- Address the forward-looking twist: does the current trajectory imply parity, a re-widening as US labs ship reasoning-heavy successors on next-gen compute, or Chinese leadership in specific verticals (code, math, multilingual, open weights)?

As submitted# Conclave Prompt — The China–US Frontier AI Gap *Prepared for Matt Murray · Substrate Capital · July 7, 2026* --- ## Prompt (paste into Conclave) **Is it true that Chinese frontier AI models are now only about six months behind US frontier models, and if so, what actually explains how they closed the gap this fast?** Treat this as a single decidable claim on the gap, followed by a causal account. Do not hedge into "it depends." A verdict of "the question is malformed, here is the better question" is a legitimate outcome and should be argued for on the record if any advocate believes it. ### Definitions the room must accept before arguing - **"Frontier US models"** = the current flagship from OpenAI, Anthropic, and Google DeepMind as of the deliberation date. - **"Frontier Chinese models"** = the current flagship from DeepSeek, Alibaba (Qwen), Moonshot (Kimi), Zhipu (GLM), and ByteDance (Doubao/Seed). - **"Six months behind"** = the elapsed time between a US capability level being first shipped in a generally available model and a Chinese lab shipping an openly available model that matches it on a basket of public evals (MMLU-Pro, GPQA, SWE-bench Verified, LiveCodeBench, AIME, Arena-Hard, long-context retrieval). Cite specific model pairs and release dates. - The room may contest these definitions in Phase 1 but must adopt a shared working definition before Phase 2. ### The gap question — answer with a confidence interval, not a single number Is "~6 months" defensible today, optimistic (gap is smaller), or stale (gap has widened or closed further)? Give a range, and break it out by capability domain — text reasoning, code, math, long-context, multimodal, agentic/tool-use, open-weights leadership. A verdict of "6 months ± 3 months on text reasoning, 12–18 months on agentic workloads, ~0 months on open weights" is more honest than a single number. Anchor to named model pairs, not vibes. ### Structure the causal deliberation in two phases **Phase 1 — Independent nomination (Briefing → Submit stages).** Each of the six advocates must, before seeing the others' submissions, nominate the **two causes they believe are most load-bearing** for the gap-closure. Advocates are explicitly instructed not to converge on a canonical Western-analyst list. Constraints on the nomination round: - At least one advocate must argue the premise itself is wrong — either the gap is not ~6 months, or "gap" is the wrong frame entirely. - At least one advocate must nominate a cause that would not appear in a standard US think-tank readout. Steelman a Chinese-industry, hardware-supply-chain, or benchmarks-skeptic view. - One advocate is designated the **red-team seat** and must argue the strongest version of "the gap is a mirage — Chinese models look close on public benchmarks because those benchmarks are saturated, contaminated, or gameable via post-training, and the real frontier (long-horizon agents, tool use, novel scientific reasoning) still has a 12–24 month gap." **Phase 2 — Collate, contest, rank (Collate → Cross-examine → Debate stages).** Merge the nominated causes into a working set. Before producing any ranking, the room must: - Produce a **one-paragraph causal sketch** identifying which causes are independent drivers vs. downstream effects of other causes. A flat ranking hides this — a sketch surfaces it. - Flag any cause that appears in fewer than two advocates' nominations as a **minority hypothesis** and give it a dedicated defender before the room is allowed to dismiss it. - Explicitly name **what is NOT in the working set and why** — one sentence per omitted candidate. At minimum, the room must consider and either adopt or reject on the record: benchmark contamination and eval overfitting; Chinese-language and industrial/government data advantages; weaker IP and copyright friction on training data; US labs deliberately delaying frontier releases for commercial or safety-review reasons; differential product-shipping cadence and regulatory asymmetry; the possibility that "6 months" is a measurement artifact and there is no gap to explain in the domains that matter. ### Candidate causes — a floor, not a ceiling The room may add to this list, must justify omissions, and must not treat it as exhaustive. 1. **Algorithmic efficiency under compute constraint** — export controls forced MoE, MLA, aggressive distillation, RL-from-verifier pipelines (DeepSeek-V3/R1, Qwen3, Kimi K2 as evidence). 2. **Open-weights flywheel** — Chinese labs release weights, absorb global fine-tuning and red-teaming for free, and iterate on a public artifact while US frontier labs iterate behind an API. 3. **Distillation from US frontier outputs** — synthetic data, trajectory mining, and post-training on US model completions compress the R&D cycle. 4. **Talent density and state-adjacent capital** — returning PhDs, hyper-competitive domestic labor market, provincial/central subsidies, and university-lab pipelines (Tsinghua, PKU, USTC). 5. **Hardware substitution and stockpiling** — H800/H20 access before controls tightened, Huawei Ascend 910B/910C ramp, SMIC 7nm yields, and domestic HBM progress reducing the compute penalty. 6. **US self-inflicted drag** — safety review cycles, alignment tax, RLHF conservatism, product-safety gating, and API-only distribution slowing observable capability release even when internal capability is higher. ### Constraints on the deliberation throughout - Every claim must be tied to a specific model, benchmark result, paper, or dated event. No abstract "China is investing heavily" filler. - Distinguish *closed capability gap* from *closed deployment gap*. A US lab sitting on a stronger internal model is not the same as parity. - Distinguish *frontier* from *open-weights leadership*. China clearly leads open weights; the question is the *frontier* delta. - Distinguish *root causes* from *amplifiers* when ranking. An amplifier that would collapse without its root cause is not itself load-bearing. - Address the forward-looking twist: does the current trajectory imply parity, a re-widening as US labs ship reasoning-heavy successors on next-gen compute, or Chinese leadership in specific verticals (code, math, multilingual, open weights)? ### What the majority verdict must contain 1. A **confidence interval on the gap**, broken out by capability domain, anchored to specific model pairs. 2. **A causal sketch** (one paragraph) distinguishing root causes from downstream effects and amplifiers. 3. **A ranked list of root causes only**, with the top two carrying the load and any demoted or rejected candidates named explicitly — including candidates from outside the six-item floor if the room nominated them. 4. **A red-team dissent paragraph** if the red-team advocate did not carry the majority, stating the strongest surviving objection to the verdict. 5. **One concrete falsifier** — a model release, benchmark result, export-control action, or hardware milestone in the next 6–12 months that would force the room to revise the verdict. Do not produce an executive summary of "considerations." Produce a decision, with its uncertainty made legible. --- ## Operational notes - Conclave hardcodes T6 (full 6 advocates + 5 judges + fresh-eyes audit) regardless of length, so the two-phase structure has enough runway to actually execute. See [Conclave README, submit pipeline](https://github.com/mdm-sfo/conclave/blob/main/README.md). - The prompt is long but well under any practical limit — the composition page's Prompt Assist will offer to rewrite it. Decline. The structure is doing load-bearing work and a rewrite will flatten it. - Named model pairs and dated releases trigger the Perplexity Sonar pre-flight grounding path in `routers/submit.py`, so the worker prepends a "FACTS — do not contradict" block and reduces hallucinated release dates. ## Design rationale (why the prompt is shaped this way) - **Single decidable claim + causal sketch** matches Conclave's convergence instrument — the agreement score only moves meaningfully when advocates are staking positions on the same proposition. - **Definitions block up front** prevents the 8-stage debate from burning Round 1 on semantics; the Briefing → Submit stages then share a frame and the Judge phase has objective criteria. - **Open nomination round (Phase 1)** treats the six-item list as a floor rather than a ceiling, mitigating the closed-set bias where advocates rank rather than discover. - **Causal sketch before ranking** surfaces entanglement (export controls → algorithmic efficiency; open weights → distillation) that a flat ranking would hide. - **Red-team seat and explicit "what is NOT in the set"** force the fresh-eyes audit (stage 7) to bite rather than rubber-stamp consensus. - **Falsifier requirement** produces a verdict that ages well and gives you something to re-litigate in six months.