367

Source document

[2026-03-18 14:08] xenonfun: ```⏺ 3 configs done, 4th running. Early results: ┌──────────────────────┬─────────────┬────────┬────────┬────────┬────────┬────────┐ │ Config │ Stage1 │ Stage2 │ Stage3 │ Stage4 │ Stage5 │ Chance │ ├──────────────────────┼─────────────┼────────┼────────┼────────┼────────┼────────┤ │ baseline (no entity) │ 1.8% │ 2.7% │ 2.5% │ 2.3% │ 1.2% │ 1.6% │ ├──────────────────────┼─────────────┼────────┼────────┼────────┼────────┼────────┤ │ stage1 10% │ 3.7% │ 1.6% │ 3.1% │ 2.5% │ 1.6% │ 1.6% │ ├──────────────────────┼─────────────┼────────┼────────┼────────┼────────┼────────┤ │ stage1 20% │ 7.4% │ 1.2% │ 6.6% │ 4.3% │ 2.2% │ 1.6% │ ├──────────────────────┼─────────────┼────────┼────────┼────────┼────────┼────────┤ │ stage2 10% │ ... running │ │ │ │ │ │ └──────────────────────┴─────────────┴────────┴────────┴────────┴────────┴────────┘ Stage 1 at 20% shows the first signal above chance — 7.4% on direct recall (chance 1.6%), 4.5× chance. It's weak but it's real. And interestingly, Stage 3 (rebinding) also lifted to 6.6% — the model is learning something about identity binding, not just memorizing the format. At 10% curriculum the signal is barely there (3.7%, only 2× chance). This suggests binding needs more curriculum pressure than retrieval did (retrieval worked at 2%). Still 5 configs to go. The G16 and late-injection results will be telling. ``` [2026-03-18 15:40] xenonfun: (files: message.txt) [2026-03-18 15:42] xenonfun: (files: plan.txt) [2026-03-18 15:48] xenonfun: (files: message.txt)