โ PROVISIONALPortal numbers predate the Option-2 sealed re-run
The figures in this portal were computed on the original run set. The hypothesis โ HCIE wins at cold-start specifically because it is O(1) embedding-free โ actually requires a train-on-few / evaluate-on-unseen-usersprotocol. The current "cold-start" here is the first-N interaction slice of a full-population run, not that protocol. Read this portal as the calibrated evidence trail that motivates the sealed re-run, not as a terminal result.
Flips after:
Benchmark/baselines re-pointed at train-on-few/eval-on-unseen runs
Graph-KT widened beyond N=10 eval users
JT 6-dim ablation re-run at results-grade (replaces smoke runs)
Each new run sealed via Stage 0 and surfaced on /infrastructure โ sealed runs
Sealed thresholds (immutable): ฮฑ_floor = 0.01 ยท signal_ratio_threshold = 0.08 ยท Frozen before any run data was observed. Run IDs are permanent in hcie-final-postgres.