Routing policy
Planning stays with Opus
The three fleet policies agree on Opus for the plan and planning-loop lead, and Fable for the fixed co-plan draft.
View policy snapshotPlayer select · 04 files
Claude · AnthropicClaude
← → Switch player
PLANNING + CO-PLANNING
Policy · Planning
Discovery85
Planning85
Build75
Review judgment45
Grounding75
Cost efficiency50
Editorial ratings · recomputed 2026-09-10. 70% evidence reading + 30% harness fit, rounded to five points.
Scoring breakdown & sources
Claude-family reading. Opus supplies the discovery/build baseline; Fable 5.1 supplies the newer review evidence. Models are named in each explanation.
Each input uses a 0–4 rubric. Evidence: negative (0), weak or indirect (1), mixed (2), strong within bounds (3), strong across tested cases (4). Fit: excluded (0), trial/reserve (1), bounded/additive (2), regular role (3), core role (4). Policy estimates are labeled below; these ratings are not measured success percentages.
Discovery 85
Evidence 3/4 · fit 4/4 · study + policy
Opus averaged 73.35 in the two discovery comparisons, with strong anchoring and retained deep-synthesis duties. GLM won pooled breadth, so this is below the top evidence band.
Planning 85
Evidence 3/4 · fit 4/4 · policy estimate
Policy-based estimate: Opus owns plan and planloop leadership; Fable is the fixed co-plan drafter. This records established responsibility, not a measured planning win rate.
Build 75
Evidence 3/4 · fit 3/4 · study + policy
Opus’s shipped W3-EMP build scored 85/100. Opus remains a build-lead candidate, but shares the route with Terra.
Review judgment 45
Evidence 2/4 · fit 1/4 · study + policy
Fable 5.1 was strong at FIND and bounded DECLARER work, but recalled only one of three ITERATE musts and lacks a dedicated confirm-seat test. Current review-lead routes list Terra.
Grounding 75
Evidence 3/4 · fit 3/4 · study + policy
Opus’s discovery anchoring was strong; Fable’s eight cells had no fabricated citations, with one immaterial location error. These support a strong scoped reading, not a family-wide zero-error claim.
Cost efficiency 50
Evidence 2/4 · fit 2/4 · study reading
Relative resource-use estimate: the lead study shifted lead-token load off Claude, while Fable xhigh used about 1.3–1.5× high’s tokens. Comparable billed cost per successful task is unavailable, so neither gets a top efficiency band.
Published study
Opus baseline: 85/100
The shipped W3-EMP implementation scored 85/100 in the lead-seat comparison. This score applies to that build.
Read studyPublished study
Fable 5.1: eight frozen cells
The September study supports FIND and bounded DECLARER work. A dedicated confirm-seat cell is still missing; the result alone does not change routing.
Read studyStudy results describe the named model and test. Roles reflect the policy checked on 2026-09-10.
Assignment · checked 2026-09-10
Planning + co-planning
Opus leads planning and the planning loop, with deep discovery work. Fable is the fixed co-plan drafter. Opus is also eligible for the build lead; review and cleanup lead routes currently list Terra.
Claude is a family: Opus, Fable and Sonnet have different assignments.
Configured seatsPolicy snapshot
Build leadeligibleOpus is a candidate alongside Terra; the router selects by pool availability and burn. Source ↗
Rote build workerconfiguredSonnet 5 is the static worker default; eligible bounded steps can route to GLM. Source ↗
Review + cleanup leadexcludedThose lead routes currently list Terra only. This does not rule out future Claude routing. Source ↗
Published evidence
Select player01 / 04