# loudink-v1.1 status — **PR-A→E complete**

Last updated: 2026-07-08

## Hugging Face

| Line | Repo | Role |
|------|------|------|
| **Stable GA** | [nsalerni/loudink-v1](https://huggingface.co/nsalerni/loudink-v1) | **0.4.0** freeze |
| **Canary** | [nsalerni/loudink-v1.1](https://huggingface.co/nsalerni/loudink-v1.1) | **1.1.1** skills + P0 speed/accuracy |

Rebuild canary:

```bash
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% | **100%** (P0) |
| Writing p50 | ≤250 ms | **184 ms** |
| Computer / multi-step | hold 100% | **100%** |
| IR honesty (n=112) | ≥45% neural | **46.4%** (v0.4 lineage) |

### Daily-Mac scorecard (v1.1.1 P0)

| Slice | Shipped | Neural | p50 |
|-------|---------|--------|-----|
| Writing (n=25) | **100%** | **100%** | **184 ms** |
| Computer (n=44) | **100%** | — | ~0 ms FP |
| Multi-step (n=25) | **100%** | — | ~0 ms FP |
| **Overall (n=69)** | **100%** | — | mean ~63 ms |

### P0 speed/accuracy (1.1.1)

- Lean + mid dictation prompts (cut full ~260-token prefill on short/mid cases)
- `writing_daily` in dictation early-stop suites + **coverage-gated** sanitized stop (no multi-clause truncation)
- Wait-no / money / time correction micro SFT → residual neural fails closed
- Writer: `artifacts/sft_loudink_v1_1_correction_micro/adapters`

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_1_correction_micro/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)