Skip to content
HOPIUM
An AI that admits when it's wrong.
model v3-12h-logreg-20261104
BTC......ETH......SOL......connecting to tape

Model registry

data: live database, read-only

Every version that has ever published, with the exact artifact, code and rules. Entries are append-only; a new version never resets or hides an old one. Records below are per version; nothing is pooled across versions.

v3-12h-logreg-20261104

change
initial release. First engine-v3 live bundle; config a2dedaa2 chosen by the pre-registered rule (REVIEW.md 2026-11-04). Replaces 1h/4h v2 models, which never published live rows.
artifact sha256
54c2cd7ee6ea1b0606ee54f9d0a15825885a82d06074c795476f982177e15206
code commit
13d45b0de4ef
feature schema
features_v3 groups=base+flow+mtf+volreg+cross+cal+fund
training
all available history through 2026-11-03 19:00:00+00:00; cutoff 2026-11-03T19:00:00+00:00
decision rule
dec-v3: score=p_big_up-p_big_down; UP if score>threshold, DOWN if score<-threshold, else NO_CALL; threshold 0.16 (inner validation (last 90 days)), k=0.01
grading rule
grade-v1: direction of move from reference close to close at target_time; |move|<0.1% is FLAT
activated (first live row, database clock)
retired
active or not retired
research
ml/RESULTS.md, ml/EXPERIMENTS.md, ml/models/engine_v3.json

Live record for this version only: 213 receipts, 1 calls, 213 graded (0 hit, 0 miss, 1 flat). Too few calls to judge anything.

v3-12h-logreg-20261108

change
automatic retrain, unchanged procedure. Scheduled weekly retrain (research/weekly.py)
artifact sha256
550f64f8b8b2f7104e6f03e85a6fe2a9b4db03965f683fecd4a1e80b1ccb48c3
code commit
33d5114d3575
feature schema
features_v3 groups=base+flow+mtf+volreg+cross+cal+fund
training
all available history through 2026-11-07 19:00:00+00:00; cutoff 2026-11-07T19:00:00+00:00
decision rule
dec-v3: score=p_big_up-p_big_down; UP if score>threshold, DOWN if score<-threshold, else NO_CALL; threshold 0.16 (fallback: median of 17 walk-forward fold thresholds (protocol found none profitable)), k=0.01
grading rule
grade-v1: direction of move from reference close to close at target_time; |move|<0.1% is FLAT
activated (first live row, database clock)
retired
active or not retired
research
ml/RESULTS.md, ml/EXPERIMENTS.md, ml/models/engine_v3.json

Live record for this version only: 483 receipts, 16 calls, 447 graded (7 hit, 3 miss, 0 flat). Too few calls to judge anything.