Not a run
Important things to note
Where the evidence stands after our Amazon rerun. Report the weak spots. Don't treat one seed as proof.
Read this first
Replicates
The professor's Amazon tuning result reproduces: the grid memory matches to six decimals, the network leads by 0.0128.
Shrinks
On held-out users the lead is +0.0044 on average (1.7%), with a positive interval on two seeds and a slight loss on the third.
Unstable
The network varies about ten times more between seeds than the grid (SD 0.0057 against 0.0006).
Readouts
Normalized and additive tie on every seed. Don't pick additive for a nicer story.
Rules
Development data only. No test split. Don't mix numbers from studies with different tuning budgets.
In order
Next steps
Proposed, for agreement with Prof. Wu.
- Share the rerun; ask for his held-out summary to compare
- Write down what counts as a win before running more
- Add seeds to settle the seed-to-seed swings
- Add the middle step: two networks, no shared like/dislike memory
- Do the literature check on what is new
- Decide whether the paper stays on FWPRec
- Only then: one locked run on the Amazon test split