Spend more computation on confident bits?
The compressor combines predictions from many models. Before its expensive mixer cascade, a small online model decides whether it is already confident enough to predict a bit. If its confidence exceeds a threshold, the compressor uses that prediction and skips the cascade, secondary estimation and BITLSTM correction for that bit. Both encoder and decoder make the same decision from the same earlier data.
The pinned upstream documentation reports about 5.1% fewer instructions for a small compression cost: roughly 2.7 KB on one 80 MB window, extrapolated there to about 18 KB on the full stream. Those are upstream measurements, not results reproduced here. They suggest a bounded experiment: spend a little more time and check whether it recovers bytes. Even the upstream estimate is far too small to make this a complete prize-winning strategy.
A frozen three-model comparison
| Model | Skip threshold | What changes |
|---|---|---|
| Upstream settings | 8 | Baseline |
| Stricter confidence | 9 | Fewer bits qualify for the shortcut |
| Adaptive skips disabled | 31 | The gate's logit is clamped to ±30, so it cannot pass this threshold |
The saved protocol fixes both canonical article inputs and the exact changed source hashes before measurement. Each candidate changes one constant. Dictionary, weights, link context, learning rules and compiler settings stay fixed. The experiment retains losses as well as gains.
More computation need not improve compression. The shortcut also determines which bits train the mixer and BITLSTM, so changing it changes their future state. The disabled control still computes and trains the small gate, and a separate character-rate shortcut remains active. This is a test of routing more bits through the cascade, not a complete removal of shortcuts or gate overhead.
Every candidate must restore the original XML exactly. A gain on both reused fixtures earns a test on the reserved third range and a native Linux comparison. Concurrent Mac timings are exploratory; a useful size/time tradeoff needs paired serial timings and ultimately the full submission's runtime limit. The experiment playbook explains those gates.
First native Mac results
| Model | Screen payload | Change | Disjoint payload | Change |
|---|---|---|---|---|
| Upstream threshold 8 | 119,065 | — | 117,438 | — |
| Threshold 9 | 119,071 | +6 | 117,436 | −2 |
| Adaptive skips disabled | 119,084 | +19 | 117,451 | +13 |
All six trials restore the exact original XML and pass context checks. Neither candidate improves both fixtures, so neither advances to the reserved third range. Disabling the shortcut makes both payloads larger: spending more computation does not necessarily improve the online ensemble's predictions. These are small fixture deltas; they do not contradict a different result on the full stream.
The screen and disjoint record retain all outcomes. Jobs overlapped with other experiments, so their timings are exploratory. The separately predeclared native Linux comparison has now finished too; the upstream threshold remains the default.
Native Linux confirms the rejection
| Model | Screen payload | Change | Disjoint payload | Change |
|---|---|---|---|---|
| Upstream threshold 8 | 119,067 | — | 117,426 | — |
| Threshold 9 | 119,069 | +2 | 117,430 | +4 |
| Adaptive skips disabled | 119,083 | +16 | 117,442 | +16 |
The Linux screen and Linux disjoint record retain native serial trials and exact original XML restoration. Both candidates lose on both fixtures. The tiny Mac disjoint improvement at threshold 9 changes sign on Linux, another reason to avoid choosing a model from a few Mac bytes. No candidate advances and the reserved third range remains untouched. This rejects these settings on these fixtures, not all possible depth strategies.