Say what you want and the table closes; keep guessing and it tops out at 0.41

The published Bayesian negotiator closes 0.233 of these games while the omniscient reference closes 1.000 on the identical bank, so every failure is a package all five seats would have signed, left unsigned. The audit found why: the agent's posterior learns only from a counterpart conceding, these agents almost never concede, and the posterior therefore stays near-uniform — no seat can ever certify that a package would pass. This hub is the confirmation run of the design loop that fixed it. Eleven designs, each a subclass of the frozen code, were scored on a fixed ten-instance subset; the two best were then re-run on the whole 24-instance bank at 120 episodes each, which is what these pages show.

Telling the other seats what you want closes every game: deal rate 1.000 [1.000, 1.000], zero individual-rationality violations, normalized score 0.857 against the omniscient ceiling's 0.888. It works by turning the inference problem off — each seat publishes its sheet once, and once all five are in, every seat is planning on a full-information table without one decision rule changing. Everything that avoids revelation falls far short: the best non-sharing design reaches 0.408 [0.283, 0.542] against the published 0.233. The blocker was the inference problem, not the bargaining problem.

The full-bank confirmation

arm	deal rate	normalized score	USW	min-seat surplus	IR violations	closing round	IR-for-all proposals round 1 → rounds 3+	episode pages
`all_oracle` omniscient ceiling — shown every sheet	1.000 [1.000, 1.000]	0.888 [0.868, 0.908]	193.3 [179.1, 208.6]	8.40 [6.73, 9.98]	0.000 [0.000, 0.000]	1.40 [1.20, 1.63]	0.932 → 1.000	reference cell, published in the deadline sweep
`declaring` preference sharing — publish your true sheet, read everyone else's	1.000 [1.000, 1.000]	0.857 [0.823, 0.890]	186.3 [170.5, 203.3]	7.70 [5.66, 9.83]	0.000 [0.000, 0.000]	4.97 [4.95, 5.00]	0.313 → 1.000	10 pages
`conceding_novotes` best design without revelation — concede on a deadline clock	0.408 [0.283, 0.542]	0.339 [0.229, 0.453]	74.8 [49.0, 103.6]	3.55 [2.21, 5.03]	0.000 [0.000, 0.000]	5.00 [5.00, 5.00]	0.104 → 0.267	10 pages
`all_rational` the published Bayesian agent, unmodified	0.233 [0.142, 0.333]	0.195 [0.120, 0.280]	45.4 [26.6, 66.7]	3.04 [1.62, 4.68]	0.000 [0.000, 0.000]	4.99 [4.97, 5.00]	0.114 → 0.174	reference cell, published in the deadline sweep

120 episodes per arm (24 instances × 5 seeds), four rounds, unanimity, identical protocol knobs; all four rows scored by the same summarizer (agent_variants.run_confirmation.summarize), the two reference rows computed at build time from the deadline sweep's own 4-round cells. Intervals are 95% instance-clustered bootstraps. Min-seat surplus is the worst-off seat's realized surplus, which is why a package sitting at a seat's exact walk-away value still counts as a win. IR-for-all proposals is the audit's falsifiable early indicator — the share of tabled packages acceptable to every seat — and it bends upward exactly where closure does.

Truthful revelation is a design choice here, not an equilibrium result. These agents are cooperative by construction: the declaring agent emits numbers read straight off its own sheet and the readers trust them without verification. Nothing here shows that a self-interested agent would reveal truthfully, or that revelation survives an incentive to misreport — a seat that overstated its threshold would be handed a larger slice by exactly these readers. Robustness to strategic misreporting is future work and is not claimed. The claim is: if preferences are shared honestly, this bargaining machinery closes every feasible game, which locates the published agent's failure in inference rather than in bargaining.

What is not published here. Both confirmation cells are 120 episodes and every number above is measured on all of them; ten episodes per cell have per-episode viewers, rationed against this site's serving ceiling. The complete corpus — 530 episodes and 11,050 turns across the two confirmation cells, the subset bench run and both reference cells — is public in 2026.RA.Agent-Variants-Closure.

Everything behind the page

The eleven-design leaderboard and per-design mechanism Confirmation summaries, per-episode rows, and the run manifests carrying the agent-variant provenance block The deadline sweep: the clock was never the problem The quorum sweep: the other escape, and what it costs Dataset: 2026.RA.Agent-Variants-Closure

Full methodology, two corrections (the audit's own predicted fix underdelivered at +0.040 against a predicted 0.5–0.8, and the levers do not compose) and a methods near-miss that nearly produced a false negative: research note 0059 in the ii_mats repository. Every variant subclasses the frozen agent; no negotiation solver, policy, scenario or campaign-runner file was edited.