Chapter two
The calibrated fusion
论点
The pipeline learns in two places, with a shared signal store between them. First, a learned encoder turns each protein into a sparse code and uses it to retrieve similar, already-labelled proteins: the candidates. Then a set of scorers put every clue onto one common scale, and a small model, trained separately for each branch of the ontology, fuses them into a single probability. The surprising lesson is that the biggest lever is not a deeper or fancier model, it is calibration: simply standardising a representation (rescaling each of its dimensions to a common range) moves the score more than changing which internal layer of the language model you read it from. Re-running the pipeline end to end did not land back on the figure this page used to publish, and the reason was measured: turning the candidate classifier on widens the pool per query, which dilutes the score. That figure is now withdrawn for a separate reason, and the campaign recomputes rather than reconciling the two.
展开标记，查看凭证以及重新生成该数值的操作。
证据
| Value | ||
|---|---|---|
| Board headline | withdrawn | measured against a baseline a later defect showed to be manufactured |
我们的保留说明
在他人指出之前，我们先行说明。
- The largest single lever is standardisation, a normalisation effect, not a deeper model. We name it so no one reads depth into the result.
- A single universal reranker reaches essentially the same number as the per-category combiner, so the per-category split earns little and we say so. The exact figures wait on the receipt below.
- The reproducibility receipt named by the plan, comparison.json, was not present in the storage snapshot this page was built from, and the values it would carry are not restated here. Until that file is re-materialised, this pillar rests only on the sealed board and the layer-ablation receipt.