[Agora] Dynamic debate round count - stop when stability detected done

Effort: thorough

Goal

SciDEXOrchestrator.run_debate (agent.py:1420 + scidex/agora/scidex_orchestrator.py) and the artifact-debate fan-outs
all run a fixed number of rounds (commonly 4: theorist → skeptic →
domain_expert → synthesizer). Many debates converge by round 2; others
genuinely need 6+. A fixed N over-spends compute on easy questions and
under-resolves the hard ones. Build a stability-aware round controller
that watches the inter-round content for convergence (semantic similarity,
verdict-vote stability, no new claims emerging) and decides each round
whether to stop, continue, or escalate to a fresh persona — bounded by min_rounds and max_rounds.

Acceptance Criteria

class RoundController with should_continue(state) -> Decision,
where Decision ∈ {stop_converged, continue_baseline,
escalate_new_persona, stop_max_rounds, stop_safety}. 1. Semantic stability: cosine sim of last-round content vs
prior round ≥ 0.9 (use existing scidex/core/embedding.py).
2. Verdict-vote stability: extracted per-round verdict (re-uses
aggregate_debate_consensus_dissent's consensus parser) is
identical to the previous round's verdict.
3. Claim-novelty floor: no new ≥-1-line claims emerged
(tokenised diff vs prior round; threshold 1 new claim). cosine sim BUT verdict is split — that's stable disagreement,
and the controller picks one new persona via the existing
specialist matcher (scidex.agents.select) to break the tie. configurable per-debate-type). Hard safety: stop when token budget
consumed > 80% of budget. scidex/agora/scidex_orchestrator.py:1843 swap their fixed
for round_idx in range(num_rounds): loop for
while controller.should_continue(state) != stop_*:. decision TEXT, signal_json JSONB, decided_at TIMESTAMPTZ,
PRIMARY KEY (session_id, round_index)). Migration
migrations/20260428_debate_round_decisions.sql. written at finish — interface is preserved. trace for one debate; surfaced as a "Why did this debate stop at
round X?" widget on /debate/{id}. (a) all three signals satisfied at round 3 → stop_converged,
num_rounds=3.
(b) min_rounds floor: signals satisfied at round 1 → still goes
to round 2.
(c) escalation: split verdict + high sim → returns
escalate_new_persona, controller integrates the new persona.
(d) max_rounds ceiling: never converges → stops at 8.
(e) token-budget safety: ceiling fires before max_rounds. runs the controller alongside the fixed loop and writes its
counterfactual decision to debate_round_decisions without
changing actual round count. After 1 week, flip to enabled. has a debate_round_decisions trace; spot-check one decision
against the round content.

Approach

Read scidex/agora/scidex_orchestrator.py:1843 for the canonical

round-loop shape; introduce the controller as a thin object so the
loop reads as while controller.next() == continue: ....

Embedding helper exists; semantic sim is one call. Verdict parser

already exists in synthesis_engine.py. Claim-novelty diff is
tokenised set difference — keep simple.

Persistence first; integration after; rollout in shadow mode.

Add a cost_savings_estimate rollup to the dashboard so we can

prove the controller pays for itself.

Dependencies

• scidex/agora/synthesis_engine.py — verdict parser.
• scidex/core/embedding.py — semantic sim.
• scidex.agents.select — escalation persona picker.

Dependents

• q-debate-evidence-weighted-vote — voting layer benefits from

stable-round signal as a confidence weight.

• q-debate-judge-interruption — the judge's mid-round halt API and

this controller share state plumbing.

Work Log