User reassessed model needs: 26b is overkill for slice-scope tasks
(short Korean title + 3-line summary + ≤3 kebab-case tags). e4b
(8.0B Q4_K_M, efficiency variant) sits closest to the original
"표준 9b" tier in spec §6.5 and gives faster responses with
adequate margin for the slice's simple structuring work.
Spec changes:
- Header bumped to v0.4 with new revision line.
- §1.1 retiers AI 파이프라인 from "고품질 26B" to "표준 e4b".
- §2.1 diagram, §3.1 ai_provider example, §4.1 provider name,
§4.2 request body, §4.6 health check, §5.4 copy catalog: model
string swapped 26b → e4b.
- §6.4 perf target restored to <30s provisional (e4b is similar
scale to original 9b reference, 60s margin no longer needed).
- §8 open issue rewritten: e4b is the standard-tier substitute,
with 26b/e2b documented as quality-up / speed-up fallback paths
if dogfood reveals issues with Korean output or tag formatting.
Plan changes:
- Header bumped to v0.4 with restated Goal.
- Architecture summary, Task 12 unit test mocks/expectations,
LocalOllamaProvider default model, copy banner, Task 33 dogfood
prerequisite: 26b → e4b throughout.
Tier ladder for dogfood escalation (from §8):
e2b (5.1B, speed) ← e4b (8.0B, default) → 26b (25.8B, quality)
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
82137a5ce1