Status: +18.3% over v2
Mean F1: 65.2%
F1 Folded: 64.3% | F1 Disordered: 66.2%
Best config: W=30, N=2, T_fill=3, T_drain=2, K=8, D=3, SH=True
Decouple the fill threshold from the drain threshold. In engineering, hysteresis prevents a system from "chattering" between two states by requiring different thresholds for state transitions in each direction. Applied to the Hatch:
- T_fill: A window with score < T_fill fills the cup (disorder evidence)
- T_drain: A window with score ≥ T_drain contributes toward the consensus drain
By setting T_fill > T_drain, windows in the "grace zone" (T_drain ≤ score < T_fill) neither fill the cup nor drain it — they are treated as ambiguous. This rewards "marginal order" in folded proteins without treating it as strong disorder evidence.
Additionally, a Super-Hatch hard reset fires when a window achieves a perfect score (7/7 features passing). A 7/7 window is a high-confidence structural anchor — a core hydrophobic packing event — and should immediately empty the cup regardless of previous state.
cup_level = 0
consecutive_ordered = 0
for each window w_i:
score = count(features meeting folded condition)
if score == 7: # Super-Hatch
cup_level = 0
consecutive_ordered = 0
elif score < T_fill: # Clog
cup_level += 1
consecutive_ordered = 0
elif score >= T_drain: # Potential drain
consecutive_ordered += 1
if consecutive_ordered == N:
cup_level = max(0, cup_level - D)
consecutive_ordered = 0
if cup_level >= K:
return DISORDERED
return FOLDED
Mean F1 jumped from 46.9% to 65.2% (+18.3%) — the largest absolute gain of any version. Both classes improved substantially and became balanced for the first time (64.3% folded, 66.2% disordered).
The dominant driver was T_fill=3, not the grace zone width. The heatmap showed a cliff between T_fill=3 (mean F1 ≈ 0.57–0.65) and T_fill=4 (mean F1 ≈ 0.43–0.51). This is a phase transition, not a gradual improvement.
Lowering T_fill from 4 to 3 stopped the model from penalizing "marginal order" windows (score=3/7) that represent breathing loops and flexible linkers in folded proteins. A score of 3/7 means the window passes 3 of the 7 biophysical conditions — it is not strongly ordered, but it is not strongly disordered either. Treating it as disorder evidence was the primary source of false DISORDERED predictions on folded proteins.
The threshold 3/7 ≈ 43% of maximum fold evidence. This value appeared consistently across all parameter sweeps and versions as the "Critical Mass" constant of structural integrity.
The Super-Hatch (SH=True vs SH=False) had negligible impact at this parameter point. The Hysteresis already solves the problem the Super-Hatch was designed for: by lowering T_fill, the cup rarely builds up enough entropy to need a hard reset. The Super-Hatch may become relevant at higher K values or with feature weighting.
T_drain=2 and T_drain=3 produced identical results in the top-10 configurations. The grace zone width (1 unit vs 0 units) is irrelevant — what matters is that T_fill dropped. The "Hysteresis" framing is mechanistically correct, but the grace zone itself is not load-bearing at this scale.
The Hysteresis is not about the gap between T_fill and T_drain — it is about setting T_fill at the right level. The model was miscalibrated in v1 and v2 by treating marginal-order windows as disorder evidence. Correcting this single threshold produced the largest performance jump in the entire experimental record.