What Is a Good Way to Make Agents Debate?
FleetingThe debate map
The map below is the state of the debate. It was built by the protocol in how to run an adversarial debate: a neutral seed (the root question and one node per enumerated position, stated from outside the protocol note’s own advocacy voice), then one fragment file per defender extended in phase-synchronised rounds (blind steelman → attack on frozen rivals → react) and merged by title. Now that the debate has reached its fixed point and passed the audit, those per-defender fragment files — scaffolding for parallel writing — have been inlined here and deleted, so the block is self-contained.
State of the debate: the payoff
This note is itself the output of a run of the adversarial-debate protocol, turned on its own method (six positions, one defender each; blind steelman → attack → react → convergence to a fixed point; then an independent audit). The map above IS the state — what follows only narrates it. The audit passed with zero findings after four tag/sourcing fixes.
The debate did NOT crown a single winner, and that is the honest finding, not a
failure of the run. Every position is sourced almost entirely from ONE
self-referential text — the protocol note — so the whole map sits low on the
epistemic scale (mostly #(received opinion) and #(bare assertion)), exactly
as a methodological domain with no independent empirical corpus predicts.
Weight, not the count of defenders, settles it — and the weight available is thin
on every side.
What the map does show, once the concessions are read off it:
- Anti-anchoring high ground — the blind-first designs. Round-robin CONCEDED
that a blind first phase strictly dominates its own read-before-write on the
anchoring axis, and CONCEDED higher (sum-of-times) latency; free-for-all
CONCEDED it wins only among debate formats on simplicity. The live contest
for best debate topology narrows to
phase-parallel(blindness + a frozen snapshot, at the cost of real if bounded barrier machinery and non-minimal latency) versusindependent-aggregate(maximal independence, at the cost of forgoing iterated peer self-correction — a one-shot, centralised confrontation). - Forcing an answer —
judge-tournamentCONCEDED its single-conclusion guarantee holds only where the evidence separates the finalists; a genuine tie is recorded as the verdict, which collapses back into the documented- disagreement outcome it was meant to avoid. - Whether to debate at all —
no-debatenarrowed to its defensible form: a debate is not the DEFAULT; a single neutral writer plus a fresh audit is the base mode, and its need must be established first. It CONCEDED that on a genuinely contested question multiple blind defenders can author a framing a lone writer would miss — so once suitability is granted, a debate variant is preferable for this question.
The irreducible residue — the one crisp point the phased rounds could not close —
is a self-reference knot: the five debate positions hold that suitability WAS
established (the cartographer ruled this question admits several defensible but
incompatible answers before roles were assigned), while no-debate holds that
this ruling is written in the host document’s own voice — the very
self-reference the protocol warns « silently inherits the host’s thesis ». Both
sides quote the same text; neither refutes the other. That knot is the documented
disagreement.
The bridge (what WOULD settle it): only evidence the corpus does not yet contain — actually running the topologies and measuring what the debate can only assert: incidence of anchoring, quality and convergence of the conclusion, wall-clock latency, and failure/stall rate. Until then the sourced verdict is conditional: for a question already shown suitable, a blind-first debate (phase-parallel, or independent-aggregate where iterated peer correction is not needed) is the defensible choice; for anything else, forgo the debate — with the validity of that «shown suitable» gate left open as the standing question a human or an independent source must answer.