loudink-v1.1 1.1.0: skills PR-A–E canary (record/replay, slots, expanded daily-mac)
d506628 verified
loudink-v1.1 status — PR-A→E complete
Last updated: 2026-07-08
Hugging Face
Rebuild canary:
python -m gemmaflow_tune.cli.build_loudink_v1_1_publish --force
python -m gemmaflow_tune.cli.publish_sota --model loudink-v1.1 \
--commit-message "loudink-v1.1 1.1.0: skills PR-A–E canary"
PR board
| PR |
Deliverable |
Status |
| PR-A |
Skill store + recorder schema + compiler + null-safe IR |
done |
| PR-B |
Paraphrase invoke + skill bench + IR labels + runner wiring |
done (19/19) |
| PR-C |
Slot fill (for / about / named) |
done |
| PR-D |
Daily-mac expand (Focus/DND, Reminders, reply, tabs, Notion/Linear, writing) |
done (69/69) |
| PR-E |
Vision fallback scaffold (disabled by default) |
done |
Headline metrics (canary)
| Metric |
Bar |
Result |
| Skill invoke (n=19) |
≥95% named / ≥85% paraphrase |
100% shipped |
| Daily-Mac overall (n=69) |
no drop vs 0.4 |
100% shipped |
| Writing neural |
hold ≥80% |
92% |
| Writing p50 |
≤250 ms |
171 ms |
| Computer / multi-step |
hold 100% |
100% |
| IR honesty (n=112) |
≥45% neural |
46.4% (v0.4 lineage) |
Daily-Mac scorecard (v1.1 expanded)
| Slice |
Shipped |
Neural |
p50 |
| Writing (n=25) |
100% |
92% |
171 ms |
| Computer (n=44) |
100% |
— |
~0 ms FP |
| Multi-step (n=25) |
100% |
— |
~0 ms FP |
| Overall (n=69) |
100% |
— |
mean ~60 ms |
Scorecard: artifacts/loudink_v1/product_scorecard.json
Skill bench: artifacts/benchmarks/loudink_v1_1/skill_invoke/
Production stack (v1.1 canary)
writer: artifacts/sft_loudink_v1_rft_daily/adapters
polish: structural (+ daily rules)
ir: artifacts/sft_loudink_v1_1_skill_invoke_ir/adapters
ir_fallback: artifacts/sft_loudink_v1_ir_honesty/adapters
skills: SkillStore (example seed in publish bundle)
runner: loudink_v1 + skill_store_path
bench: configs/benchmark_loudink_v1_daily_mac.yaml
configs/benchmark_loudink_v1_1_skill_invoke.yaml
Out of scope (later product)
- Client recorder UI (Flowcast owns capture UX)
- Optional real vision weights/resolver
- AirDrop / Bluetooth / long-horizon 1:1 prep suites (roadmap backlog)
- Training cloud models on private user recordings (forbidden by default)
