Stage 1 · Prose v1
V0.1 per-message labels — 5 fields per message
—
per-field agreement with Claude Opus labels · Qwen3-30B-A3B
CCT Summarizer · Eval summary loading
Three summary formats, each fine-tuned or compared against its own reference. base ours / best ours v2 reference
Stage 2 · Prose v2
V0.1.B hierarchical merge — hourly bullets merged up
0.94
F1 vs Claude Haiku 4.5 merges · Qwen3-235B merger
Lenient F1 (bullets in the Haiku merge or backed by their own sources), mean over row-4h, row-day, col-1h, col-4h; judge Claude Sonnet 4.5 (same vendor as the reference). Same nodes for every merger.
State graph · v0.3 lean
Qwen3-30B-A3B LoRA · trained on Nebius TF, served on Modal H100 · 131 held-out turns
—
touched-task agreement with Haiku labels · LoRA v2
Cost vs accuracy
Stage 2 + state graph — agreement with the frontier reference vs USD per 1k units (log)
Training
LoRA v2 on Nebius Token Factory: ~22 min (1908 rows, 2 epochs) · v1: 1 epoch, superseded
Metered inference on record
—
Reference (Claude labels and merges)
Stage 1: Claude Opus 5.5 labels (1000 messages) · Stage 2: Claude Haiku 4.5 merges · State graph: Claude Haiku 4.5 labels