outrigger: evidence-grounded long-horizon lab—and the blind gate that earned demotion
Jul–Sep Patel lab: spec interview + adversarial plan review before code; held-out sealed suites; exec-loop ledger. Three-arm result demoted blind merge gate (5.9× cost, same defect)—fresh adversarial eyes still pay.
Source: outrigger — experiments in evidence-grounded machinery for long-horizon coding agents
Dwijen Patel’s outrigger is deliberately a lab, not a product pitch: mechanisms for long-horizon coding agents enter only when graded evidence licenses them, carry dated conclusions, and get demoted on the record when measurements reverse the thesis. The working stack is CLI tools wired by schema-validated files—spec-interview (ratified machine-checkable plan), plan-preflight, heldout-suite (fails-on-base proof + tamper-evident seal), merge-gate, exec-loop (walk a ratified plan unattended), run-ledger (append-only predictions and null arms), shadow-pilot. The load-bearing early finding: spec ambiguity survives implementers, reviewers, and test authors alike; the cheap intercept is a planning interview plus an adversarial plan review before any code (live catch of a build-breaking seam, 2026-07-17). The productive complication that earned the lab its keep: a three-arm experiment (gated vs diligent vs frontier-solo, 2026-07-16) found the blind merge gate underperformed on well-specified work—all three arms shipped the same genuine defect, while the gate cost 5.9× the ungated arm—so the gate was demoted to a profile for weak specs or high stakes. Fresh adversarial eyes on every artifact remain the one mechanism no experiment managed to demote. Active skills graduated into the one-punch Claude Code plugin; outrigger stays the evidence base those skills cite.
Why it matters for code review
Always-on factories that bolt on a blind verifier because FOMO says “never trust the worker” can tax the night without catching the defect class that actually kills merges—ambiguous specs inherited by every role. Review overnight harnesses for outrigger’s discipline: dated warrants, null arms, and deletion criteria for checks that catch nothing; promote adversarial plan review and held-out seals where the evidence says they pay; demote expensive gates that fail to beat ungated diligent work on well-specified tasks. Steady harnessed throughput compounds when verification spend tracks measured defect classes; bursty mastery that adds another sealed suite without a ledger does not. Pair with AgentLoop’s held-out oracle and OvernightAgent’s four-gate pipeline when deciding which verification walls stay up overnight.