| """ | |
| K6 — Encompassment per-cell (R8 fix). | |
| R7 bug: K6 = 0.6138 uniform across all cells | |
| Cause: encompassment_engine_v2 uses a global pattern index. | |
| R8 fix: K6[cell] = fraction of layer-distinct fingerprints | |
| that this cell has seen (via fp_history). | |
| """ | |
| import numpy as np | |
| def _hamming_sim(a, b): | |
| """Similarity = 1 - normalized hamming distance on hex strings.""" | |
| if not a or not b or len(a) != len(b): | |
| return 0.0 | |
| try: | |
| x = int(a, 16) ^ int(b, 16) | |
| except ValueError: | |
| return 0.0 | |
| n_bits = len(a) * 4 | |
| dist = bin(x).count("1") / n_bits | |
| return 1.0 - dist | |
| def build_layer_fingerprints(cells, history_window=100): | |
| """Group distinct fingerprints by layer.""" | |
| out = {} | |
| for c in cells: | |
| h = getattr(c, "fp_history", []) | |
| if not h: | |
| continue | |
| out.setdefault(c.layer_idx, set()).update(h[-history_window:]) | |
| return out | |
| def compute_k6_percell(cells, layer_fps, sim_threshold=0.7, | |
| history_window=100): | |
| """ | |
| For each cell: | |
| K6 = fraction of layer's distinct fingerprints | |
| that the cell has 'seen' (similarity >= sim_threshold). | |
| Returns {cell_id: K6 in [0, 1]}. | |
| """ | |
| out = {} | |
| for c in cells: | |
| lfps = layer_fps.get(c.layer_idx, set()) | |
| if not lfps: | |
| out[c.cell_id] = 0.0 | |
| continue | |
| history = getattr(c, "fp_history", []) | |
| if not history: | |
| out[c.cell_id] = 0.0 | |
| continue | |
| recent = history[-history_window:] | |
| matched = 0 | |
| for fp in lfps: | |
| for h in recent: | |
| if _hamming_sim(fp, h) >= sim_threshold: | |
| matched += 1 | |
| break | |
| out[c.cell_id] = matched / max(1, len(lfps)) | |
| return out |