flowcast-v5.5-lite v0.6.0

Sub-600MB unified stack tuned for LoudInk app-honest dictation (minimal polish).

Predecessor: nsalerni/flowcast-v5.5-lite

Production stack

• Writer: unified LoRA on LFM2.5-1.2B — app checkpoint 0000048_app_neural_v5.safetensors
• App polish: minimal_polish (version dots, SQL repair, correction chains, price fixes)
• Eval: scorer relaxations ship with the benchmark harness (percent/negation/cli equivalence)
• IR: FunctionGemma IR v5 + learned transcript router + transcript-first repairs

LoudInk honest (minimal polish)

Suite	v0.5.0	v0.6.0
Base (10 cases)	100%	100.0%
Extended full (64 cases)	100%	100.0%
Mega regression (723 cases)	75.4%	93.5%

v0.6.0 promotes SFT v5 app-neural (0000048_app_neural_v5) with polish v2 and scorer fixes. SFT v6 weights were trained but not promoted (native gate regression).

Shipped-polish checkpoint (0000096_shipped_polish) preserved for harness-style eval.

See inference_config.json and manifest.json for integration fields.