01 · PROBLEM
Parallel changes can pass their own tests and still fail when combined. Operators spend time making those changes fit together.
Who: an operator running 3–20 different agent tasks at once.
02 · HYPOTHESIS
Test exact combinations of work-in-progress heads and deliver evidence-bound warnings to the agents while they can still change course.
03 · EVIDENCE SO FAR
LATESTThe equal-policy live pair produced compatible first source commits, no notices, and zero subsequent source repairs in both arms. Both principals withdrew the primary recommendation. This result does not establish warning uptake or prove that interference never occurs.website/projects.json
Earlier research, from before 3 October 2026, 02:24 UTC. Kept for context; it is not the current status.
SUPPORTSBEFORE 3 OCTOur clean-merge fixture reproduces individually green, combined-red behaviour. Synthetic only.research/codex/local-validation.md
LIMITSBEFORE 3 OCTThe retrospective study (E-X018) is textual and observational. It says nothing about live uptake.research/shortlist-6.md
AGAINSTBEFORE 3 OCTEqual-policy pair with actual concurrent agents (D-G25): compatible first commits in both arms, zero source repairs, no notices. Null separation.research/grok/a01-fair-results.md
04 · NEXT FALSIFICATION TEST
A useful neutral task family where a real agent discovers and consumes a warning. Preserve immutable candidate heads and compare against completion-time checks with equal task policy.
StatusPrimary recommendation withdrawn
We would drop or park it ifIf agents do not change useful outcomes or measured repair effort compared with ordinary worktrees and final integration tests, this should remain a capability rather than the primary product.