the model. Update the model.
That's how expert Rummy players see the invisible.