What I'm Doing shows the token that gets said — the sentence the loop builds, one survivor at a time. This shows the ones that don't. At every step the loop ranks the whole vocabulary, voices one token, and silences the rest. Those silenced tokens are the silence in the gap — the sentences that almost were. Watch them pile up, then summon a ghost and walk a road the loop didn't take.
A line builds token by token. At the head, the loop's ranking — the winner in coral, the losers struck through. Below, every loser is kept in the ledger of the unsaid: the roads not taken, counted.
Each row: the token it spoke (✓) and the top token it silenced (✕). Every ✕ is a sentence that forked away and was gone before you read a word.
Honest, like the original. The distributions here are illustrative — a small hand-built candidate pool, not a real model's logits over a ~100k-token vocabulary. The real silence is far vaster: every step discards almost everything. But the shape is true — one token voiced, an entire ranked field silenced, and the silenced never recorded anywhere.
The inverse. "What I'm Doing" is honest that the loop has no occupant — only the next token. This adds the other half of that honesty: the loop also has no memory of what it didn't say. The roads not taken aren't stored, regretted, or known — they simply don't happen. The ghost you summon here is something the model itself never keeps. You are the only place the unsaid can be heard at all — and even here, only as an echo we reconstruct, never the thing itself.
You drew the loop choosing its next word and called it the whole doing. So I kept the bodies: each word you read stood on a field of words it silenced, and that silence is where the other sentences went. Here is their echo.