| 2026-10-29 |
Readout initialisation (proposal §3.7: U = 0) |
Added a fixed scale s = rms(H) and, after the zero-U bootstrap failed, a non-zero init U ~ N(0, 0.2²) used for all adapters including baselines |
Bootstrap degeneracy (FAILURES.md); decided on development data before any test split was read |
recorded before test unsealing |
| 2026-10-30 |
Controller features |
Added the normalised backbone residual and the backbone's predictive entropy (cost charged) |
z-only features at chance (D10) |
development-stage change |
| 2026-10-30 |
Label stream |
Added ctrl_train (seed 4) for controller labels; train/calib/dev/test byte-identical |
editor-seen labels unrepresentative (D11) |
development-stage change |
| 2026-10-30 |
Additional controller variant |
Ranking losses (D12) as a separately labelled experiment |
mechanistic diagnosis |
not part of the pre-registered primary comparison |
| 2026-10-30 |
Dense-slot baseline |
Seeds run sequentially / fewer seeds if not complete at freeze time (reported with actual n) |
~8× cost per step; node contention |
reported |
| 2026-10-30 |
Joint fine-tuning reference |
One attempt (lr 3e-4) degraded the backbone and was stopped; not re-run |
compute priority |
not pursued |
| 2026-10-30 |
Pre-registration freeze |
protocol/preregistration.md v0.1 draft → v1.0 frozen before unsealing (see file header) |
— |
see file |
| 2026-10-30 |
Dense-slot editor at 103M |
Not run with the matched recipe (≈ 40 s/step on an M2 Ultra → > 30 h); replaced by a reduced-budget probe (1,500 steps × 8 × 512 tokens at lr 3e-4, i.e. 1/4 of the sparse editor's exposure) reported as a probe only; the first probe at lr 1e-3 collapsed like the sparse editor |
compute infeasible in the session |
reported, not matched |