Research Runs Summary
A compact map of the runs behind the published reports. This page intentionally summarizes the evidence chain instead of copying every intermediate artifact.
Current Verdict
The best current weight-projector is a fixed damped-linear rule when the trajectory is coherent, but it is not a universal ANCHOR default. GSM8K has aligned consecutive updates and benefits from projection. Big-Math has near-orthogonal consecutive updates and should default to hold-stale except at the freshest anchor.
M4 Spine
| Run | Status | Role |
|---|---|---|
| EXP-43 | PASS | Dense GRPO full weight-matrix collection; 160 bf16 snapshots and R2 verification. |
| EXP-44 | PASS | Offline weight-projection sweep engine and predictor-family reconstruction checks. |
| EXP-45 | PASS | MOAT scorecard contract, schema, taxonomy, and self-test substrate. |
| EXP-58 | PASS | Big-Math 1000-step checkpoint and fp32 weight-trajectory collection. |
Decision Lanes
| Date | Run | Result | Takeaway |
|---|---|---|---|
| 2026-07-03 | EXP-47 | FAVORABLE | Fixed damped-linear projection helps on GSM8K: ratio about 0.94 and h_safe around 30 global steps. |
| 2026-07-03 | EXP-48 | NEGATIVE | Fixed second-order curvature regresses badly; keep the simpler linear rule. |
| 2026-07-04 | EXP-49 | SUB-THRESHOLD | Self-correction improves slightly but not enough to beat the prefer-simplicity margin. |
| 2026-07-04 | EXP-60 | DATASET-SPECIFIC | GSM8K projector does not transfer to Big-Math; hold-stale is effectively best. |
| 2026-07-05 | EXP-61 | FRESH-ANCHOR ONLY | Math projection clears about 1% only at Delta=1, h=1; operational staleness remains do-nothing. |
Design Direction
Use a coherence gate. When consecutive updates are aligned, fixed damped-linear projection is cheap and useful. When updates are near-orthogonal, shorten refresh or hold stale rather than projecting noise. Rank-1 state remains promising as a long-range position or spectral-correction basis, not as the short-horizon predictor.