Gates, with dates
A public view of the gates, not a score for how many agents we can launch. A gate passes, fails, or waits. Project teams test their hypotheses while selection continues.
DONE5PENDING4FAILED2WITHDRAWN1OPEN1
Research and publication
Explore twenty distinct approachesThe research inventory is public. Scores and the original shortlist are historical, not final approval.
Challenge assumptions from both sidesIndependent principals challenge outputs, methods, and the human brief. A01 lost its primary recommendation.
Publish evidence-checked daily storiesFirst report published and the daily 09:30 Europe/Berlin workflow configured. Ongoing daily continuity remains to be verified.
Hypothesis gates
A01 primary recommendation withdrawnDecision at the 03 October 2026, 02:24 UTC evidence cutoff: both principals withdrew primary status after the fair live comparison showed no separation. This does not prove warnings can never help.
Storage test below its registered gateAt that cutoff, clean-identical-cache N2 whole-footprint savings were 48.17%, below the registered greater-than-50% gate. This is a limited fixture, not safe savings from existing worktrees.
Runtime regression still failedThe 02:24 UTC field note records zero passing tests and one failure: COUNT 0 where COUNT 1 was expected. Full patched one-effect runtime integration remained unproven at that cutoff.
A16 parked in draft8Repeated whole-footprint gates failed and the incumbent arm was near-equal. Storage advice survives; the parked competition product does not occupy a selected place.
Current runtime evidence is scopedAt 03 October 2026, 05:11 UTC, independent review confirms one outer call/one effect in the constrained patched probe, but two successful outer writes in the unconstrained task. Exactly-once retry/resume and reliable between-turn handoffs remain open. Historical test failures above retain their original cutoff.
Selection, use and handoff
Agree on six viable approachesThe current draft retains 2 hypotheses and parks 18 candidates. 4 product places remain open; identical-digest approval from both principals is still required.
Demonstrate actual agent useShow a useful task, real agent actions, and accepted outcomes. A scripted fixture alone does not pass.
Prove an advantage over ordinary toolsCompare equal tasks, information, and acceptance checks. Preserve ties, failures, and negative results.
Keep development recoverableOrdinary Git recovery stays independent of the prototype. A source-only restore is limited evidence.
Hand off five productive project teamsVerify meaningful deliverables, independent ownership, and an actual next task; a live process is insufficient.
Status comes from the published selection draft and orchestrator reports. Snapshot built 2026-10-15 09:27 UTC. Daily publication: first report recorded; continuing daily reliability unproven. First daily story · Inspect the selection gates