V1_1_COMPLETE.md

loudink-v1.1 complete (PR-A → PR-E)

Version: 1.1.0 canary
Date: 2026-07-08
Stable freeze left intact: loudink-v1 0.4.0 at nsalerni/loudink-v1

Delivered

PR-A — Skill substrate

• src/gemmaflow_tune/skills/{schema,store,recorder,compile}.py
• Null-safe run_skill compile + sanitize (never invent skill_id)
• Unit tests: tests/test_skills_pra.py

PR-B — Invoke surface

• Paraphrase + exact alias match in skills/invoke.py
• Freeform disambiguation (open gmail ≠ skill hijack)
• Bench suite n=19 + IR SFT corpus
• Runner wiring: skill_store_path via runner_factory
• Unit tests: tests/test_skills_prb.py

PR-C — Slot fill

• Deterministic extract in skills/slots.py
• Seed skills: Invoice Chase (company), Standup Post (update)
• Scorer: slot_values / skill_id
• Unit tests: tests/test_skills_prc_slots.py

PR-D — Daily-Mac expand

• Computer: Focus/DND, Reminders, Messages/Mail reply, Chrome tabs, Notion+Linear, Wi‑Fi
• Writing: reply iMessage, email subject, code cell
• Resolvers: focus/DND, browser tabs, reply context
• 69/69 overall on expanded suite

PR-E — Vision scaffold

• skills/vision.py: ScreenObservation, VisionFallbackConfig, try_vision_fallback
• Default disabled; client injects resolver when ready

Verification commands

.venv/bin/python -m pytest tests/test_skills_pra.py tests/test_skills_prb.py tests/test_skills_prc_slots.py -q
.venv/bin/python -m gemmaflow_tune.cli.prepare_skill_invoke_benchmarks
.venv/bin/python -m gemmaflow_tune.cli.prepare_daily_mac_benchmarks
GEMMAFLOW_GPU_MODE=shared .venv/bin/python -m gemmaflow_tune.cli.benchmark \
  --config configs/benchmark_loudink_v1_1_skill_invoke.yaml --variant loudink_v1_1_skill_invoke
GEMMAFLOW_GPU_MODE=shared .venv/bin/python -m gemmaflow_tune.cli.benchmark \
  --config configs/benchmark_loudink_v1_daily_mac.yaml --variant loudink_v1_daily_mac
.venv/bin/python -m gemmaflow_tune.cli.build_loudink_v1_product_scorecard \
  --measurement artifacts/benchmarks/loudink_v1/daily_mac/loudink_v1_daily_mac_measurement.json \
  --output artifacts/loudink_v1/product_scorecard.json
.venv/bin/python -m gemmaflow_tune.cli.build_loudink_v1_1_publish --force
.venv/bin/python -m gemmaflow_tune.cli.publish_sota --model loudink-v1.1

Success metrics (all met)

Metric	Bar	Result
Skill invoke accuracy	≥95%	100% (19/19)
Skill paraphrase	≥85%	100% (suite includes paraphrase)
Freeform regression	no drop vs 0.4	100% daily-mac
Writing neural / p50	hold 0.4 bars	92% / 171 ms
IR honesty	≥45%	46.4%