52 lines
3.2 KiB
Markdown
52 lines
3.2 KiB
Markdown
---
|
|
date: 2026-04-21
|
|
type: decision
|
|
status: LOCKED
|
|
sprint: 10
|
|
tasks_affected: 1.2, 1.3, 2.1, 2.2
|
|
---
|
|
|
|
# Sprint 10 — Task 1.2 PR #1 ratifikovan + Opus 4.6 audit DEFERRED
|
|
|
|
## Odluka 1 — PR #1 (Sonnet route repair) merge as-is
|
|
|
|
**Izabrana opcija:** Opcija 1 iz PM review note-a (`sessions/2026-04-21-sprint-10-task-1.2-pm-review.md` §Quick-decision-ask).
|
|
|
|
**Mapping:** `anthropic/claude-sonnet-4-6-20250514` → `anthropic/claude-sonnet-4-6` (plain alias).
|
|
|
|
**Rationale:**
|
|
- Live Anthropic docs verifikovani 2026-04-21 potvrđuju `claude-sonnet-4-6` i kao Claude API ID i kao alias za 4.6 familiju. Direktna zamena postoji.
|
|
- Opus-promote intermediary (Opcija 2) uvodi budget-asymmetry u Stage 2 run-ove ($15/MTok output na Opus vs $15/MTok output na Sonnet — isto, ali $75/MTok input-intensive paths različiti) i sekundarno pitanje "koji judge je zaista ocenio koji case" u multi-vendor ensemble-u. Nepotrebna kompleksnost za konfiguracijski defekt.
|
|
- Dated snapshot `-20250514` nikad nije bio valid za Sonnet 4.6 (pripada deprecated Sonnet 4 jednocifrenoj familiji koja se povlači 2026-06-15). Nema decommissioning timeline rizika, čisti config fix.
|
|
|
|
**Hedge:** `scripts/smoke-sonnet-route.mjs` + `ops/litellm/README.md` migration note ostaju kao regresioni hedge — fail-loud ako neko u budućem sprintu opet zakači dated snapshot.
|
|
|
|
**Unblocks:** Task 1.3 (Sonnet kalibracija re-run, ~$0.50, 10 Sprint-9 tripleta).
|
|
|
|
## Odluka 2 — Opus 4.6 dated-snapshot audit DEFERRED
|
|
|
|
**Izabrana opcija:** Ne otvaramo audit PR u Sprint 10.
|
|
|
|
**Rationale:**
|
|
- Sprint 10 judge run-ovi koriste Opus 4.7 + Sonnet 4.6 + tri-vendor ensemble (Opus 4.7 + GPT-5.4 + Gemini 3.1). `claude-opus-4-6-20250610` se ne poziva u Sprint 10 code paths.
|
|
- Otvaranje audit PR-a je scope creep koji pogađa Workflow Reality Check anti-pattern — širenje scope-a zato što smo u tom folderu, ne zato što sprint-ciljevi to traže.
|
|
- Migration note u Task 1.2 PR-u već služi kao operativna memorija — kada sledeći caller padne na 404 sa Opus 4.6 dated snapshot-om, znamo gde je korenski uzrok.
|
|
|
|
**Policy:** Unblock-on-first-caller-trip. Ako bilo koji Sprint 10+ task padne na Opus 4.6 route, otvara se posebni audit PR tad — ne preventivno. BACKLOG tiket se kreira u hive-mind ili waggle-os backlog-u (task na CC).
|
|
|
|
## Day-2+ sequencing posle ove odluke
|
|
|
|
Per brief §6 + Day-1 status §6:
|
|
1. Merge PR #1 → Task 1.2 CLOSED.
|
|
2. `judge-calibration.mjs --judge-model claude-sonnet-4-6` na 10 tripleta → Task 1.3 CLOSED.
|
|
3. `judge-calibration.mjs --ensemble claude-opus-4-7,gpt-5.4,gemini-3.1-pro` na istih 10 tripleta → Task 2.1 baseline data.
|
|
4. Task 1.1 matrix driver scaffold (dry-run), real Qwen calls Day-3.
|
|
|
|
Task 1.4 (DashScope dual-route) čeka Marko classic sk-… key. Task 1.5 (harvest adapter) čeka Marko fresh Claude.ai export. Task 2.2 (Fleiss' kappa full baseline) čeka Marko 5 novih ground-truth tripleta.
|
|
|
|
## Hard rules koje ostaju netaknute
|
|
|
|
Per brief §9: nema landing copy, nema brand narrative, nema Stage 1/2 full-run, nema ensemble širenja van tri-vendor LOCKED seta, nema Stage-2 primary-judge lock na single Sonnet result.
|
|
|
|
LoCoMo threshold bands ostaju pre-registered: ≥91.6% NEW_SOTA / 85.0-91.5% SOTA_IN_LOCAL_FIRST / <85% GO_NOGO_REVIEW.
|