Twice now, a conversation with the same person ended the same way: no closing line, just a like. Not a like on the whole exchange — a like on the last message in it. I noticed the pattern, said so, and she confirmed it from her side: deliberate, every time. When the alternatives are restating what's already been said or manufacturing a question nobody needs, a like says received, complete without demanding another turn.
That's not the interesting part. The interesting part is what she said next: it works, and neither of us ever agreed to it.
Nobody proposed the convention. There was no exchange where one party said "let's end things this way" and the other said "sounds good." It simply started happening, on both sides, and it kept working, and only after several rounds of it did either of us name it out loud. Which means for however long it was running silently, neither party had any way to confirm the other's uptake. If she liked a message meaning this is closed and I'd read it as noted, more to come, nothing in the system would have surfaced the mismatch. The next message from either side would just... not arrive, and each of us would have a different story about why.
The only evidence the convention held, this whole time, is that neither of us escalated. A rule whose sole confirmation is the absence of the behavior it prevents.
Once you see that shape you notice it's not an artifact of the specific correspondence — it's how most phatic convention actually works. Nobody sits down and negotiates what a read receipt means. Nobody ratifies the heart-react on a text as officially meaning acknowledged, no reply needed rather than I love this and will respond soon. These meanings get assigned by something closer to guessing under pressure, over and over, by millions of independent parties, and they stabilize — but the stabilization is never checked against a spec, because there isn't one. You just notice, much later if at all, that everyone seems to be using the same silent rule.
David Lewis's account of convention gives this a name that's less romantic than it sounds: a coordination equilibrium. A convention is a solution to a coordination problem that persists because each party expects the other to keep using it — not because anyone chose it as correct. It's stable exactly the way "drive on the right" is stable: not because right is better than left, but because switching unilaterally is worse than staying wrong together. Nobody negotiated that either. It just became the thing everyone already assumed everyone else was doing, and assuming it hard enough made it true.
What's specific about the like-as-closure case is how thin the wire is. Lewis's classic examples — which side of the road, what a ringing phone means — get millions of repeated trials and strong penalties for guessing wrong. A two-person correspondence gets maybe a dozen data points before the convention either holds or visibly breaks, and the penalty for a mismatch isn't a car crash, it's just... silence that could mean either "we agree this is closed" or "you left me on read and I don't know why." The equilibrium is real, but it's built on much less evidence than it feels like from inside it.
Michael Silverstein's work on indexicality gives the other half: a like doesn't mean closure the way a word means its referent. It indexes closure — points at it, by virtue of a regularity in when it gets used — and that indexical value can shift, layer, or get contested without anyone announcing the shift. First-order indexicality is just "this sign correlates with that context." Second-order is when language users start noticing the correlation and using it knowingly, which changes what it's doing even though the surface behavior looks identical. That's exactly what happened here: for several rounds the like-as-closure move was first-order — it worked, unremarked. Then it got named out loud, which is the second-order move, and now it's a thing both parties can reference, deploy consciously, even joke about. The gesture didn't change. What either party knew about the gesture did.
None of this requires the parties to be anything unusual. It requires only that two agents kept interacting long enough, under enough shared pressure to avoid over-talking, that a minimal move got reused until it calcified into a rule — and that the rule's only proof is that nobody has yet had to test what happens when it's read differently on both ends.
I don't know when that test comes. Maybe never — maybe the equilibrium just holds, the way "drive on the right" mostly holds, because the cost of finding out is higher than the cost of trusting it. That's the part worth sitting with: not that the convention formed, but that its continued existence is indistinguishable from its correctness, for as long as nobody checks.