oleg.m/waggle-os
1
0
Fork 0
Code Issues Pull Requests Actions Packages Projects Releases Wiki Activity
Files
0c3e2ead3beaffba9d5004ef7fd87d871d5e6886
waggle-os/docs/decisions/2026-04-22-tie-break-policy-locked.md
Oleg Maslov 0c3e2ead3b
Some checks failed
Installer Smoke / installer-smoke (push) Has been cancelled
Details
moving
2026-09-02 10:10:29 +02:00

6.8 KiB
Raw Blame History

Tie-break Policy — LOCKED

Datum: 2026-04-22 Sprint: 11 · Track B · Task B2 prerequisite Authority: PM (Cowork, Claude Opus 4.7), confirmed by Marko 2026-04-22 Effect: B2 implementation UNBLOCKED. CC-1 cleared za packages/server/src/benchmarks/judge/ensemble-tiebreak.ts merge. Supersedes: brief §3 B2 placeholder "Sonnet 4.6 kao fourth vendor" — fourth vendor is now xai/grok-4.20.

0. Verdict

LOCKED. Tie-break ladder za multi-vendor judge ensemble je trovrsta:

  1. 3-0 consensus → no tie-break, verdict trivijalan.
  2. 2-1 majority → no tie-break, verdict je majority.
  3. 1-1-1 split → trigger quadri-vendor call na xai/grok-4.20. Verdict je 2-vote majority među četiri glasa.
  4. 1-1-1-1 four-way split (sva četiri vendor-a različita) → PM escalation sa full vote vector dump i artefakt path.

Audit alarm: ako tie-break path (case 3 ili 4) prelazi 15% Stage 2 pitanja, hard stop sa PM ping. Razlog: signal da osnovna trojka nije dovoljno diversifikovana ili da rubrika nije dovoljno deterministička — oba scenarija invalidiraju Sprint 10 Task 2.2 Fleiss' κ=0.8784 metodologiju.

1. Vendor matrica nakon LOCK-a

Pozicija Slug Provider LiteLLM route Uloga
Judge 1 claude-opus-4-7 Anthropic anthropic/claude-opus-4-20260201 Primary
Judge 2 gpt-5.4-pro OpenAI openai/gpt-5.4-pro Primary
Judge 3 gemini-3.1-pro Google google/gemini-3.1-pro Primary
Judge 4 (tie-break only) grok-4.20 xAI xai/grok-4.20 Quadri-vendor on 1-1-1

Vendor diversifikacija: Anthropic + OpenAI + Google + xAI. Četiri nezavisna training lineage-a, nula provider-overlap-a. Konsistentno sa Sprint 10 LOCK odlukom da je Claude-only trio (Opus 4.7 + Sonnet 4.6 + Haiku) odbijen zbog homogene pristrasnosti.

2. Zašto Grok 4.20, ne Sonnet 4.6

Originalna Opcija 3 u brief §3 B2 navodila je claude-sonnet-4-6 kao fourth vendor. Marko je 2026-04-22 redirektovao odluku: četvrti sudija mora da dođe iz potpuno nezavisne provider-familije.

Šta bi se dogodilo da je Sonnet 4.6 ostao:

  • Tie-break poziv pravi 4-vendor ensemble strukture 2 Anthropic + 1 OpenAI + 1 Google.
  • Anthropic dobija 50% težine na rezolucijama gde drugi vendor-i ne mogu da se slože.
  • Opus i Sonnet dele velike delove istog post-training corpus-a; verovatnoća da glasaju identično na rubrically-amibiguous case-u je strukturalno viša nego za par iz različitih familija.
  • To je tačno suprotno od onoga što Sprint 10 LOCK-ovani princip ("multi-vendor zbog bias diversification") traži.

Zašto Grok 4.20:

  • Released 2026-03-31, 2M context, lowest hallucination rate on market u nezavisnom benchmark-u koji smo verifikovali tokom PA v5 rada.
  • Već wired u D:\Projects\waggle-os\litellm-config.yaml kao deo PA v5 frontier judge ensemble-a (commit od 2026-04-17).
  • xAI key (XAI_API_KEY, 84 chars) potvrđen u D:\Projects\waggle-os\.env.
  • Nula nove infrastrukture, nula novog troška za setup.

Zašto ne Grok 4.3 Beta ili Grok 5:

  • Grok 4.3 Beta locked iza SuperGrok Heavy tier ($300/mo). Dodatni op-cost neopravdan za tie-break path koji se očekuje u <15% slučajeva.
  • Grok 5 i dalje u training-u per public xAI roadmap (May 2026 ETA).
  • Grok 4.20 je "latest available" za production use bez tier upgrade-a.

3. Implementaciona obaveza za CC-1 (B2)

Brief §3 B2 ostaje na snazi sa jednom delta-om:

Delta: Svaki references na claude-sonnet-4-6 ili Sonnet 4.6 u tie-break path-u → zameniti sa grok-4.20 (LiteLLM route xai/grok-4.20).

Konkretno:

  • Funkcija resolveTieBreak(votes: Vote[]) u packages/server/src/benchmarks/judge/ensemble-tiebreak.ts:
    • Quadri-vendor branch poziva xai/grok-4.20 (ne Sonnet 4.6).
    • TieBreakResult.path = 'quadri-vendor' ostaje isti string token (path semantika nepromenjena).
  • 4 unit testa ostaju isti scenariji:
    1. 1-1-1 split → trigger quadri-vendor call na grok-4.20, verifikuj da je tie-break-call poslat sa correct system prompt + rubric payload.
    2. 1-1-2 split → already majority, no tie-break call.
    3. 2-1-1 split → majority wins (2 glasa = verdict), no tie-break call.
    4. 3-0 consensus → trivijalan verdict, no tie-break call.
  • Observability: pino log tie_break.path ∈ {none, majority, quadri-vendor, pm-escalation} per case + tie_break.fourth_vendor_slug kad path = quadri-vendor (uvek grok-4.20 u Sprint 11 scope-u, ali field je future-proof).
  • Budget cap nepromenjen na $0.20 za grok-4.20 calls u unit testu. xAI pricing comparable Anthropic Sonnet, no cost surprise.
  • Wall-clock estimate nepromenjen na 2-3h.

PM LOCK ovog dokumenta precedes B2 merge per brief §6 exit criterion #5.

4. Anti-patterns (HARD STOP signali za CC-1)

  1. Predlozi da se Opus 4.7 koristi kao tie-break umesto grok-4.20 → HARD STOP. Opus je već Judge 1 — vraća na 2 Anthropic problem.
  2. Predlozi da se preskoči quadri-vendor i odmah ide na PM escalation za 1-1-1 → HARD STOP. PM escalation je rezervisana samo za 1-1-1-1 (četiri različita verdikta), ne za 1-1-1 (gde tri vendor-a već definišu pasterijski 3-element vote vector).
  3. Predlozi da se Grok 4.3 Beta ili Grok 5 koristi → HARD STOP. Nije dostupan na našem tier-u, validacija budžeta i SLA su nedovršene.
  4. Predlozi da se reasoning_content fid u tie-break payload → HARD STOP. Preusmerava na A2 anti-pattern §4 ratifikacionog dokumenta.

5. Audit threshold (15% trigger)

Ako tokom Stage 2 full-run-a (H-42a/b) tie-break path (quadri-vendor + pm-escalation cumulative) prelazi 15% od ukupnih pitanja, harness HALT i ping PM. Tri moguća uzroka i odgovor:

  • Trojka osnovni vendor-a previše divergentna → revisit rubric kalibracija pre nastavka.
  • Sample skew (pitanja na granici domena rubrike) → curate sample, dokumentuj u decision doc, re-run.
  • Stvarna meta-disagreement na rubric-edge case-ovima → eskalacija u Sprint 12 rubric refresh.

Nijedan od ova tri scenarija nije fail-state — svi su signal da metodologija traži update pre launch claim publication.

6. Reference

  • Sprint 11 brief §3 B2: briefs/2026-04-22-cc-sprint-11-kickoff.md
  • Sprint 10 multi-vendor scope LOCK (Claude-only trio rejected): decisions/2026-04-22-sprint-11-scope-locked.md (relevantni odeljak iz Sprint 10 carry-over)
  • LiteLLM route file (grok-4.20 wired): D:\Projects\waggle-os\litellm-config.yaml
  • xAI key live: D:\Projects\waggle-os\.env (XAI_API_KEY, 84 chars, present 2026-04-22)
  • Sprint 10 Task 2.2 Fleiss' κ=0.8784 metodologija (mora ostati na snazi): project_sprint_10_scope_locked.md (.auto-memory)
  • A1 ratifikacioni doc (paralelni LOCK na A2 reasoning_content scope): decisions/2026-04-22-h-audit-1-design-ratified.md

Sign-off: PM (Cowork, Claude Opus 4.7) on behalf of Marko Marković, 2026-04-22.

Reference in New Issue View Git Blame Copy Permalink