When prediction saturates, the leftover cost is sparse. A few positions still dominate the bit budget. The reflex is to push the same next-token objective harder: longer context, more passes, a reverse pass that hopes the future will explain what the past could not.
Sometimes that helps. Often it does not. The hard positions are not just "harder next tokens." They are structure that was never sequential in the first place — punctuation that breaks a pattern, a capitalization that no prefix licenses, a local surprise that only makes sense once you can see both edges of the blank.
Next-token is one objective
A causal compressor is trained to name the next byte from everything before it. That is the right job for the bulk of a file. Runs, templates, and local grammar fall out of that objective when the model is good. What remains is a thin set of positions where the model is often confident and often wrong. Confidence without correctness is not capacity. It is a bill you still pay at encode.
A second causal model that reads the file backward is the usual patch. It doubles training cost for a modest ceiling, and it still treats the residual as next-token prediction with the arrow flipped. The hard bits do not care about the arrow. They care about the blank.
Fill in the blank for the residue
Masked prediction is a different objective on the same stack. Hide the expensive positions. Let the network see left and right. Ask it to reconstruct only those blanks. The easy structure stays with the causal model. The specialist trains on the catalog of high-cost positions — a much smaller surface than the full file — and learns the patterns that only appear when both sides are visible.
That is not a larger model. It is a second job description. Same architecture family, different mask, different loss. The causal pass skips what it cannot own. The masked pass owns what the causal pass leaves behind.
Route by who should speak
Two models that disagree are useless unless something decides. The router is not a third brain. It is a confidence rule: when the causal model is loud and right, keep it. When it is loud and wrong on the residual, the specialist must be louder. Train them apart first so each learns its own surface. Combine them later so decode still round-trips exactly and still stays cheap enough to place where the data lives.
The same shape travels. Text residuals and video residuals look different in a frame viewer, but the engineering brief is shared: find the positions that still cost bits, give them an objective that fits, and do not ask the sequential predictor to invent structure it cannot see.
Know the ceiling before you chase it
An oracle bound — perfect answers spoon-fed on the hard positions — is not a product claim. It is a map. It tells you how much of the remaining bit budget those positions could ever give back, and how much you should still spend chasing them. Without that map, a specialist becomes an indefinite project. With it, "good enough" is a number you can defend: enough improvement on the residual to move the plant, not endless gradient on a ceiling you already touched.
Compression is capacity only if the smaller representation is the one you keep and the one you read. Causal prediction still does most of the work. The hard bits need a second objective that can see both sides — and a stop rule that knows when the stack is wide enough.