A compact replication of the Knowing–Using Gap on Qwen2.5-1.5B: direct recall separates sharply from two-hop use; activation patching reveals latent signal but does not recover the answers.
On 1,000 synthetic chains, atomic recall reached 100% at epoch 4 while two-hop accuracy stayed at 0% through epoch 31.
Patching moved answer-relevant signal—best source states clustered in early and late layers—but never made the answer rank first.
Red cells improve the correct token's log-probability; stars mark each case's best source→target pair. Exact recovery remained 0/4.
| Stage | Scale | Result |
|---|---|---|
| Toy end-to-end | 8 chains | 100% recall · 12.5% use |
| Full behavioral run | 1,000 chains | 100% recall · 0% use |
| Patching at saturation | 4 scans + 1,000 fixed-pair tests | fixed pairs: 0% |
| Oracle validation | 16 scans | superseded after semantic mismatch |
| Faithful residual-post oracle | 12,544 patches | +16.16 nats · 0/16 recovered |
| Layer-swap heatmaps | 4 × 784 patches | +16.72 nats · 0/4 recovered |
Synthetic single-token labels replace the unreleased STaRK experiment data. Results therefore support the paper's central behavioral claim, not its cross-domain or exact-recovery numbers.