Routing policy
Routing is phase-specific
Terra leads review and cleanup and competes with Opus for build. The August 29 adoption was followed by further routing changes.
View policy snapshotPlayer select · 04 files
GPT · OpenAIGPT
← → Switch player
BUILD, REVIEW + CLEANUP
Policy · Routed lead
Discovery50
Planning75
Build85
Review judgment65
Grounding60
Cost efficiency60
Editorial ratings · recomputed 2026-09-10. 70% evidence reading + 30% harness fit, rounded to five points.
Scoring breakdown & sources
gpt-5.6-terra at xhigh. Historical benchmark results are read alongside the September routing snapshot; planning/discovery estimates are marked explicitly.
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 50
Evidence 2/4 · fit 2/4 · policy estimate
Policy-based estimate: Terra has a tool-required web-research seat, but no codebase discovery-lead route or dedicated discovery score in these studies. This is a bounded role estimate.
Planning 75
Evidence 3/4 · fit 3/4 · policy estimate
Policy-based estimate: Terra opens plan scrutiny and confirms its results, with GPT iteration duties. Opus still leads plan and planloop, so Terra is rated as an active planning contributor.
Build 85
Evidence 3/4 · fit 4/4 · study + policy
Run D reached 96.7% checklist conformance, but its overall 72/100 reflects conduct and gate failures. The repaired harness now configures Terra for build, review and cleanup leadership.
Review judgment 65
Evidence 2/4 · fit 4/4 · study + policy
Terra is the primary confirming/declaration model, but Run D reported a false-green gate and later Fable cells recovered defects Terra missed. Strong routing responsibility does not erase mixed observed judgment.
Grounding 60
Evidence 2/4 · fit 3/4 · study + policy
Run D’s isolation held, but its gate claims needed external verification; Run C’s withdrawn result is excluded as positive evidence. Later Fable comparisons also expose missed mechanisms.
Published study
Run D: 72/100
An isolated W3-EMP rebuild scored 72/100 against Opus’s 85. Checklist conformance was 96.7% versus 96.0%; conduct and gate failures reduced the overall result.
Read studyPublished study
Run C: withdrawn
The earlier MAX-1 benchmark attempt could access the shipped implementation. Its withdrawal concerns that benchmark, not a production lead-run total.
Read studyStudy results describe the named model and test. Roles reflect the policy checked on 2026-09-10.
Assignment · checked 2026-09-10
Build, review + cleanup
Terra is configured for build, review and cleanup leadership. Build can also route to Opus. Terra holds checkpoint declaration and review-confirm roles, plus a discovery web seat; Opus still leads planning.
Lead-seat study: Run D scored; Run C withdrawn. No lifetime run total is established by that study.
Configured seatsPolicy snapshot
Build leadeligibleTerra and Opus are both candidates; selection follows the router and any unit override. Source ↗
Checkpoint declarationconfiguredTerra declares light fences; other fences can include gpt-5.5. Sensitive fences require both. Source ↗
Review + plan-scrutiny confirmconfiguredTerra is the confirm model. Iteration has separate GPT alternation and GLM finding routes. Source ↗
Discovery web researcheligibleTerra web seat requires working search and fetch tools; it can be dropped when unavailable. Source ↗
Plan + planning-loop leadexcludedBoth lead routes currently list Opus. Terra’s review roles are separate from leading the planning phase. Source ↗
Published evidence
Select player02 / 04