alephlm-0 / s0a /s0a_erank.json
AbstractPhil's picture
S0a erank control
3a99912 verified
Raw
History Blame Contribute Delete
1.06 kB
{
"chunk": 0,
"n_sample": 100000,
"device": "cuda",
"ts": 1785717563.1047525,
"bert_in_domain": {
"erank_raw": 39.62,
"erank_normalized": 40.74,
"rows_used": 100000,
"dim": 768
},
"consensus_in_domain": {
"erank_raw": 28.07,
"erank_normalized": 28.07,
"rows_used": 100000,
"dim": 768
},
"bert_stsb_out_of_domain": {
"erank_raw": 34.35,
"erank_normalized": 40.25,
"rows_used": 2910,
"n_sentences": 2910
},
"bert_over_consensus_normalized_gauge": 1.45,
"bert_over_consensus_raw_gauge_diagnostic": 1.41,
"verdict_thresholds": {
"agreement_gt": 2.0,
"encoder_lt": 1.3,
"gauge": "normalized",
"provenance": "operationalization chosen in-script 2026-08-02 BEFORE the first run, NOT in the plan; plan S0a (2026-08-02_alephlm_anchored_trunk.md \u00a73) registers only the erank measurement \u2014 the >>/~ qualitative reading is this script's header"
},
"verdict": "INTERMEDIATE \u2014 partial: both the construction and the member contribute"
}