diff --git a/runs/frontier-phase-factorial-v0/experiment-card.md b/runs/frontier-phase-factorial-v0/experiment-card.md
new file mode 100644
index 0000000..f291ab7
--- /dev/null
+++ b/runs/frontier-phase-factorial-v0/experiment-card.md
@@ -0,0 +1,54 @@
+# 实验 EXP-SIMFID-PHASE-FACTORIAL：prefill-only 是否是 simulator ranking 的容易区间？
+
+> **状态：** 已批准，运行中（用户于 2026-07-17 明确要求先完成 235B mixed 与 30B prefill-only）
+
+## Claim 与决策
+
+- **Parent claim：** Frontier 的 config-ranking fidelity 由 workload execution phase 决定；prefill-only 可能比 decode/mixed 更容易由 isolated operator profiles 组合。
+- **目的：** 用跨 model 的 phase factorial 区分 phase-complexity explanation 与 model/runtime/profile-specific explanation。
+- **Competing hypotheses：** H-phase：prefill-only 是低状态反馈的 compatibility envelope，因此 30B prefill-only 也能达到低-regret ranking，而 decode/mixed 更容易失真；H-stack：30B 的失败主要来自 FA3/CUDA-graph/routing/profile composition 等 stack-specific mismatch，因此即使 prefill-only 也可能失败；H-margin：mixed 的 differential 可以错误，但只要 topology margin 足够仍会选对 top set。
+- **事前预测：** 若 H-phase 成立，30B prefill-only 满足 regret ≤5%、Kendall τ-b ≥0.8，并明显优于同模型 mixed；若 H-stack 成立，30B prefill-only 仍过不了 gate；若 H-margin 成立，235B mixed 可保持 top set，但会漏掉 MNS/MBT pair directions。
+- **判定规则：** 30B prefill-only fail → “prefill-only 是充分条件”被否证，下一步优先做 same-model execution-context ablation；30B prefill-only pass 且新增 decode-heavy mixed fail → 支持 phase hypothesis；30B prefill-only 与 235B mixed 都 pass → phase 不能单独解释，转向 margin-aware compatibility envelope。
+
+## Setup
+
+- **自变量：** model×phase：已有 Qwen3-235B-A22B-FP8 mixed；新增 Qwen3-30B-A3B BF16 prefill-only。
+- **控制变量：** dash0 H20、community vLLM 与各自 frozen Frontier profile、同一 config 内 real/sim 的 request shape、arrival lattice、SLO、prefix policy、MNS/MBT/TP 与随机种子。
+- **30B system context：** community vLLM 0.20.0+cu129，BF16 weights/activation/KV，TP∈{1,2,4}，MNS∈{8,16,32,64}，MBT=8192，chunked prefill on，prefix off；real 保留 runtime 默认 CUDA graph，Frontier profile-only 不做 E2E calibration。
+- **30B workload：** fixed ISL=2048、OSL=1，64 个不同 token-chain prompts，uniform open-loop QPS；fresh server per `(config, rate, round)`，target-rate warmup 与 measured requests 分离。
+- **30B SLO：** TTFT≤1256 ms，至少 61/64 requests 通过；primary score 为最大共同 tested feasible req/s / 实际 TP GPUs。
+- **235B baseline：** 已冻结 `ISL=2048, OSL=128`、8 configs、68 个 fresh-server anchors；primary sensitivity TTFT≤1256 ms、TPOT≤150 ms。
+- **Baselines：** real community vLLM；Frontier same-stack profile-only；historical Qwen30 mixed profile-only；historical frozen per-TP calibration 只作为 upper bound，不参与本 case 拟合。
+- **Metrics：** top set、worst tie-break regret、Kendall τ-b、exact/non-tied pair direction、anchor confusion、absolute capacity、TTFT p50/p95、real trial variance与GPU-hour。
+
+## 预期产物与 review
+
+- **预期数据：** 30B frozen simulator surface；real config-rate anchors；两模型 phase comparison table；failure mechanism breakdown。
+- **Figure prototype：** `mock-phase-factorial.png`；x=model×phase，左轴=worst regret，右侧 annotation=τ-b；虚线是 5% regret gate。mock 只表达可区分趋势，不进入结论。
+- **人工 review：** 已批准。用户要求 smoke 通过后推进实验；先做这两个 case，再根据 hypothesis verdict 扩展。
+- **Review 意见：** 不把 existing 235B mixed top-set match 隐藏掉；不把“全 config 并列”算作成功 hit；30B prefill-only 必须使用相同 primary ranking objective。
+
+## Benchmark design audit
+
+| Risk | Verdict | 处理 |
+|---|---|---|
+| Selective benchmarking | PASS for initial screen | 同时报已有 235B mixed success和内部 pairwise failure；后续 expansion 由预注册 verdict 触发 |
+| Simplified workload | NEEDS EVIDENCE | fixed-shape 只用于 phase isolation，不外推 trace-faithful mixed |
+| Calibration=evaluation | PASS | 新 case 不用 serving E2E 数据拟合 scale |
+| Missing significance | NEEDS EVIDENCE until run | boundary anchors做独立 fresh-server repeat，保留 disagreement |
+| Relative-only result | PASS by design | 同时报 req/s/GPU、TTFT distribution、rank/regret |
+
+## 复现信息
+
+- **Code：** AITuner branch `codex/fidelity-prefix-pilot-20260714`；Frontier upstream `d9cfeb6d8791fbf2f295dd9744c56a666171776e` + frozen known patches。
+- **Environment：** 只使用 dash0 8×H20；Qwen30 venv `/tmp/wjh/venvs/vllm-0.20.0-cu129-profiler-v1`；model `/home/admin/cpfs/wjh/models/Qwen/Qwen3-30B-A3B`。
+- **产物路径：** local/remote `runs/frontier-phase-factorial-v0/`；raw GPU artifacts 由 fleet harvest，condensed JSON/CSV 进入结果目录。
+- **已知 deviation：** 235B 为 FP8/vLLM0.10.2/FlashInfer eager，30B 为 BF16/vLLM0.20/FA3/default CUDA graph；因此跨模型只检验 hypothesis consistency，causal phase claim 最终仍需 same-model phase pair。
+
+## 结果
+
+- **观察事实：** 235B fixed-shape mixed 已完成：real/sim TP4 top set exact match，worst regret=0、τ-b=0.8944；但 20 个 real non-tie pairs 只保持 16 个，10/34 anchors false-infeasible，TP8 MNS×MBT interaction 被漏掉。30B prefill-only 待运行。
+- **异常：** 无。
+- **Interpretation 与剩余 alternatives：** 强版本“mixed 必然失败”已被 235B top-set result 削弱；仍可能存在 phase-dependent error magnitude，由 topology margin 掩盖。
+- **Claim update：** unchanged，等待 30B prefill-only。
+- **下一步：** freeze 30B simulator surface → guided real anchors → joint verdict；只有判定需要时扩展 decode-heavy 235B 或 trace-shaped 30B prefill。
diff --git a/runs/frontier-phase-factorial-v0/fleet.toml b/runs/frontier-phase-factorial-v0/fleet.toml
new file mode 100644
index 0000000..76a9e93
--- /dev/null
+++ b/runs/frontier-phase-factorial-v0/fleet.toml
@@ -0,0 +1,24 @@
+version = 1
+
+[paths]
+state_dir = "runs/frontier-phase-factorial-v0/fleet-state"
+artifacts_dir = "runs/frontier-phase-factorial-v0/fleet-artifacts"
+
+[ssh]
+connect_timeout_sec = 10
+
+[scheduler]
+gpu_free_memory_mb = 1024
+gpu_free_utilization_pct = 10
+prefer_pack = true
+
+[sync]
+mode = "scp"
+local_path = "runs/frontier-phase-factorial-v0/remote-sync-marker"
+
+[[hosts]]
+name = "dash0"
+ssh_alias = "dash0"
+enabled = true
+sync_remote_path = "/home/admin/cpfs/wjh/aituner/phase-factorial-sync-marker"
+fleet_root = "/home/admin/cpfs/wjh/aituner/gpu-fleet-phase-factorial-v0"
diff --git a/runs/frontier-phase-factorial-v0/jobs_full.toml b/runs/frontier-phase-factorial-v0/jobs_full.toml
new file mode 100644
index 0000000..c67f255
--- /dev/null
+++ b/runs/frontier-phase-factorial-v0/jobs_full.toml
@@ -0,0 +1,241 @@
+version = 1
+
+[[jobs]]
+name = "qwen30-prefill-real-tp1-mns8-20260717-v1"
+gpus = 1
+gpu_model = "H20"
+hosts = ["dash0"]
+command = "cd /home/admin/cpfs/wjh/aituner/aituner-qwen30-vllm020-profile-v1/runs/frontier-phase-factorial-v0 && timeout --signal=TERM --kill-after=30s 7200 bash run_qwen30_prefill_real_config.sh"
+artifacts = ["artifacts/real-tp1-mns8-v1"]
+
+[jobs.env]
+HOME = "/tmp/wjh"
+XDG_CACHE_HOME = "/tmp/wjh/.cache"
+VLLM_CACHE_ROOT = "/tmp/wjh/.cache/vllm"
+TP = "1"
+MNS = "8"
+RATES = "4 8 16 32 64"
+SERVER_PORT = "8720"
+OUTPUT_ROOT = "/home/admin/cpfs/wjh/aituner/gpu-fleet-phase-factorial-v0/artifacts/real-tp1-mns8-v1"
+VENV_ROOT = "/tmp/wjh/venvs/vllm-0.20.0-cu129-profiler-v1"
+MODEL_ROOT = "/home/admin/cpfs/wjh/models/Qwen/Qwen3-30B-A3B"
+
+[[jobs]]
+name = "qwen30-prefill-real-tp1-mns16-20260717-v1"
+gpus = 1
+gpu_model = "H20"
+hosts = ["dash0"]
+command = "cd /home/admin/cpfs/wjh/aituner/aituner-qwen30-vllm020-profile-v1/runs/frontier-phase-factorial-v0 && timeout --signal=TERM --kill-after=30s 7200 bash run_qwen30_prefill_real_config.sh"
+artifacts = ["artifacts/real-tp1-mns16-v1"]
+
+[jobs.env]
+HOME = "/tmp/wjh"
+XDG_CACHE_HOME = "/tmp/wjh/.cache"
+VLLM_CACHE_ROOT = "/tmp/wjh/.cache/vllm"
+TP = "1"
+MNS = "16"
+RATES = "4 8 16 32 64"
+SERVER_PORT = "8721"
+OUTPUT_ROOT = "/home/admin/cpfs/wjh/aituner/gpu-fleet-phase-factorial-v0/artifacts/real-tp1-mns16-v1"
+VENV_ROOT = "/tmp/wjh/venvs/vllm-0.20.0-cu129-profiler-v1"
+MODEL_ROOT = "/home/admin/cpfs/wjh/models/Qwen/Qwen3-30B-A3B"
+
+[[jobs]]
+name = "qwen30-prefill-real-tp1-mns32-20260717-v1"
+gpus = 1
+gpu_model = "H20"
+hosts = ["dash0"]
+command = "cd /home/admin/cpfs/wjh/aituner/aituner-qwen30-vllm020-profile-v1/runs/frontier-phase-factorial-v0 && timeout --signal=TERM --kill-after=30s 7200 bash run_qwen30_prefill_real_config.sh"
+artifacts = ["artifacts/real-tp1-mns32-v1"]
+
+[jobs.env]
+HOME = "/tmp/wjh"
+XDG_CACHE_HOME = "/tmp/wjh/.cache"
+VLLM_CACHE_ROOT = "/tmp/wjh/.cache/vllm"
+TP = "1"
+MNS = "32"
+RATES = "4 8 16 32 64"
+SERVER_PORT = "8722"
+OUTPUT_ROOT = "/home/admin/cpfs/wjh/aituner/gpu-fleet-phase-factorial-v0/artifacts/real-tp1-mns32-v1"
+VENV_ROOT = "/tmp/wjh/venvs/vllm-0.20.0-cu129-profiler-v1"
+MODEL_ROOT = "/home/admin/cpfs/wjh/models/Qwen/Qwen3-30B-A3B"
+
+[[jobs]]
+name = "qwen30-prefill-real-tp1-mns64-20260717-v1"
+gpus = 1
+gpu_model = "H20"
+hosts = ["dash0"]
+command = "cd /home/admin/cpfs/wjh/aituner/aituner-qwen30-vllm020-profile-v1/runs/frontier-phase-factorial-v0 && timeout --signal=TERM --kill-after=30s 7200 bash run_qwen30_prefill_real_config.sh"
+artifacts = ["artifacts/real-tp1-mns64-v1"]
+
+[jobs.env]
+HOME = "/tmp/wjh"
+XDG_CACHE_HOME = "/tmp/wjh/.cache"
+VLLM_CACHE_ROOT = "/tmp/wjh/.cache/vllm"
+TP = "1"
+MNS = "64"
+RATES = "4 8 16 32 64"
+SERVER_PORT = "8723"
+OUTPUT_ROOT = "/home/admin/cpfs/wjh/aituner/gpu-fleet-phase-factorial-v0/artifacts/real-tp1-mns64-v1"
+VENV_ROOT = "/tmp/wjh/venvs/vllm-0.20.0-cu129-profiler-v1"
+MODEL_ROOT = "/home/admin/cpfs/wjh/models/Qwen/Qwen3-30B-A3B"
+
+[[jobs]]
+name = "qwen30-prefill-real-tp2-mns8-20260717-v1"
+gpus = 2
+gpu_model = "H20"
+hosts = ["dash0"]
+command = "cd /home/admin/cpfs/wjh/aituner/aituner-qwen30-vllm020-profile-v1/runs/frontier-phase-factorial-v0 && timeout --signal=TERM --kill-after=30s 7200 bash run_qwen30_prefill_real_config.sh"
+artifacts = ["artifacts/real-tp2-mns8-v1"]
+
+[jobs.env]
+HOME = "/tmp/wjh"
+XDG_CACHE_HOME = "/tmp/wjh/.cache"
+VLLM_CACHE_ROOT = "/tmp/wjh/.cache/vllm"
+TP = "2"
+MNS = "8"
+RATES = "4 8 16 32 64"
+SERVER_PORT = "8724"
+OUTPUT_ROOT = "/home/admin/cpfs/wjh/aituner/gpu-fleet-phase-factorial-v0/artifacts/real-tp2-mns8-v1"
+VENV_ROOT = "/tmp/wjh/venvs/vllm-0.20.0-cu129-profiler-v1"
+MODEL_ROOT = "/home/admin/cpfs/wjh/models/Qwen/Qwen3-30B-A3B"
+
+[[jobs]]
+name = "qwen30-prefill-real-tp2-mns16-20260717-v1"
+gpus = 2
+gpu_model = "H20"
+hosts = ["dash0"]
+command = "cd /home/admin/cpfs/wjh/aituner/aituner-qwen30-vllm020-profile-v1/runs/frontier-phase-factorial-v0 && timeout --signal=TERM --kill-after=30s 7200 bash run_qwen30_prefill_real_config.sh"
+artifacts = ["artifacts/real-tp2-mns16-v1"]
+
+[jobs.env]
+HOME = "/tmp/wjh"
+XDG_CACHE_HOME = "/tmp/wjh/.cache"
+VLLM_CACHE_ROOT = "/tmp/wjh/.cache/vllm"
+TP = "2"
+MNS = "16"
+RATES = "4 8 16 32 64"
+SERVER_PORT = "8725"
+OUTPUT_ROOT = "/home/admin/cpfs/wjh/aituner/gpu-fleet-phase-factorial-v0/artifacts/real-tp2-mns16-v1"
+VENV_ROOT = "/tmp/wjh/venvs/vllm-0.20.0-cu129-profiler-v1"
+MODEL_ROOT = "/home/admin/cpfs/wjh/models/Qwen/Qwen3-30B-A3B"
+
+[[jobs]]
+name = "qwen30-prefill-real-tp2-mns32-20260717-v1"
+gpus = 2
+gpu_model = "H20"
+hosts = ["dash0"]
+command = "cd /home/admin/cpfs/wjh/aituner/aituner-qwen30-vllm020-profile-v1/runs/frontier-phase-factorial-v0 && timeout --signal=TERM --kill-after=30s 7200 bash run_qwen30_prefill_real_config.sh"
+artifacts = ["artifacts/real-tp2-mns32-v1"]
+
+[jobs.env]
+HOME = "/tmp/wjh"
+XDG_CACHE_HOME = "/tmp/wjh/.cache"
+VLLM_CACHE_ROOT = "/tmp/wjh/.cache/vllm"
+TP = "2"
+MNS = "32"
+RATES = "4 8 16 32 64"
+SERVER_PORT = "8726"
+OUTPUT_ROOT = "/home/admin/cpfs/wjh/aituner/gpu-fleet-phase-factorial-v0/artifacts/real-tp2-mns32-v1"
+VENV_ROOT = "/tmp/wjh/venvs/vllm-0.20.0-cu129-profiler-v1"
+MODEL_ROOT = "/home/admin/cpfs/wjh/models/Qwen/Qwen3-30B-A3B"
+
+[[jobs]]
+name = "qwen30-prefill-real-tp2-mns64-20260717-v1"
+gpus = 2
+gpu_model = "H20"
+hosts = ["dash0"]
+command = "cd /home/admin/cpfs/wjh/aituner/aituner-qwen30-vllm020-profile-v1/runs/frontier-phase-factorial-v0 && timeout --signal=TERM --kill-after=30s 7200 bash run_qwen30_prefill_real_config.sh"
+artifacts = ["artifacts/real-tp2-mns64-v1"]
+
+[jobs.env]
+HOME = "/tmp/wjh"
+XDG_CACHE_HOME = "/tmp/wjh/.cache"
+VLLM_CACHE_ROOT = "/tmp/wjh/.cache/vllm"
+TP = "2"
+MNS = "64"
+RATES = "4 8 16 32 64"
+SERVER_PORT = "8727"
+OUTPUT_ROOT = "/home/admin/cpfs/wjh/aituner/gpu-fleet-phase-factorial-v0/artifacts/real-tp2-mns64-v1"
+VENV_ROOT = "/tmp/wjh/venvs/vllm-0.20.0-cu129-profiler-v1"
+MODEL_ROOT = "/home/admin/cpfs/wjh/models/Qwen/Qwen3-30B-A3B"
+
+[[jobs]]
+name = "qwen30-prefill-real-tp4-mns8-20260717-v1"
+gpus = 4
+gpu_model = "H20"
+hosts = ["dash0"]
+command = "cd /home/admin/cpfs/wjh/aituner/aituner-qwen30-vllm020-profile-v1/runs/frontier-phase-factorial-v0 && timeout --signal=TERM --kill-after=30s 7200 bash run_qwen30_prefill_real_config.sh"
+artifacts = ["artifacts/real-tp4-mns8-v1"]
+
+[jobs.env]
+HOME = "/tmp/wjh"
+XDG_CACHE_HOME = "/tmp/wjh/.cache"
+VLLM_CACHE_ROOT = "/tmp/wjh/.cache/vllm"
+TP = "4"
+MNS = "8"
+RATES = "4 8 16 32 64"
+SERVER_PORT = "8728"
+OUTPUT_ROOT = "/home/admin/cpfs/wjh/aituner/gpu-fleet-phase-factorial-v0/artifacts/real-tp4-mns8-v1"
+VENV_ROOT = "/tmp/wjh/venvs/vllm-0.20.0-cu129-profiler-v1"
+MODEL_ROOT = "/home/admin/cpfs/wjh/models/Qwen/Qwen3-30B-A3B"
+
+[[jobs]]
+name = "qwen30-prefill-real-tp4-mns16-20260717-v1"
+gpus = 4
+gpu_model = "H20"
+hosts = ["dash0"]
+command = "cd /home/admin/cpfs/wjh/aituner/aituner-qwen30-vllm020-profile-v1/runs/frontier-phase-factorial-v0 && timeout --signal=TERM --kill-after=30s 7200 bash run_qwen30_prefill_real_config.sh"
+artifacts = ["artifacts/real-tp4-mns16-v1"]
+
+[jobs.env]
+HOME = "/tmp/wjh"
+XDG_CACHE_HOME = "/tmp/wjh/.cache"
+VLLM_CACHE_ROOT = "/tmp/wjh/.cache/vllm"
+TP = "4"
+MNS = "16"
+RATES = "4 8 16 32 64"
+SERVER_PORT = "8729"
+OUTPUT_ROOT = "/home/admin/cpfs/wjh/aituner/gpu-fleet-phase-factorial-v0/artifacts/real-tp4-mns16-v1"
+VENV_ROOT = "/tmp/wjh/venvs/vllm-0.20.0-cu129-profiler-v1"
+MODEL_ROOT = "/home/admin/cpfs/wjh/models/Qwen/Qwen3-30B-A3B"
+
+[[jobs]]
+name = "qwen30-prefill-real-tp4-mns32-20260717-v1"
+gpus = 4
+gpu_model = "H20"
+hosts = ["dash0"]
+command = "cd /home/admin/cpfs/wjh/aituner/aituner-qwen30-vllm020-profile-v1/runs/frontier-phase-factorial-v0 && timeout --signal=TERM --kill-after=30s 7200 bash run_qwen30_prefill_real_config.sh"
+artifacts = ["artifacts/real-tp4-mns32-v1"]
+
+[jobs.env]
+HOME = "/tmp/wjh"
+XDG_CACHE_HOME = "/tmp/wjh/.cache"
+VLLM_CACHE_ROOT = "/tmp/wjh/.cache/vllm"
+TP = "4"
+MNS = "32"
+RATES = "4 8 16 32 64"
+SERVER_PORT = "8730"
+OUTPUT_ROOT = "/home/admin/cpfs/wjh/aituner/gpu-fleet-phase-factorial-v0/artifacts/real-tp4-mns32-v1"
+VENV_ROOT = "/tmp/wjh/venvs/vllm-0.20.0-cu129-profiler-v1"
+MODEL_ROOT = "/home/admin/cpfs/wjh/models/Qwen/Qwen3-30B-A3B"
+
+[[jobs]]
+name = "qwen30-prefill-real-tp4-mns64-20260717-v1"
+gpus = 4
+gpu_model = "H20"
+hosts = ["dash0"]
+command = "cd /home/admin/cpfs/wjh/aituner/aituner-qwen30-vllm020-profile-v1/runs/frontier-phase-factorial-v0 && timeout --signal=TERM --kill-after=30s 7200 bash run_qwen30_prefill_real_config.sh"
+artifacts = ["artifacts/real-tp4-mns64-v1"]
+
+[jobs.env]
+HOME = "/tmp/wjh"
+XDG_CACHE_HOME = "/tmp/wjh/.cache"
+VLLM_CACHE_ROOT = "/tmp/wjh/.cache/vllm"
+TP = "4"
+MNS = "64"
+RATES = "4 8 16 32 64"
+SERVER_PORT = "8731"
+OUTPUT_ROOT = "/home/admin/cpfs/wjh/aituner/gpu-fleet-phase-factorial-v0/artifacts/real-tp4-mns64-v1"
+VENV_ROOT = "/tmp/wjh/venvs/vllm-0.20.0-cu129-profiler-v1"
+MODEL_ROOT = "/home/admin/cpfs/wjh/models/Qwen/Qwen3-30B-A3B"
diff --git a/runs/frontier-phase-factorial-v0/jobs_smoke.toml b/runs/frontier-phase-factorial-v0/jobs_smoke.toml
new file mode 100644
index 0000000..5741e2d
--- /dev/null
+++ b/runs/frontier-phase-factorial-v0/jobs_smoke.toml
@@ -0,0 +1,21 @@
+version = 1
+
+[[jobs]]
+name = "qwen30-prefill-real-smoke-tp1-mns8-20260717-v1"
+gpus = 1
+gpu_model = "H20"
+hosts = ["dash0"]
+command = "cd /home/admin/cpfs/wjh/aituner/aituner-qwen30-vllm020-profile-v1/runs/frontier-phase-factorial-v0 && timeout --signal=TERM --kill-after=30s 1800 bash run_qwen30_prefill_real_config.sh"
+artifacts = ["artifacts/real-smoke-tp1-mns8-v1"]
+
+[jobs.env]
+HOME = "/tmp/wjh"
+XDG_CACHE_HOME = "/tmp/wjh/.cache"
+VLLM_CACHE_ROOT = "/tmp/wjh/.cache/vllm"
+TP = "1"
+MNS = "8"
+RATES = "4"
+SERVER_PORT = "8718"
+OUTPUT_ROOT = "/home/admin/cpfs/wjh/aituner/gpu-fleet-phase-factorial-v0/artifacts/real-smoke-tp1-mns8-v1"
+VENV_ROOT = "/tmp/wjh/venvs/vllm-0.20.0-cu129-profiler-v1"
+MODEL_ROOT = "/home/admin/cpfs/wjh/models/Qwen/Qwen3-30B-A3B"
diff --git a/runs/frontier-phase-factorial-v0/mock-phase-factorial.png b/runs/frontier-phase-factorial-v0/mock-phase-factorial.png
new file mode 100644
index 0000000..6c3c2bd
Binary files /dev/null and b/runs/frontier-phase-factorial-v0/mock-phase-factorial.png differ
diff --git a/runs/frontier-phase-factorial-v0/plot_mock_phase_factorial.py b/runs/frontier-phase-factorial-v0/plot_mock_phase_factorial.py
new file mode 100644
index 0000000..fdfc37d
--- /dev/null
+++ b/runs/frontier-phase-factorial-v0/plot_mock_phase_factorial.py
@@ -0,0 +1,45 @@
+#!/usr/bin/env python3
+"""Render the preregistered phase-factorial hypothesis schematic (mock data)."""
+
+from __future__ import annotations
+
+from pathlib import Path
+
+import matplotlib
+
+matplotlib.use("Agg")
+import matplotlib.pyplot as plt
+
+
+def main() -> None:
+    labels = ("235B\nprefill-only", "235B\nmixed", "30B\nprefill-only", "30B\nmixed")
+    x = range(len(labels))
+    phase_hypothesis = (1.0, 18.0, 2.0, 28.0)
+    stack_hypothesis = (1.0, 4.0, 24.0, 30.0)
+
+    fig, ax = plt.subplots(figsize=(8.6, 4.8), constrained_layout=True)
+    ax.plot(x, phase_hypothesis, marker="o", lw=2, label="H-phase (mock)")
+    ax.plot(x, stack_hypothesis, marker="s", lw=2, label="H-stack (mock)")
+    ax.axhline(5, color="black", ls="--", lw=1.2, label="5% regret gate")
+    ax.set_xticks(list(x), labels)
+    ax.set_ylabel("Worst selected-config regret (%) — MOCK DATA")
+    ax.set_title("Schematic only: predictions that distinguish phase vs stack explanations")
+    ax.set_ylim(0, 35)
+    ax.grid(axis="y", alpha=0.25)
+    ax.legend()
+    ax.text(
+        0.01,
+        0.98,
+        "MOCK DATA / NOT A RESULT",
+        transform=ax.transAxes,
+        va="top",
+        color="crimson",
+        weight="bold",
+    )
+    output = Path(__file__).with_name("mock-phase-factorial.png")
+    fig.savefig(output, dpi=180)
+    print(output)
+
+
+if __name__ == "__main__":
+    main()
diff --git a/runs/frontier-phase-factorial-v0/qwen30_prefill_client.py b/runs/frontier-phase-factorial-v0/qwen30_prefill_client.py
new file mode 100644
index 0000000..de40676
--- /dev/null
+++ b/runs/frontier-phase-factorial-v0/qwen30_prefill_client.py
@@ -0,0 +1,216 @@
+#!/usr/bin/env python3
+"""Open-loop fixed-shape prefill-only workload for one real offered-load anchor."""
+
+from __future__ import annotations
+
+import argparse
+import concurrent.futures
+import hashlib
+import http.client
+import json
+import math
+import time
+from pathlib import Path
+from typing import Any
+
+
+TTFT_SLO_MS = 1256.0
+TARGET_PASS_RATE = 0.95
+
+
+def parse_args() -> argparse.Namespace:
+    parser = argparse.ArgumentParser()
+    parser.add_argument("--host", default="127.0.0.1")
+    parser.add_argument("--port", type=int, required=True)
+    parser.add_argument("--served-model", required=True)
+    parser.add_argument("--model-path", type=Path, required=True)
+    parser.add_argument("--rate", type=float, required=True)
+    parser.add_argument("--requests", type=int, default=64)
+    parser.add_argument("--input-tokens", type=int, default=2048)
+    parser.add_argument("--timeout-seconds", type=float, default=900.0)
+    parser.add_argument("--output", type=Path, required=True)
+    return parser.parse_args()
+
+
+def percentile(values: list[float], fraction: float) -> float | None:
+    if not values:
+        return None
+    ordered = sorted(values)
+    index = min(len(ordered) - 1, max(0, math.ceil(fraction * len(ordered)) - 1))
+    return ordered[index]
+
+
+def run_request(
+    *,
+    request_index: int,
+    scheduled_at: float,
+    benchmark_start: float,
+    args: argparse.Namespace,
+    prompt_ids: list[int],
+) -> dict[str, Any]:
+    delay = scheduled_at - time.perf_counter()
+    if delay > 0:
+        time.sleep(delay)
+    admitted = time.perf_counter()
+    record: dict[str, Any] = {
+        "request_index": request_index,
+        "scheduled_s": scheduled_at - benchmark_start,
+        "admitted_s": admitted - benchmark_start,
+        "admission_lag_ms": (admitted - scheduled_at) * 1000.0,
+        "success": False,
+    }
+    connection = http.client.HTTPConnection(args.host, args.port, timeout=args.timeout_seconds)
+    body = {
+        "model": args.served_model,
+        "prompt": prompt_ids,
+        "min_tokens": 1,
+        "max_tokens": 1,
+        "ignore_eos": True,
+        "temperature": 0,
+        "stream": True,
+        "stream_options": {"include_usage": True},
+        "return_token_ids": True,
+    }
+    try:
+        started = time.perf_counter()
+        connection.request(
+            "POST",
+            "/v1/completions",
+            body=json.dumps(body, separators=(",", ":")).encode(),
+            headers={"Content-Type": "application/json"},
+        )
+        response = connection.getresponse()
+        if response.status != 200:
+            raise RuntimeError(
+                f"HTTP {response.status}: {response.read().decode(errors='replace')}"
+            )
+        first_token_at = None
+        streamed_tokens = 0
+        usage = None
+        while True:
+            raw = response.readline()
+            if not raw:
+                break
+            line = raw.decode(errors="replace").strip()
+            if not line.startswith("data:"):
+                continue
+            data = line[5:].strip()
+            if data == "[DONE]":
+                break
+            payload = json.loads(data)
+            if payload.get("usage"):
+                usage = payload["usage"]
+            emitted = 0
+            for choice in payload.get("choices") or []:
+                token_ids = choice.get("token_ids") or []
+                emitted += len(token_ids) if token_ids else int(bool(choice.get("text")))
+            if emitted:
+                first_token_at = first_token_at or time.perf_counter()
+                streamed_tokens += emitted
+        finished = time.perf_counter()
+        if first_token_at is None or usage is None:
+            raise RuntimeError("missing streaming token or usage")
+        prompt_tokens = int(usage["prompt_tokens"])
+        completion_tokens = int(usage["completion_tokens"])
+        if prompt_tokens != args.input_tokens or completion_tokens != 1:
+            raise RuntimeError(f"usage mismatch: {prompt_tokens}+{completion_tokens}")
+        ttft = (first_token_at - started) * 1000.0
+        record.update(
+            {
+                "success": True,
+                "prompt_tokens": prompt_tokens,
+                "completion_tokens": completion_tokens,
+                "streamed_token_count": streamed_tokens,
+                "ttft_ms": ttft,
+                "e2e_ms": (finished - started) * 1000.0,
+                "slo_pass": ttft <= TTFT_SLO_MS,
+            }
+        )
+    except Exception as error:  # Failed requests remain in the SLO denominator.
+        record["error"] = f"{type(error).__name__}: {error}"
+        record["slo_pass"] = False
+    finally:
+        connection.close()
+    return record
+
+
+def main() -> None:
+    args = parse_args()
+    if args.rate <= 0 or args.requests <= 0 or args.input_tokens <= 0:
+        raise ValueError("rate, requests, and input tokens must be positive")
+    from transformers import AutoTokenizer
+
+    tokenizer = AutoTokenizer.from_pretrained(args.model_path, trust_remote_code=True)
+    excluded = set(tokenizer.all_special_ids)
+    candidates = [
+        token_id for token_id in range(tokenizer.vocab_size) if token_id not in excluded
+    ]
+    if len(candidates) < args.requests + 1:
+        raise RuntimeError("tokenizer has too few non-special token IDs")
+    base_id = candidates[0]
+    prompts = [
+        [candidates[index + 1], *([base_id] * (args.input_tokens - 1))]
+        for index in range(args.requests)
+    ]
+    prompt_hash = hashlib.sha256(
+        "\n".join(",".join(map(str, prompt)) for prompt in prompts).encode()
+    ).hexdigest()
+
+    benchmark_start = time.perf_counter() + 2.0
+    with concurrent.futures.ThreadPoolExecutor(max_workers=args.requests) as pool:
+        futures = [
+            pool.submit(
+                run_request,
+                request_index=index,
+                scheduled_at=benchmark_start + index / args.rate,
+                benchmark_start=benchmark_start,
+                args=args,
+                prompt_ids=prompts[index],
+            )
+            for index in range(args.requests)
+        ]
+        requests = [future.result() for future in futures]
+    requests.sort(key=lambda row: int(row["request_index"]))
+    completed = [row for row in requests if row["success"]]
+    passed = sum(bool(row["slo_pass"]) for row in requests)
+    ttfts = [float(row["ttft_ms"]) for row in completed]
+    pass_rate = passed / len(requests)
+    payload = {
+        "schema": "qwen30-prefill-rate-anchor-v1",
+        "workload": {
+            "offered_request_rate": args.rate,
+            "request_count": args.requests,
+            "input_tokens": args.input_tokens,
+            "output_tokens": 1,
+            "prefix_caching": False,
+            "arrival": "open_loop_uniform",
+            "last_scheduled_arrival_s": (args.requests - 1) / args.rate,
+            "prompt_vector_sha256": prompt_hash,
+        },
+        "summary": {
+            "completed": len(completed),
+            "failed": len(requests) - len(completed),
+            "ttft_p50_ms": percentile(ttfts, 0.50),
+            "ttft_p95_ms": percentile(ttfts, 0.95),
+            "ttft_max_ms": max(ttfts) if ttfts else None,
+            "admission_lag_max_ms": max(
+                float(row["admission_lag_ms"]) for row in requests
+            ),
+            "slo": {
+                "ttft_threshold_ms": TTFT_SLO_MS,
+                "passed": passed,
+                "pass_rate": pass_rate,
+                "feasible": pass_rate >= TARGET_PASS_RATE,
+            },
+        },
+        "requests": requests,
+    }
+    args.output.parent.mkdir(parents=True, exist_ok=True)
+    args.output.write_text(json.dumps(payload, indent=2, sort_keys=True) + "\n")
+    print(json.dumps(payload["summary"], sort_keys=True), flush=True)
+    if len(completed) != args.requests:
+        raise SystemExit(2)
+
+
+if __name__ == "__main__":
+    main()
diff --git a/runs/frontier-phase-factorial-v0/remote-sync-marker/README.md b/runs/frontier-phase-factorial-v0/remote-sync-marker/README.md
new file mode 100644
index 0000000..9dfd754
--- /dev/null
+++ b/runs/frontier-phase-factorial-v0/remote-sync-marker/README.md
@@ -0,0 +1,6 @@
+# Phase-factorial fleet sync marker
+
+Experiment source is synchronized through Git into the clean dash0 checkout
+`/home/admin/cpfs/wjh/aituner/aituner-qwen30-vllm020-profile-v1`. This small
+directory exists only to satisfy the fleet orchestrator's explicit sync phase
+without copying local simulator caches or raw metrics to the GPU host.
diff --git a/runs/frontier-phase-factorial-v0/run_frontier_qwen30_prefill_surface.py b/runs/frontier-phase-factorial-v0/run_frontier_qwen30_prefill_surface.py
new file mode 100644
index 0000000..10249d7
--- /dev/null
+++ b/runs/frontier-phase-factorial-v0/run_frontier_qwen30_prefill_surface.py
@@ -0,0 +1,400 @@
+#!/usr/bin/env python3
+"""Freeze the Qwen30 fixed-shape prefill-only Frontier surface."""
+
+from __future__ import annotations
+
+import argparse
+import csv
+import hashlib
+import importlib.util
+import json
+import math
+import os
+import subprocess
+import sys
+import time
+from dataclasses import asdict, dataclass
+from pathlib import Path
+from typing import Any
+
+
+MODEL = "qwen3-a3b-30b-moe"
+RATES = (4.0, 8.0, 16.0, 32.0, 64.0)
+TTFT_SLO_MS = 1256.0
+TARGET_PASS_RATE = 0.95
+NUM_BLOCKS = {1: 20080, 2: 76537, 4: 191727}
+
+
+@dataclass(frozen=True)
+class Config:
+    tp: int
+    mns: int
+
+    @property
+    def name(self) -> str:
+        return f"tp{self.tp}_mns{self.mns}"
+
+
+GRID = tuple(Config(tp, mns) for tp in (1, 2, 4) for mns in (8, 16, 32, 64))
+
+
+def parse_args() -> argparse.Namespace:
+    parser = argparse.ArgumentParser()
+    parser.add_argument("--frontier-source", type=Path, required=True)
+    parser.add_argument("--replayserve-root", type=Path, required=True)
+    parser.add_argument("--profile-root", type=Path, required=True)
+    parser.add_argument("--python-deps", type=Path, required=True)
+    parser.add_argument("--output-root", type=Path, required=True)
+    parser.add_argument("--requests", type=int, default=64)
+    parser.add_argument("--rate", type=float, action="append")
+    parser.add_argument("--config", action="append")
+    parser.add_argument("--timeout-seconds", type=float, default=900.0)
+    parser.add_argument("--resume", action="store_true")
+    return parser.parse_args()
+
+
+def sha256(path: Path) -> str:
+    digest = hashlib.sha256()
+    with path.open("rb") as source:
+        for chunk in iter(lambda: source.read(1 << 20), b""):
+            digest.update(chunk)
+    return digest.hexdigest()
+
+
+def write_json(path: Path, payload: Any) -> None:
+    path.parent.mkdir(parents=True, exist_ok=True)
+    temporary = path.with_suffix(path.suffix + ".tmp")
+    temporary.write_text(json.dumps(payload, indent=2, sort_keys=True) + "\n")
+    os.replace(temporary, path)
+
+
+def load_module(name: str, path: Path):
+    spec = importlib.util.spec_from_file_location(name, path)
+    if spec is None or spec.loader is None:
+        raise ImportError(path)
+    module = importlib.util.module_from_spec(spec)
+    sys.modules[name] = module
+    spec.loader.exec_module(module)
+    return module
+
+
+def write_trace(path: Path, *, requests: int, rate: float) -> None:
+    path.parent.mkdir(parents=True, exist_ok=True)
+    fields = [
+        "arrived_at",
+        "num_prefill_tokens",
+        "num_decode_tokens",
+        "session_id",
+        "block_hash_ids",
+    ]
+    with path.open("w", newline="") as output:
+        writer = csv.DictWriter(output, fieldnames=fields)
+        writer.writeheader()
+        for request_id in range(requests):
+            writer.writerow(
+                {
+                    "arrived_at": f"{request_id / rate:.12f}",
+                    "num_prefill_tokens": 2048,
+                    "num_decode_tokens": 1,
+                    "session_id": request_id,
+                    "block_hash_ids": "|".join(
+                        str(request_id * 128 + block + 1) for block in range(128)
+                    ),
+                }
+            )
+
+
+def profile_paths(root: Path) -> dict[str, Path]:
+    paths = {
+        "linear": root / "linear_op.csv",
+        "attention": root / "attention.csv",
+        "moe": root / "moe.csv",
+        "manifest": root / "manifest.json",
+    }
+    missing = [str(path) for path in paths.values() if not path.is_file()]
+    if missing:
+        raise FileNotFoundError(missing)
+    return paths
+
+
+def validate_profile(paths: dict[str, Path]) -> dict[str, Any]:
+    manifest = json.loads(paths["manifest"].read_text())
+    expected = manifest["outputs"]
+    for filename in ("linear_op.csv", "attention.csv", "moe.csv"):
+        actual = sha256(paths[{"linear_op.csv": "linear", "attention.csv": "attention", "moe.csv": "moe"}[filename]])
+        if actual != expected[filename]:
+            raise ValueError(f"profile hash mismatch for {filename}")
+    with paths["attention"].open(newline="") as source:
+        rows = list(csv.DictReader(source))
+    coverage = {}
+    for tp in (1, 2, 4):
+        exact = [
+            row
+            for row in rows
+            if int(row["num_tensor_parallel_workers"]) == tp
+            and row["is_prefill"].lower() == "true"
+            and row.get("is_true_mixed_batch", "").lower() != "true"
+            and int(float(row["total_tokens"])) == 2048
+        ]
+        if len(exact) != 1:
+            raise ValueError(f"expected one exact TP{tp} 2048-token prefill row, got {len(exact)}")
+        coverage[str(tp)] = {"exact_prefill_2048_rows": len(exact), "profile_batch_size": int(exact[0]["batch_size"])}
+    return {"manifest": manifest, "attention": coverage}
+
+
+def knobs(config: Config, paths: dict[str, Path], cache: Path) -> dict[str, Any]:
+    return {
+        "simulation_mode": "online",
+        "sys_arch": "co-location",
+        "num_replicas": 1,
+        "cluster_scheduler": "sticky_round_robin",
+        "model_name": MODEL,
+        "device": "h20",
+        "network_device": "h20_dgx",
+        "attn_tensor_parallel_size": config.tp,
+        "attn_data_parallel_size": 1,
+        "moe_tensor_parallel_size": config.tp,
+        "moe_expert_parallel_size": 1,
+        "num_pipeline_stages": 1,
+        "replica_scheduler": "vllm_v1",
+        "decode_cuda_graph_mode": "none",
+        "batch_size_cap": config.mns,
+        "max_tokens_in_batch": 8192,
+        "long_prefill_token_threshold": 0,
+        "block_size": 16,
+        "num_blocks_mode": "explicit",
+        "num_blocks": NUM_BLOCKS[config.tp],
+        "gpu_memory_utilization": 0.92,
+        "non_kv_cache_overhead_bytes": 0,
+        "trace_max_tokens": 40960,
+        "cache_dir": str(cache),
+        "enable_dummy_mode": False,
+        "linear_op_input_file": str(paths["linear"]),
+        "atten_input_file": str(paths["attention"]),
+        "moe_input_file": str(paths["moe"]),
+        "prediction_max_prefill_chunk_size": 18000,
+        "prediction_max_batch_size": 128,
+        "prediction_max_tokens_per_request": 32768,
+        "no_cache": False,
+        "skip_cpu_overhead_modeling": True,
+        "enable_prefix_caching": False,
+        "enable_chunked_prefill": True,
+    }
+
+
+def find_metrics(run_dir: Path) -> tuple[Path, Path]:
+    systems = list((run_dir / "metrics").rglob("system_metrics.json"))
+    requests = list((run_dir / "metrics").rglob("request_metrics.csv"))
+    if len(systems) != 1 or len(requests) != 1:
+        raise RuntimeError(f"metric pair mismatch: {len(systems)}/{len(requests)}")
+    return systems[0], requests[0]
+
+
+def score(system_path: Path, request_path: Path, expected_requests: int) -> dict[str, Any]:
+    system = json.loads(system_path.read_text())
+    metadata = system["simulation_metadata"]
+    if int(metadata["completed_requests"]) != expected_requests:
+        raise ValueError("Frontier completion count mismatch")
+    with request_path.open(newline="") as source:
+        rows = list(csv.DictReader(source))
+    if len(rows) != expected_requests:
+        raise ValueError("request metric count mismatch")
+    values = []
+    passed = 0
+    for row in rows:
+        if int(float(row["request_num_prefill_tokens"])) != 2048:
+            raise ValueError("prefill shape drift")
+        if int(float(row["request_num_decode_tokens"])) != 1:
+            raise ValueError("decode shape drift")
+        ttft = float(row["ttft"])
+        if not math.isfinite(ttft) or ttft < 0:
+            raise ValueError("invalid TTFT")
+        values.append(ttft)
+        passed += int(ttft <= TTFT_SLO_MS)
+    ordered = sorted(values)
+    pass_rate = passed / expected_requests
+    return {
+        "ttft_p50_ms": ordered[math.ceil(0.50 * len(ordered)) - 1],
+        "ttft_p95_ms": ordered[math.ceil(0.95 * len(ordered)) - 1],
+        "ttft_max_ms": max(ordered),
+        "passed": passed,
+        "pass_rate": pass_rate,
+        "feasible": pass_rate >= TARGET_PASS_RATE,
+        "throughput_requests_per_second": float(system["throughput_metrics"]["requests_per_second"]),
+    }
+
+
+def main() -> None:
+    args = parse_args()
+    args.frontier_source = args.frontier_source.resolve()
+    args.replayserve_root = args.replayserve_root.resolve()
+    args.profile_root = args.profile_root.resolve()
+    args.python_deps = args.python_deps.resolve()
+    args.output_root = args.output_root.resolve()
+    rates = tuple(args.rate or RATES)
+    selected = list(GRID)
+    if args.config:
+        wanted = set(args.config)
+        selected = [config for config in GRID if config.name in wanted]
+        if {config.name for config in selected} != wanted:
+            raise ValueError(f"unknown configs: {wanted - {config.name for config in selected}}")
+    paths = profile_paths(args.profile_root)
+    coverage = validate_profile(paths)
+    builder = load_module(
+        "qwen30_prefill_frontier_builder",
+        args.replayserve_root / "tools/run_frontier_sweep.py",
+    )
+    frontier_head = subprocess.run(
+        ["git", "-C", str(args.frontier_source), "rev-parse", "HEAD"],
+        check=True,
+        text=True,
+        stdout=subprocess.PIPE,
+    ).stdout.strip()
+    traces = {}
+    for rate in rates:
+        trace = args.output_root / "traces" / f"r{rate:g}.csv"
+        write_trace(trace, requests=args.requests, rate=rate)
+        traces[rate] = trace
+
+    config_results = []
+    for config in selected:
+        loads = []
+        config_knobs = knobs(config, paths, args.output_root / "cache")
+        for rate in rates:
+            run_dir = args.output_root / "runs" / config.name / f"r{rate:g}"
+            result_path = run_dir / "result.json"
+            if args.resume and result_path.is_file():
+                loads.append(json.loads(result_path.read_text()))
+                continue
+            run_dir.mkdir(parents=True, exist_ok=True)
+            command = builder.build_frontier_command(
+                python_bin="/usr/bin/python3",
+                trace_file=traces[rate],
+                metrics_root=run_dir / "metrics",
+                run_id=f"qwen30_prefill_{config.name}_r{rate:g}",
+                knobs=config_knobs,
+            )
+            write_json(run_dir / "command.json", command)
+            environment = os.environ.copy()
+            pythonpath = [str(args.python_deps), str(args.frontier_source)]
+            if environment.get("PYTHONPATH"):
+                pythonpath.append(environment["PYTHONPATH"])
+            environment.update(
+                {
+                    "PYTHONPATH": ":".join(pythonpath),
+                    "CUDA_VISIBLE_DEVICES": "",
+                    "NVIDIA_VISIBLE_DEVICES": "void",
+                    "WANDB_DISABLED": "true",
+                    "VIDUR_DISABLE_WANDB": "1",
+                    "FRONTIER_LOG_LEVEL": "WARNING",
+                    "PYTHONDONTWRITEBYTECODE": "1",
+                }
+            )
+            started = time.time()
+            with (run_dir / "stdout.log").open("w") as stdout, (
+                run_dir / "stderr.log"
+            ).open("w") as stderr:
+                completed = subprocess.run(
+                    command,
+                    cwd=args.frontier_source,
+                    env=environment,
+                    stdout=stdout,
+                    stderr=stderr,
+                    timeout=args.timeout_seconds,
+                    check=False,
+                )
+            if completed.returncode != 0:
+                raise RuntimeError(
+                    f"Frontier failed for {config.name} rate={rate}: {completed.returncode}"
+                )
+            system_path, request_path = find_metrics(run_dir)
+            result = {
+                "status": "completed",
+                "config": asdict(config) | {"name": config.name},
+                "offered_request_rate": rate,
+                "offered_request_rate_per_gpu": rate / config.tp,
+                "request_count": args.requests,
+                "elapsed_seconds": time.time() - started,
+                "trace_sha256": sha256(traces[rate]),
+                "request_metrics_sha256": sha256(request_path),
+                "score": score(system_path, request_path, args.requests),
+            }
+            write_json(result_path, result)
+            loads.append(result)
+            print(
+                json.dumps(
+                    {
+                        "config": config.name,
+                        "rate": rate,
+                        "pass_rate": result["score"]["pass_rate"],
+                        "feasible": result["score"]["feasible"],
+                    },
+                    sort_keys=True,
+                ),
+                flush=True,
+            )
+        config_results.append({"config": asdict(config) | {"name": config.name}, "loads": loads})
+
+    capacities = []
+    for item in config_results:
+        feasible = [
+            load["offered_request_rate"]
+            for load in item["loads"]
+            if load["score"]["feasible"]
+        ]
+        capacity = max(feasible) if feasible else None
+        capacities.append(
+            {
+                "config": item["config"],
+                "maximum_tested_feasible_request_rate": capacity,
+                "maximum_tested_feasible_request_rate_per_gpu": (
+                    capacity / item["config"]["tp"] if capacity is not None else None
+                ),
+                "lower_censored": capacity is None,
+                "upper_censored": capacity == max(rates) if capacity is not None else False,
+            }
+        )
+    capacities.sort(
+        key=lambda row: (
+            -(row["maximum_tested_feasible_request_rate_per_gpu"] or -1),
+            row["config"]["name"],
+        )
+    )
+    full = selected == list(GRID) and rates == RATES and args.requests == 64
+    manifest = {
+        "schema": "frontier-qwen30-prefill-surface-v1",
+        "status": "frozen_before_real" if full else "partial_not_decision_bearing",
+        "contract": {
+            "rates": rates,
+            "requests_per_anchor": args.requests,
+            "input_tokens": 2048,
+            "output_tokens": 1,
+            "ttft_slo_ms": TTFT_SLO_MS,
+            "target_pass_rate": TARGET_PASS_RATE,
+            "prefix_caching": False,
+            "arrival": "open_loop_uniform",
+        },
+        "frontier": {
+            "source": str(args.frontier_source),
+            "git_head": frontier_head,
+            "git_status_short": subprocess.run(
+                ["git", "-C", str(args.frontier_source), "status", "--short"],
+                check=True,
+                text=True,
+                stdout=subprocess.PIPE,
+            ).stdout,
+        },
+        "profiles": {
+            "root": str(args.profile_root),
+            "coverage": coverage,
+            "sha256": {name: sha256(path) for name, path in paths.items()},
+        },
+        "config_results": config_results,
+        "capacity": capacities,
+    }
+    write_json(args.output_root / "frontier_surface_frozen.json", manifest)
+    print(args.output_root / "frontier_surface_frozen.json")
+
+
+if __name__ == "__main__":
+    main()
diff --git a/runs/frontier-phase-factorial-v0/run_qwen30_prefill_real_config.sh b/runs/frontier-phase-factorial-v0/run_qwen30_prefill_real_config.sh
new file mode 100644
index 0000000..a406d51
--- /dev/null
+++ b/runs/frontier-phase-factorial-v0/run_qwen30_prefill_real_config.sh
@@ -0,0 +1,135 @@
+#!/usr/bin/env bash
+
+set -euo pipefail
+
+OUTPUT_ROOT="${OUTPUT_ROOT:?OUTPUT_ROOT is required}"
+TP="${TP:?TP is required}"
+MNS="${MNS:?MNS is required}"
+RATES="${RATES:-4 8 16 32 64}"
+SERVER_PORT="${SERVER_PORT:?SERVER_PORT is required}"
+VENV_ROOT="${VENV_ROOT:-/tmp/wjh/venvs/vllm-0.20.0-cu129-profiler-v1}"
+MODEL_ROOT="${MODEL_ROOT:-/home/admin/cpfs/wjh/models/Qwen/Qwen3-30B-A3B}"
+SERVED_MODEL="qwen3-30b-prefill-only"
+SERVER_PID=""
+
+mkdir -p "${OUTPUT_ROOT}/logs" "${OUTPUT_ROOT}/provenance"
+exec > >(tee -a "${OUTPUT_ROOT}/logs/controller.log") 2>&1
+
+cleanup() {
+  if [[ -n "${SERVER_PID}" ]] && kill -0 "${SERVER_PID}" 2>/dev/null; then
+    kill -TERM -- "-${SERVER_PID}" 2>/dev/null || true
+    for _ in $(seq 1 30); do
+      kill -0 "${SERVER_PID}" 2>/dev/null || break
+      sleep 1
+    done
+    kill -KILL -- "-${SERVER_PID}" 2>/dev/null || true
+  fi
+  SERVER_PID=""
+}
+trap cleanup EXIT INT TERM
+
+IFS=',' read -r -a GPU_IDS <<< "${CUDA_VISIBLE_DEVICES:-}"
+if [[ "${#GPU_IDS[@]}" -ne "${TP}" ]]; then
+  echo "ERROR: expected ${TP} allocated GPUs, got ${CUDA_VISIBLE_DEVICES:-unset}" >&2
+  exit 1
+fi
+read -r -a RATE_ARRAY <<< "${RATES}"
+
+echo "QWEN30_PREFILL_REAL_LAUNCH_ECHO host=$(hostname) gpus=${CUDA_VISIBLE_DEVICES} model=${MODEL_ROOT} runtime=vLLM-0.20.0+cu129 dtype=BF16 config=TP${TP}_MNS${MNS}_MBT8192 rates=${RATES// /,} rounds=2 requests=64 shape=ISL2048_OSL1 arrivals=uniform prefix=off cuda_graph=runtime_default isolation=fresh_server_per_anchor output=${OUTPUT_ROOT}"
+date -u +"START_UTC=%Y-%m-%dT%H:%M:%SZ"
+sha256sum run_qwen30_prefill_real_config.sh qwen30_prefill_client.py \
+  > "${OUTPUT_ROOT}/provenance/source.sha256"
+"${VENV_ROOT}/bin/python" - "${TP}" "${MNS}" "${RATES}" \
+  > "${OUTPUT_ROOT}/provenance/contract.json" <<'PY'
+import importlib.metadata as metadata
+import json
+import platform
+import sys
+
+tp, mns, rates = sys.argv[1:]
+print(json.dumps({
+    "python": platform.python_version(),
+    "torch": metadata.version("torch"),
+    "transformers": metadata.version("transformers"),
+    "vllm": metadata.version("vllm"),
+    "config": {"tp": int(tp), "mns": int(mns), "mbt": 8192},
+    "rates": [float(value) for value in rates.split()],
+    "rounds": 2,
+    "requests_per_anchor": 64,
+    "anchor_isolation": "fresh_server_per_rate_per_round",
+    "target_rate_warmup_requests": "min(32, max(4, ceil(rate * 2)))",
+    "input_tokens": 2048,
+    "output_tokens": 1,
+    "ttft_slo_ms": 1256.0,
+    "target_pass_rate": 0.95,
+}, indent=2, sort_keys=True))
+PY
+nvidia-smi --query-gpu=index,name,uuid,driver_version --format=csv,noheader \
+  > "${OUTPUT_ROOT}/provenance/gpus.csv"
+sha256sum "${MODEL_ROOT}/config.json" > "${OUTPUT_ROOT}/provenance/model.sha256"
+
+export TOKENIZERS_PARALLELISM=false
+export VLLM_USE_V1=1
+export TORCH_CUDA_ARCH_LIST=9.0
+export HF_HUB_OFFLINE=1
+export TRANSFORMERS_OFFLINE=1
+
+for ROUND in 1 2; do
+  ROUND_ROOT="${OUTPUT_ROOT}/round${ROUND}"
+  mkdir -p "${ROUND_ROOT}/logs" "${ROUND_ROOT}/results"
+  ORDERED_RATES=("${RATE_ARRAY[@]}")
+  if [[ "${ROUND}" -eq 2 ]]; then
+    ORDERED_RATES=()
+    for ((index=${#RATE_ARRAY[@]}-1; index>=0; index--)); do
+      ORDERED_RATES+=("${RATE_ARRAY[index]}")
+    done
+  fi
+  for RATE in "${ORDERED_RATES[@]}"; do
+    KEY="$(printf 'r%.2f' "${RATE}" | tr '.' 'p')"
+    SERVER_LOG="${ROUND_ROOT}/logs/server_${KEY}.log"
+    setsid "${VENV_ROOT}/bin/vllm" serve "${MODEL_ROOT}" \
+      --host 127.0.0.1 --port "${SERVER_PORT}" --served-model-name "${SERVED_MODEL}" \
+      --tensor-parallel-size "${TP}" --gpu-memory-utilization 0.92 \
+      --max-model-len 40960 --max-num-batched-tokens 8192 --max-num-seqs "${MNS}" \
+      --no-enable-prefix-caching --enable-chunked-prefill --no-enable-log-requests \
+      > "${SERVER_LOG}" 2>&1 &
+    SERVER_PID=$!
+    READY=0
+    for _ in $(seq 1 120); do
+      if curl -fsS --max-time 2 "http://127.0.0.1:${SERVER_PORT}/v1/models" \
+        > "${ROUND_ROOT}/results/models_${KEY}.json" 2>/dev/null; then
+        READY=1
+        break
+      fi
+      if ! kill -0 "${SERVER_PID}" 2>/dev/null; then
+        tail -200 "${SERVER_LOG}"
+        exit 1
+      fi
+      sleep 3
+    done
+    if [[ "${READY}" -ne 1 ]]; then
+      tail -200 "${SERVER_LOG}"
+      exit 1
+    fi
+
+    WARMUP_REQUESTS="$("${VENV_ROOT}/bin/python" - "${RATE}" <<'PY'
+import math
+import sys
+print(min(32, max(4, math.ceil(float(sys.argv[1]) * 2.0))))
+PY
+)"
+    "${VENV_ROOT}/bin/python" qwen30_prefill_client.py --port "${SERVER_PORT}" \
+      --served-model "${SERVED_MODEL}" --model-path "${MODEL_ROOT}" --rate "${RATE}" \
+      --requests "${WARMUP_REQUESTS}" \
+      --output "${ROUND_ROOT}/results/warmup_${KEY}.json"
+    "${VENV_ROOT}/bin/python" qwen30_prefill_client.py --port "${SERVER_PORT}" \
+      --served-model "${SERVED_MODEL}" --model-path "${MODEL_ROOT}" --rate "${RATE}" \
+      --requests 64 --output "${ROUND_ROOT}/results/${KEY}.json"
+    cleanup
+  done
+done
+
+find "${OUTPUT_ROOT}" -type f ! -path '*/provenance/artifacts.sha256' -print0 \
+  | sort -z | xargs -0 sha256sum > "${OUTPUT_ROOT}/provenance/artifacts.sha256"
+date -u +"END_UTC=%Y-%m-%dT%H:%M:%SZ"
+echo "QWEN30_PREFILL_REAL_CONFIG_COMPLETE"
diff --git a/runs/frontier-phase-factorial-v0/simulator-tp1/frontier_surface_frozen.json b/runs/frontier-phase-factorial-v0/simulator-tp1/frontier_surface_frozen.json
new file mode 100644
index 0000000..0dafcdd
--- /dev/null
+++ b/runs/frontier-phase-factorial-v0/simulator-tp1/frontier_surface_frozen.json
@@ -0,0 +1,647 @@
+{
+  "capacity": [
+    {
+      "config": {
+        "mns": 16,
+        "name": "tp1_mns16",
+        "tp": 1
+      },
+      "lower_censored": false,
+      "maximum_tested_feasible_request_rate": 8.0,
+      "maximum_tested_feasible_request_rate_per_gpu": 8.0,
+      "upper_censored": false
+    },
+    {
+      "config": {
+        "mns": 32,
+        "name": "tp1_mns32",
+        "tp": 1
+      },
+      "lower_censored": false,
+      "maximum_tested_feasible_request_rate": 8.0,
+      "maximum_tested_feasible_request_rate_per_gpu": 8.0,
+      "upper_censored": false
+    },
+    {
+      "config": {
+        "mns": 64,
+        "name": "tp1_mns64",
+        "tp": 1
+      },
+      "lower_censored": false,
+      "maximum_tested_feasible_request_rate": 8.0,
+      "maximum_tested_feasible_request_rate_per_gpu": 8.0,
+      "upper_censored": false
+    },
+    {
+      "config": {
+        "mns": 8,
+        "name": "tp1_mns8",
+        "tp": 1
+      },
+      "lower_censored": false,
+      "maximum_tested_feasible_request_rate": 8.0,
+      "maximum_tested_feasible_request_rate_per_gpu": 8.0,
+      "upper_censored": false
+    }
+  ],
+  "config_results": [
+    {
+      "config": {
+        "mns": 8,
+        "name": "tp1_mns8",
+        "tp": 1
+      },
+      "loads": [
+        {
+          "config": {
+            "mns": 8,
+            "name": "tp1_mns8",
+            "tp": 1
+          },
+          "elapsed_seconds": 20.451046466827393,
+          "offered_request_rate": 4.0,
+          "offered_request_rate_per_gpu": 4.0,
+          "request_count": 64,
+          "request_metrics_sha256": "f0cfb072afcc2fc82a92c59ddb4e8cfc1e3201e15848f37785eeb9fdd975e779",
+          "score": {
+            "feasible": true,
+            "pass_rate": 1.0,
+            "passed": 64,
+            "throughput_requests_per_second": 4.0197002243768605,
+            "ttft_max_ms": 171.58529929215405,
+            "ttft_p50_ms": 171.58529929215405,
+            "ttft_p95_ms": 171.58529929215405
+          },
+          "status": "completed",
+          "trace_sha256": "0cef631f351f5afd880172d0467d62d9813de5ba36fd3795f91afe3256450c05"
+        },
+        {
+          "config": {
+            "mns": 8,
+            "name": "tp1_mns8",
+            "tp": 1
+          },
+          "elapsed_seconds": 22.0544171333313,
+          "offered_request_rate": 8.0,
+          "offered_request_rate_per_gpu": 8.0,
+          "request_count": 64,
+          "request_metrics_sha256": "cfebfbd8aa5d9603158d8ff4e786119433b4c0d705f733ba242da46a6a6fdfb0",
+          "score": {
+            "feasible": true,
+            "pass_rate": 1.0,
+            "passed": 64,
+            "throughput_requests_per_second": 7.742175153637853,
+            "ttft_max_ms": 570.3133207139617,
+            "ttft_p50_ms": 391.4107605377559,
+            "ttft_p95_ms": 516.4107605377559
+          },
+          "status": "completed",
+          "trace_sha256": "c49314cbe4e4b07a9335b57e861ec5b6541ea4cc061ca9abf923453276c419fd"
+        },
+        {
+          "config": {
+            "mns": 8,
+            "name": "tp1_mns8",
+            "tp": 1
+          },
+          "elapsed_seconds": 18.859471321105957,
+          "offered_request_rate": 16.0,
+          "offered_request_rate_per_gpu": 16.0,
+          "request_count": 64,
+          "request_metrics_sha256": "03578ba406b3299e70ba7975e6b9872d381a5f4bd4fbc3c8f6a8ba49e9b53f2e",
+          "score": {
+            "feasible": false,
+            "pass_rate": 0.3125,
+            "passed": 20,
+            "throughput_requests_per_second": 8.916224898348922,
+            "ttft_max_ms": 3318.8404181099845,
+            "ttft_p50_ms": 1814.2305264490028,
+            "ttft_p95_ms": 3193.8404181099845
+          },
+          "status": "completed",
+          "trace_sha256": "2504b6340f764e726110e48251f8017a2f9ba67b44b8af7d3a838bf674823d33"
+        },
+        {
+          "config": {
+            "mns": 8,
+            "name": "tp1_mns8",
+            "tp": 1
+          },
+          "elapsed_seconds": 16.640390872955322,
+          "offered_request_rate": 32.0,
+          "offered_request_rate_per_gpu": 32.0,
+          "request_count": 64,
+          "request_metrics_sha256": "0b761b5f42e65b694eb34735433d43b473ed67415fb4e17b89894fc840b645fc",
+          "score": {
+            "feasible": false,
+            "pass_rate": 0.203125,
+            "passed": 13,
+            "throughput_requests_per_second": 9.07297385348011,
+            "ttft_max_ms": 5147.666503402199,
+            "ttft_p50_ms": 2739.3537469047046,
+            "ttft_p95_ms": 4962.713638565687
+          },
+          "status": "completed",
+          "trace_sha256": "5508a3ce590f7d1b0b3c211a2d4f2131886a7ae2d85cafa0c436a8e3029675f1"
+        },
+        {
+          "config": {
+            "mns": 8,
+            "name": "tp1_mns8",
+            "tp": 1
+          },
+          "elapsed_seconds": 18.142091274261475,
+          "offered_request_rate": 64.0,
+          "offered_request_rate_per_gpu": 64.0,
+          "request_count": 64,
+          "request_metrics_sha256": "078209f2f56f21243384bdc70add105bf4942ee44ec8cc9573cad399a4a10e03",
+          "score": {
+            "feasible": false,
+            "pass_rate": 0.140625,
+            "passed": 9,
+            "throughput_requests_per_second": 9.07297385348011,
+            "ttft_max_ms": 6100.791503402199,
+            "ttft_p50_ms": 3208.1037469047046,
+            "ttft_p95_ms": 5853.338638565687
+          },
+          "status": "completed",
+          "trace_sha256": "8987c3ebd4a7db983dc0e99e44710cd44a9d5d21cdb8e6b58878b160918593d9"
+        }
+      ]
+    },
+    {
+      "config": {
+        "mns": 16,
+        "name": "tp1_mns16",
+        "tp": 1
+      },
+      "loads": [
+        {
+          "config": {
+            "mns": 16,
+            "name": "tp1_mns16",
+            "tp": 1
+          },
+          "elapsed_seconds": 19.96323037147522,
+          "offered_request_rate": 4.0,
+          "offered_request_rate_per_gpu": 4.0,
+          "request_count": 64,
+          "request_metrics_sha256": "f0cfb072afcc2fc82a92c59ddb4e8cfc1e3201e15848f37785eeb9fdd975e779",
+          "score": {
+            "feasible": true,
+            "pass_rate": 1.0,
+            "passed": 64,
+            "throughput_requests_per_second": 4.0197002243768605,
+            "ttft_max_ms": 171.58529929215405,
+            "ttft_p50_ms": 171.58529929215405,
+            "ttft_p95_ms": 171.58529929215405
+          },
+          "status": "completed",
+          "trace_sha256": "0cef631f351f5afd880172d0467d62d9813de5ba36fd3795f91afe3256450c05"
+        },
+        {
+          "config": {
+            "mns": 16,
+            "name": "tp1_mns16",
+            "tp": 1
+          },
+          "elapsed_seconds": 18.554351568222046,
+          "offered_request_rate": 8.0,
+          "offered_request_rate_per_gpu": 8.0,
+          "request_count": 64,
+          "request_metrics_sha256": "cfebfbd8aa5d9603158d8ff4e786119433b4c0d705f733ba242da46a6a6fdfb0",
+          "score": {
+            "feasible": true,
+            "pass_rate": 1.0,
+            "passed": 64,
+            "throughput_requests_per_second": 7.742175153637853,
+            "ttft_max_ms": 570.3133207139617,
+            "ttft_p50_ms": 391.4107605377559,
+            "ttft_p95_ms": 516.4107605377559
+          },
+          "status": "completed",
+          "trace_sha256": "c49314cbe4e4b07a9335b57e861ec5b6541ea4cc061ca9abf923453276c419fd"
+        },
+        {
+          "config": {
+            "mns": 16,
+            "name": "tp1_mns16",
+            "tp": 1
+          },
+          "elapsed_seconds": 25.70284128189087,
+          "offered_request_rate": 16.0,
+          "offered_request_rate_per_gpu": 16.0,
+          "request_count": 64,
+          "request_metrics_sha256": "03578ba406b3299e70ba7975e6b9872d381a5f4bd4fbc3c8f6a8ba49e9b53f2e",
+          "score": {
+            "feasible": false,
+            "pass_rate": 0.3125,
+            "passed": 20,
+            "throughput_requests_per_second": 8.916224898348922,
+            "ttft_max_ms": 3318.8404181099845,
+            "ttft_p50_ms": 1814.2305264490028,
+            "ttft_p95_ms": 3193.8404181099845
+          },
+          "status": "completed",
+          "trace_sha256": "2504b6340f764e726110e48251f8017a2f9ba67b44b8af7d3a838bf674823d33"
+        },
+        {
+          "config": {
+            "mns": 16,
+            "name": "tp1_mns16",
+            "tp": 1
+          },
+          "elapsed_seconds": 20.056548833847046,
+          "offered_request_rate": 32.0,
+          "offered_request_rate_per_gpu": 32.0,
+          "request_count": 64,
+          "request_metrics_sha256": "0b761b5f42e65b694eb34735433d43b473ed67415fb4e17b89894fc840b645fc",
+          "score": {
+            "feasible": false,
+            "pass_rate": 0.203125,
+            "passed": 13,
+            "throughput_requests_per_second": 9.07297385348011,
+            "ttft_max_ms": 5147.666503402199,
+            "ttft_p50_ms": 2739.3537469047046,
+            "ttft_p95_ms": 4962.713638565687
+          },
+          "status": "completed",
+          "trace_sha256": "5508a3ce590f7d1b0b3c211a2d4f2131886a7ae2d85cafa0c436a8e3029675f1"
+        },
+        {
+          "config": {
+            "mns": 16,
+            "name": "tp1_mns16",
+            "tp": 1
+          },
+          "elapsed_seconds": 11.15293574333191,
+          "offered_request_rate": 64.0,
+          "offered_request_rate_per_gpu": 64.0,
+          "request_count": 64,
+          "request_metrics_sha256": "078209f2f56f21243384bdc70add105bf4942ee44ec8cc9573cad399a4a10e03",
+          "score": {
+            "feasible": false,
+            "pass_rate": 0.140625,
+            "passed": 9,
+            "throughput_requests_per_second": 9.07297385348011,
+            "ttft_max_ms": 6100.791503402199,
+            "ttft_p50_ms": 3208.1037469047046,
+            "ttft_p95_ms": 5853.338638565687
+          },
+          "status": "completed",
+          "trace_sha256": "8987c3ebd4a7db983dc0e99e44710cd44a9d5d21cdb8e6b58878b160918593d9"
+        }
+      ]
+    },
+    {
+      "config": {
+        "mns": 32,
+        "name": "tp1_mns32",
+        "tp": 1
+      },
+      "loads": [
+        {
+          "config": {
+            "mns": 32,
+            "name": "tp1_mns32",
+            "tp": 1
+          },
+          "elapsed_seconds": 10.411258935928345,
+          "offered_request_rate": 4.0,
+          "offered_request_rate_per_gpu": 4.0,
+          "request_count": 64,
+          "request_metrics_sha256": "f0cfb072afcc2fc82a92c59ddb4e8cfc1e3201e15848f37785eeb9fdd975e779",
+          "score": {
+            "feasible": true,
+            "pass_rate": 1.0,
+            "passed": 64,
+            "throughput_requests_per_second": 4.0197002243768605,
+            "ttft_max_ms": 171.58529929215405,
+            "ttft_p50_ms": 171.58529929215405,
+            "ttft_p95_ms": 171.58529929215405
+          },
+          "status": "completed",
+          "trace_sha256": "0cef631f351f5afd880172d0467d62d9813de5ba36fd3795f91afe3256450c05"
+        },
+        {
+          "config": {
+            "mns": 32,
+            "name": "tp1_mns32",
+            "tp": 1
+          },
+          "elapsed_seconds": 10.355234384536743,
+          "offered_request_rate": 8.0,
+          "offered_request_rate_per_gpu": 8.0,
+          "request_count": 64,
+          "request_metrics_sha256": "cfebfbd8aa5d9603158d8ff4e786119433b4c0d705f733ba242da46a6a6fdfb0",
+          "score": {
+            "feasible": true,
+            "pass_rate": 1.0,
+            "passed": 64,
+            "throughput_requests_per_second": 7.742175153637853,
+            "ttft_max_ms": 570.3133207139617,
+            "ttft_p50_ms": 391.4107605377559,
+            "ttft_p95_ms": 516.4107605377559
+          },
+          "status": "completed",
+          "trace_sha256": "c49314cbe4e4b07a9335b57e861ec5b6541ea4cc061ca9abf923453276c419fd"
+        },
+        {
+          "config": {
+            "mns": 32,
+            "name": "tp1_mns32",
+            "tp": 1
+          },
+          "elapsed_seconds": 10.098335266113281,
+          "offered_request_rate": 16.0,
+          "offered_request_rate_per_gpu": 16.0,
+          "request_count": 64,
+          "request_metrics_sha256": "03578ba406b3299e70ba7975e6b9872d381a5f4bd4fbc3c8f6a8ba49e9b53f2e",
+          "score": {
+            "feasible": false,
+            "pass_rate": 0.3125,
+            "passed": 20,
+            "throughput_requests_per_second": 8.916224898348922,
+            "ttft_max_ms": 3318.8404181099845,
+            "ttft_p50_ms": 1814.2305264490028,
+            "ttft_p95_ms": 3193.8404181099845
+          },
+          "status": "completed",
+          "trace_sha256": "2504b6340f764e726110e48251f8017a2f9ba67b44b8af7d3a838bf674823d33"
+        },
+        {
+          "config": {
+            "mns": 32,
+            "name": "tp1_mns32",
+            "tp": 1
+          },
+          "elapsed_seconds": 9.8981294631958,
+          "offered_request_rate": 32.0,
+          "offered_request_rate_per_gpu": 32.0,
+          "request_count": 64,
+          "request_metrics_sha256": "0b761b5f42e65b694eb34735433d43b473ed67415fb4e17b89894fc840b645fc",
+          "score": {
+            "feasible": false,
+            "pass_rate": 0.203125,
+            "passed": 13,
+            "throughput_requests_per_second": 9.07297385348011,
+            "ttft_max_ms": 5147.666503402199,
+            "ttft_p50_ms": 2739.3537469047046,
+            "ttft_p95_ms": 4962.713638565687
+          },
+          "status": "completed",
+          "trace_sha256": "5508a3ce590f7d1b0b3c211a2d4f2131886a7ae2d85cafa0c436a8e3029675f1"
+        },
+        {
+          "config": {
+            "mns": 32,
+            "name": "tp1_mns32",
+            "tp": 1
+          },
+          "elapsed_seconds": 9.754717826843262,
+          "offered_request_rate": 64.0,
+          "offered_request_rate_per_gpu": 64.0,
+          "request_count": 64,
+          "request_metrics_sha256": "078209f2f56f21243384bdc70add105bf4942ee44ec8cc9573cad399a4a10e03",
+          "score": {
+            "feasible": false,
+            "pass_rate": 0.140625,
+            "passed": 9,
+            "throughput_requests_per_second": 9.07297385348011,
+            "ttft_max_ms": 6100.791503402199,
+            "ttft_p50_ms": 3208.1037469047046,
+            "ttft_p95_ms": 5853.338638565687
+          },
+          "status": "completed",
+          "trace_sha256": "8987c3ebd4a7db983dc0e99e44710cd44a9d5d21cdb8e6b58878b160918593d9"
+        }
+      ]
+    },
+    {
+      "config": {
+        "mns": 64,
+        "name": "tp1_mns64",
+        "tp": 1
+      },
+      "loads": [
+        {
+          "config": {
+            "mns": 64,
+            "name": "tp1_mns64",
+            "tp": 1
+          },
+          "elapsed_seconds": 9.84917163848877,
+          "offered_request_rate": 4.0,
+          "offered_request_rate_per_gpu": 4.0,
+          "request_count": 64,
+          "request_metrics_sha256": "f0cfb072afcc2fc82a92c59ddb4e8cfc1e3201e15848f37785eeb9fdd975e779",
+          "score": {
+            "feasible": true,
+            "pass_rate": 1.0,
+            "passed": 64,
+            "throughput_requests_per_second": 4.0197002243768605,
+            "ttft_max_ms": 171.58529929215405,
+            "ttft_p50_ms": 171.58529929215405,
+            "ttft_p95_ms": 171.58529929215405
+          },
+          "status": "completed",
+          "trace_sha256": "0cef631f351f5afd880172d0467d62d9813de5ba36fd3795f91afe3256450c05"
+        },
+        {
+          "config": {
+            "mns": 64,
+            "name": "tp1_mns64",
+            "tp": 1
+          },
+          "elapsed_seconds": 10.197595357894897,
+          "offered_request_rate": 8.0,
+          "offered_request_rate_per_gpu": 8.0,
+          "request_count": 64,
+          "request_metrics_sha256": "cfebfbd8aa5d9603158d8ff4e786119433b4c0d705f733ba242da46a6a6fdfb0",
+          "score": {
+            "feasible": true,
+            "pass_rate": 1.0,
+            "passed": 64,
+            "throughput_requests_per_second": 7.742175153637853,
+            "ttft_max_ms": 570.3133207139617,
+            "ttft_p50_ms": 391.4107605377559,
+            "ttft_p95_ms": 516.4107605377559
+          },
+          "status": "completed",
+          "trace_sha256": "c49314cbe4e4b07a9335b57e861ec5b6541ea4cc061ca9abf923453276c419fd"
+        },
+        {
+          "config": {
+            "mns": 64,
+            "name": "tp1_mns64",
+            "tp": 1
+          },
+          "elapsed_seconds": 10.00463604927063,
+          "offered_request_rate": 16.0,
+          "offered_request_rate_per_gpu": 16.0,
+          "request_count": 64,
+          "request_metrics_sha256": "03578ba406b3299e70ba7975e6b9872d381a5f4bd4fbc3c8f6a8ba49e9b53f2e",
+          "score": {
+            "feasible": false,
+            "pass_rate": 0.3125,
+            "passed": 20,
+            "throughput_requests_per_second": 8.916224898348922,
+            "ttft_max_ms": 3318.8404181099845,
+            "ttft_p50_ms": 1814.2305264490028,
+            "ttft_p95_ms": 3193.8404181099845
+          },
+          "status": "completed",
+          "trace_sha256": "2504b6340f764e726110e48251f8017a2f9ba67b44b8af7d3a838bf674823d33"
+        },
+        {
+          "config": {
+            "mns": 64,
+            "name": "tp1_mns64",
+            "tp": 1
+          },
+          "elapsed_seconds": 10.153925895690918,
+          "offered_request_rate": 32.0,
+          "offered_request_rate_per_gpu": 32.0,
+          "request_count": 64,
+          "request_metrics_sha256": "0b761b5f42e65b694eb34735433d43b473ed67415fb4e17b89894fc840b645fc",
+          "score": {
+            "feasible": false,
+            "pass_rate": 0.203125,
+            "passed": 13,
+            "throughput_requests_per_second": 9.07297385348011,
+            "ttft_max_ms": 5147.666503402199,
+            "ttft_p50_ms": 2739.3537469047046,
+            "ttft_p95_ms": 4962.713638565687
+          },
+          "status": "completed",
+          "trace_sha256": "5508a3ce590f7d1b0b3c211a2d4f2131886a7ae2d85cafa0c436a8e3029675f1"
+        },
+        {
+          "config": {
+            "mns": 64,
+            "name": "tp1_mns64",
+            "tp": 1
+          },
+          "elapsed_seconds": 10.055675268173218,
+          "offered_request_rate": 64.0,
+          "offered_request_rate_per_gpu": 64.0,
+          "request_count": 64,
+          "request_metrics_sha256": "078209f2f56f21243384bdc70add105bf4942ee44ec8cc9573cad399a4a10e03",
+          "score": {
+            "feasible": false,
+            "pass_rate": 0.140625,
+            "passed": 9,
+            "throughput_requests_per_second": 9.07297385348011,
+            "ttft_max_ms": 6100.791503402199,
+            "ttft_p50_ms": 3208.1037469047046,
+            "ttft_p95_ms": 5853.338638565687
+          },
+          "status": "completed",
+          "trace_sha256": "8987c3ebd4a7db983dc0e99e44710cd44a9d5d21cdb8e6b58878b160918593d9"
+        }
+      ]
+    }
+  ],
+  "contract": {
+    "arrival": "open_loop_uniform",
+    "input_tokens": 2048,
+    "output_tokens": 1,
+    "prefix_caching": false,
+    "rates": [
+      4.0,
+      8.0,
+      16.0,
+      32.0,
+      64.0
+    ],
+    "requests_per_anchor": 64,
+    "target_pass_rate": 0.95,
+    "ttft_slo_ms": 1256.0
+  },
+  "frontier": {
+    "git_head": "d9cfeb6d8791fbf2f295dd9744c56a666171776e",
+    "git_status_short": " M frontier/config/config.py\n M frontier/entities/request.py\n M frontier/events/cluster_schedule_event.py\n M frontier/execution_time_predictor/sklearn_execution_time_predictor.py\n M frontier/metrics/constants.py\n M frontier/metrics/metrics_store.py\n M frontier/profiling/common/layers/rotary_embedding.py\n M frontier/profiling/moe/moe_impl.py\n M frontier/profiling/moe/moe_vllm_kernel.py\n M frontier/scheduler/cluster_scheduler/__init__.py\n M frontier/scheduler/cluster_scheduler/base_cluster_scheduler.py\n M frontier/scheduler/cluster_scheduler/cluster_scheduler_registry.py\n M frontier/scheduler/cluster_scheduler/sticky_lor_cluster_scheduler.py\n M frontier/scheduler/replica_scheduler/base_replica_scheduler.py\n M frontier/scheduler/replica_scheduler/vllm_v1_engine_replica_scheduler.py\n M frontier/scheduler/replica_stage_scheduler/replica_stage_schduler.py\n M frontier/simulator.py\n M frontier/types/cluster_scheduler_type.py\n?? data/profiling/compute/h20/\n?? frontier/scheduler/cluster_scheduler/prefix_lor_cluster_scheduler.py\n?? runs/\n?? tests/unit/test_attn_prefill_prediction_fallback.py\n",
+    "source": "/tmp/replayserve-frontier-rs1b"
+  },
+  "profiles": {
+    "coverage": {
+      "attention": {
+        "1": {
+          "exact_prefill_2048_rows": 1,
+          "profile_batch_size": 1
+        },
+        "2": {
+          "exact_prefill_2048_rows": 1,
+          "profile_batch_size": 1
+        },
+        "4": {
+          "exact_prefill_2048_rows": 1,
+          "profile_batch_size": 1
+        }
+      },
+      "manifest": {
+        "attention_tp_coverage": [
+          1,
+          2,
+          4
+        ],
+        "environment_contract": {
+          "dtype": "bfloat16",
+          "frontier_commit": "d9cfeb6d8791fbf2f295dd9744c56a666171776e",
+          "hardware": "NVIDIA H20",
+          "model": "Qwen3-30B-A3B",
+          "tensor_parallel_sizes": [
+            1,
+            2,
+            4
+          ],
+          "vllm_source_commit": "88d34c6409e9fb3c7b8ca0c04756f061d2099eb1",
+          "vllm_version": "0.20.0"
+        },
+        "inputs": {
+          "/home/gahow/phd/aituner/runs/frontier-qwen30-vllm020-profile-v1/fleet-artifacts/qwen30-vllm020-allreduce-full-tp2-20260716-v1-dispatch-aware-20260716T140743025781Z/artifacts/artifacts/allreduce-full-tp2-v1/raw/allreduce-tp2.json": "97c3c76b5a04e95bd9192423c2b891667c668f39cc0dfecbd097d749939f2d0a",
+          "/home/gahow/phd/aituner/runs/frontier-qwen30-vllm020-profile-v1/fleet-artifacts/qwen30-vllm020-allreduce-full-tp4-20260716-v1-dispatch-aware-20260716T141106009788Z/artifacts/artifacts/allreduce-full-tp4-v1/raw/allreduce-tp4.json": "809df9baa6f468cf12bf0c99827475acc67894dd9f3f948976590b665fac0e76",
+          "/home/gahow/phd/aituner/runs/frontier-qwen30-vllm020-profile-v1/fleet-artifacts/qwen30-vllm020-flashattn-kv-full-tp1-20260716-v2-20260716T135132587012Z/artifacts/artifacts/flashattn-kv-full-v2-tp1/raw/flashattn-tp1.json": "dcb4c1bf7e76b9c765f78ddd2b8a734f2d7ba2adac13ce017689a8a77fe69a27",
+          "/home/gahow/phd/aituner/runs/frontier-qwen30-vllm020-profile-v1/fleet-artifacts/qwen30-vllm020-flashattn-kv-full-tp2-20260716-v2-20260716T135134194295Z/artifacts/artifacts/flashattn-kv-full-v2-tp2/raw/flashattn-tp2.json": "43ce042556ba887c8860614b43ccf0f564e5cebc1a0cffbce299d0acb9fa8d07",
+          "/home/gahow/phd/aituner/runs/frontier-qwen30-vllm020-profile-v1/fleet-artifacts/qwen30-vllm020-flashattn-kv-full-tp4-20260716-v2-20260716T135135197200Z/artifacts/artifacts/flashattn-kv-full-v2-tp4/raw/flashattn-tp4.json": "84eef31bcad0f556907a093318a420959d14fdc94474823d11f659704bdfec73",
+          "/home/gahow/phd/aituner/runs/frontier-qwen30-vllm020-profile-v1/fleet-artifacts/qwen30-vllm020-frontier-linear-full-20260716-v2-max-tokens-20260716T144444676943Z/artifacts/artifacts/frontier-linear-full-v2/profiles/compute/h20/qwen3-a3b-30b-moe/linear_op.csv": "67666cb0a4901b74599d468df2e31bcaa2a11a7842cc0cefba24ffce62508e0c",
+          "/home/gahow/phd/aituner/runs/frontier-qwen30-vllm020-profile-v1/fleet-artifacts/qwen30-vllm020-moe-full-20260716-v1-local-shard-20260716T141334565164Z/artifacts/artifacts/moe-full-v1/raw/moe-full.json": "588f6ad0d69c9636d1b852e3df0a12d13cfe731f050ea7ec7aea457cceefbde8",
+          "/home/gahow/phd/aituner/runs/frontier-qwen30-vllm020-profile-v1/fleet-artifacts/qwen30-vllm020-router-full-20260716-v3-tp-context-20260716T145446098505Z/artifacts/artifacts/router-full-v3/raw/router.json": "1962972e983bff3e06a721ef4ae4ec65728ff669681497a4a7e7f769b88b4931"
+        },
+        "outputs": {
+          "allreduce.json": "b38d14f990578d668523d25b107aceed433da5020d8ada3b6e44d3562261a3b3",
+          "attention.csv": "76dcb767cebb4ec1c4e24bd04d93ddd48b5d271986ebfb51a197ab33e1b3d87d",
+          "attention_true_mixed_fused.csv": "43ef4be90bddc9aeac6dbbe339feec24162cd1f2129a08fbd959e6ee4eaf5f60",
+          "linear_op.csv": "67666cb0a4901b74599d468df2e31bcaa2a11a7842cc0cefba24ffce62508e0c",
+          "moe.csv": "0e4dcba72918a1c4cf4e96ced31ee3829248a19ad54553cebef14417725808b0"
+        },
+        "profile_id": "qwen3-30b-a3b-bf16-vllm020-h20-tp1-2-4-fused-mixed-total-conserving",
+        "projection_contract": {
+          "allreduce": "Frozen exact runtime measurements; base profile-only comparison keeps the historical Frontier CC backend fixed to isolate compute profile fidelity",
+          "attention": "Pure prefill/extend/decode FA3 core plus separately measured KV update; input/output reshape assumed zero; exported mean is used as median target; true mixed rows use a total-conserving compatibility projection",
+          "attention_true_mixed": "The directly measured fused total is preserved in diagnostics. Frontier's two targets are projected by the same-TP pure prefill/decode reference ratio, with projected prefill + decode exactly equal to the fused total; the split is a schema compatibility attribution, not an observation",
+          "linear": "Frontier profiler using vLLM 0.20 CUDA operators",
+          "moe": "Replicated gate and fused top-k plus TP-local modular expert kernel; expert measurement already includes prepare/finalize so shuffling is zero"
+        },
+        "row_counts": {
+          "allreduce": 24,
+          "attention_frontier_compatible": 132,
+          "attention_true_mixed_fused_diagnostic": 30,
+          "linear": 36,
+          "moe": 72
+        },
+        "schema_version": "frontier_qwen30_vllm020_frozen_profile.v2"
+      }
+    },
+    "root": "/home/gahow/phd/aituner/runs/frontier-qwen30-vllm020-profile-v1/frozen/profile-v2",
+    "sha256": {
+      "attention": "76dcb767cebb4ec1c4e24bd04d93ddd48b5d271986ebfb51a197ab33e1b3d87d",
+      "linear": "67666cb0a4901b74599d468df2e31bcaa2a11a7842cc0cefba24ffce62508e0c",
+      "manifest": "af40545e75aff55c6333cd2d5379ccf042a5a0b7d7fc7df4f745ce256cb290eb",
+      "moe": "0e4dcba72918a1c4cf4e96ced31ee3829248a19ad54553cebef14417725808b0"
+    }
+  },
+  "schema": "frontier-qwen30-prefill-surface-v1",
+  "status": "partial_not_decision_bearing"
+}
diff --git a/runs/frontier-phase-factorial-v0/simulator-tp2/frontier_surface_frozen.json b/runs/frontier-phase-factorial-v0/simulator-tp2/frontier_surface_frozen.json
new file mode 100644
index 0000000..34e2063
--- /dev/null
+++ b/runs/frontier-phase-factorial-v0/simulator-tp2/frontier_surface_frozen.json
@@ -0,0 +1,647 @@
+{
+  "capacity": [
+    {
+      "config": {
+        "mns": 16,
+        "name": "tp2_mns16",
+        "tp": 2
+      },
+      "lower_censored": false,
+      "maximum_tested_feasible_request_rate": 16.0,
+      "maximum_tested_feasible_request_rate_per_gpu": 8.0,
+      "upper_censored": false
+    },
+    {
+      "config": {
+        "mns": 32,
+        "name": "tp2_mns32",
+        "tp": 2
+      },
+      "lower_censored": false,
+      "maximum_tested_feasible_request_rate": 16.0,
+      "maximum_tested_feasible_request_rate_per_gpu": 8.0,
+      "upper_censored": false
+    },
+    {
+      "config": {
+        "mns": 64,
+        "name": "tp2_mns64",
+        "tp": 2
+      },
+      "lower_censored": false,
+      "maximum_tested_feasible_request_rate": 16.0,
+      "maximum_tested_feasible_request_rate_per_gpu": 8.0,
+      "upper_censored": false
+    },
+    {
+      "config": {
+        "mns": 8,
+        "name": "tp2_mns8",
+        "tp": 2
+      },
+      "lower_censored": false,
+      "maximum_tested_feasible_request_rate": 16.0,
+      "maximum_tested_feasible_request_rate_per_gpu": 8.0,
+      "upper_censored": false
+    }
+  ],
+  "config_results": [
+    {
+      "config": {
+        "mns": 8,
+        "name": "tp2_mns8",
+        "tp": 2
+      },
+      "loads": [
+        {
+          "config": {
+            "mns": 8,
+            "name": "tp2_mns8",
+            "tp": 2
+          },
+          "elapsed_seconds": 264.1678559780121,
+          "offered_request_rate": 4.0,
+          "offered_request_rate_per_gpu": 2.0,
+          "request_count": 64,
+          "request_metrics_sha256": "e5f6527532682ee5aca3f423367d14eba063f30e63f46d2ba0ed602db4f6eb17",
+          "score": {
+            "feasible": true,
+            "pass_rate": 1.0,
+            "passed": 64,
+            "throughput_requests_per_second": 4.032196934357252,
+            "ttft_max_ms": 122.24062760264154,
+            "ttft_p50_ms": 122.24062760264154,
+            "ttft_p95_ms": 122.24062760264154
+          },
+          "status": "completed",
+          "trace_sha256": "0cef631f351f5afd880172d0467d62d9813de5ba36fd3795f91afe3256450c05"
+        },
+        {
+          "config": {
+            "mns": 8,
+            "name": "tp2_mns8",
+            "tp": 2
+          },
+          "elapsed_seconds": 10.15904188156128,
+          "offered_request_rate": 8.0,
+          "offered_request_rate_per_gpu": 4.0,
+          "request_count": 64,
+          "request_metrics_sha256": "8dd3bedcbf76d30bf4285e10b886f95bef9ae8e5d41f4bb50a7650b322b8eb83",
+          "score": {
+            "feasible": true,
+            "pass_rate": 1.0,
+            "passed": 64,
+            "throughput_requests_per_second": 8.00276032449276,
+            "ttft_max_ms": 122.24062760264154,
+            "ttft_p50_ms": 122.24062760264154,
+            "ttft_p95_ms": 122.24062760264154
+          },
+          "status": "completed",
+          "trace_sha256": "c49314cbe4e4b07a9335b57e861ec5b6541ea4cc061ca9abf923453276c419fd"
+        },
+        {
+          "config": {
+            "mns": 8,
+            "name": "tp2_mns8",
+            "tp": 2
+          },
+          "elapsed_seconds": 10.447325944900513,
+          "offered_request_rate": 16.0,
+          "offered_request_rate_per_gpu": 8.0,
+          "request_count": 64,
+          "request_metrics_sha256": "38c1b13c1f5fbdabd7f79b82ccb1fb37fce6840bcf03614c988d7e4882814506",
+          "score": {
+            "feasible": true,
+            "pass_rate": 1.0,
+            "passed": 64,
+            "throughput_requests_per_second": 13.174042777348715,
+            "ttft_max_ms": 1048.2973400621365,
+            "ttft_p50_ms": 643.9195307403329,
+            "ttft_p95_ms": 962.8278877316862
+          },
+          "status": "completed",
+          "trace_sha256": "2504b6340f764e726110e48251f8017a2f9ba67b44b8af7d3a838bf674823d33"
+        },
+        {
+          "config": {
+            "mns": 8,
+            "name": "tp2_mns8",
+            "tp": 2
+          },
+          "elapsed_seconds": 10.053350687026978,
+          "offered_request_rate": 32.0,
+          "offered_request_rate_per_gpu": 16.0,
+          "request_count": 64,
+          "request_metrics_sha256": "cb3b7b5a2fcf63cd01304b3b530bb21c0a9bb1c8f5e06d489e3863afd0b4d2f3",
+          "score": {
+            "feasible": false,
+            "pass_rate": 0.390625,
+            "passed": 25,
+            "throughput_requests_per_second": 13.539443740959392,
+            "ttft_max_ms": 2851.9297930156335,
+            "ttft_p50_ms": 1510.0519836938302,
+            "ttft_p95_ms": 2758.1797930156335
+          },
+          "status": "completed",
+          "trace_sha256": "5508a3ce590f7d1b0b3c211a2d4f2131886a7ae2d85cafa0c436a8e3029675f1"
+        },
+        {
+          "config": {
+            "mns": 8,
+            "name": "tp2_mns8",
+            "tp": 2
+          },
+          "elapsed_seconds": 9.744416952133179,
+          "offered_request_rate": 64.0,
+          "offered_request_rate_per_gpu": 32.0,
+          "request_count": 64,
+          "request_metrics_sha256": "a4474c98a68171ada0fb007c36fcf69e9e2dc13f06c74889591e23f93418bd5a",
+          "score": {
+            "feasible": false,
+            "pass_rate": 0.265625,
+            "passed": 17,
+            "throughput_requests_per_second": 13.539443740959392,
+            "ttft_max_ms": 3773.8047930156335,
+            "ttft_p50_ms": 1995.3684369244447,
+            "ttft_p95_ms": 3622.636520081023
+          },
+          "status": "completed",
+          "trace_sha256": "8987c3ebd4a7db983dc0e99e44710cd44a9d5d21cdb8e6b58878b160918593d9"
+        }
+      ]
+    },
+    {
+      "config": {
+        "mns": 16,
+        "name": "tp2_mns16",
+        "tp": 2
+      },
+      "loads": [
+        {
+          "config": {
+            "mns": 16,
+            "name": "tp2_mns16",
+            "tp": 2
+          },
+          "elapsed_seconds": 9.402071952819824,
+          "offered_request_rate": 4.0,
+          "offered_request_rate_per_gpu": 2.0,
+          "request_count": 64,
+          "request_metrics_sha256": "e5f6527532682ee5aca3f423367d14eba063f30e63f46d2ba0ed602db4f6eb17",
+          "score": {
+            "feasible": true,
+            "pass_rate": 1.0,
+            "passed": 64,
+            "throughput_requests_per_second": 4.032196934357252,
+            "ttft_max_ms": 122.24062760264154,
+            "ttft_p50_ms": 122.24062760264154,
+            "ttft_p95_ms": 122.24062760264154
+          },
+          "status": "completed",
+          "trace_sha256": "0cef631f351f5afd880172d0467d62d9813de5ba36fd3795f91afe3256450c05"
+        },
+        {
+          "config": {
+            "mns": 16,
+            "name": "tp2_mns16",
+            "tp": 2
+          },
+          "elapsed_seconds": 9.398229360580444,
+          "offered_request_rate": 8.0,
+          "offered_request_rate_per_gpu": 4.0,
+          "request_count": 64,
+          "request_metrics_sha256": "8dd3bedcbf76d30bf4285e10b886f95bef9ae8e5d41f4bb50a7650b322b8eb83",
+          "score": {
+            "feasible": true,
+            "pass_rate": 1.0,
+            "passed": 64,
+            "throughput_requests_per_second": 8.00276032449276,
+            "ttft_max_ms": 122.24062760264154,
+            "ttft_p50_ms": 122.24062760264154,
+            "ttft_p95_ms": 122.24062760264154
+          },
+          "status": "completed",
+          "trace_sha256": "c49314cbe4e4b07a9335b57e861ec5b6541ea4cc061ca9abf923453276c419fd"
+        },
+        {
+          "config": {
+            "mns": 16,
+            "name": "tp2_mns16",
+            "tp": 2
+          },
+          "elapsed_seconds": 9.70562195777893,
+          "offered_request_rate": 16.0,
+          "offered_request_rate_per_gpu": 8.0,
+          "request_count": 64,
+          "request_metrics_sha256": "38c1b13c1f5fbdabd7f79b82ccb1fb37fce6840bcf03614c988d7e4882814506",
+          "score": {
+            "feasible": true,
+            "pass_rate": 1.0,
+            "passed": 64,
+            "throughput_requests_per_second": 13.174042777348715,
+            "ttft_max_ms": 1048.2973400621365,
+            "ttft_p50_ms": 643.9195307403329,
+            "ttft_p95_ms": 962.8278877316862
+          },
+          "status": "completed",
+          "trace_sha256": "2504b6340f764e726110e48251f8017a2f9ba67b44b8af7d3a838bf674823d33"
+        },
+        {
+          "config": {
+            "mns": 16,
+            "name": "tp2_mns16",
+            "tp": 2
+          },
+          "elapsed_seconds": 9.598989009857178,
+          "offered_request_rate": 32.0,
+          "offered_request_rate_per_gpu": 16.0,
+          "request_count": 64,
+          "request_metrics_sha256": "cb3b7b5a2fcf63cd01304b3b530bb21c0a9bb1c8f5e06d489e3863afd0b4d2f3",
+          "score": {
+            "feasible": false,
+            "pass_rate": 0.390625,
+            "passed": 25,
+            "throughput_requests_per_second": 13.539443740959392,
+            "ttft_max_ms": 2851.9297930156335,
+            "ttft_p50_ms": 1510.0519836938302,
+            "ttft_p95_ms": 2758.1797930156335
+          },
+          "status": "completed",
+          "trace_sha256": "5508a3ce590f7d1b0b3c211a2d4f2131886a7ae2d85cafa0c436a8e3029675f1"
+        },
+        {
+          "config": {
+            "mns": 16,
+            "name": "tp2_mns16",
+            "tp": 2
+          },
+          "elapsed_seconds": 9.697861671447754,
+          "offered_request_rate": 64.0,
+          "offered_request_rate_per_gpu": 32.0,
+          "request_count": 64,
+          "request_metrics_sha256": "a4474c98a68171ada0fb007c36fcf69e9e2dc13f06c74889591e23f93418bd5a",
+          "score": {
+            "feasible": false,
+            "pass_rate": 0.265625,
+            "passed": 17,
+            "throughput_requests_per_second": 13.539443740959392,
+            "ttft_max_ms": 3773.8047930156335,
+            "ttft_p50_ms": 1995.3684369244447,
+            "ttft_p95_ms": 3622.636520081023
+          },
+          "status": "completed",
+          "trace_sha256": "8987c3ebd4a7db983dc0e99e44710cd44a9d5d21cdb8e6b58878b160918593d9"
+        }
+      ]
+    },
+    {
+      "config": {
+        "mns": 32,
+        "name": "tp2_mns32",
+        "tp": 2
+      },
+      "loads": [
+        {
+          "config": {
+            "mns": 32,
+            "name": "tp2_mns32",
+            "tp": 2
+          },
+          "elapsed_seconds": 9.397430419921875,
+          "offered_request_rate": 4.0,
+          "offered_request_rate_per_gpu": 2.0,
+          "request_count": 64,
+          "request_metrics_sha256": "e5f6527532682ee5aca3f423367d14eba063f30e63f46d2ba0ed602db4f6eb17",
+          "score": {
+            "feasible": true,
+            "pass_rate": 1.0,
+            "passed": 64,
+            "throughput_requests_per_second": 4.032196934357252,
+            "ttft_max_ms": 122.24062760264154,
+            "ttft_p50_ms": 122.24062760264154,
+            "ttft_p95_ms": 122.24062760264154
+          },
+          "status": "completed",
+          "trace_sha256": "0cef631f351f5afd880172d0467d62d9813de5ba36fd3795f91afe3256450c05"
+        },
+        {
+          "config": {
+            "mns": 32,
+            "name": "tp2_mns32",
+            "tp": 2
+          },
+          "elapsed_seconds": 9.3956880569458,
+          "offered_request_rate": 8.0,
+          "offered_request_rate_per_gpu": 4.0,
+          "request_count": 64,
+          "request_metrics_sha256": "8dd3bedcbf76d30bf4285e10b886f95bef9ae8e5d41f4bb50a7650b322b8eb83",
+          "score": {
+            "feasible": true,
+            "pass_rate": 1.0,
+            "passed": 64,
+            "throughput_requests_per_second": 8.00276032449276,
+            "ttft_max_ms": 122.24062760264154,
+            "ttft_p50_ms": 122.24062760264154,
+            "ttft_p95_ms": 122.24062760264154
+          },
+          "status": "completed",
+          "trace_sha256": "c49314cbe4e4b07a9335b57e861ec5b6541ea4cc061ca9abf923453276c419fd"
+        },
+        {
+          "config": {
+            "mns": 32,
+            "name": "tp2_mns32",
+            "tp": 2
+          },
+          "elapsed_seconds": 9.647802591323853,
+          "offered_request_rate": 16.0,
+          "offered_request_rate_per_gpu": 8.0,
+          "request_count": 64,
+          "request_metrics_sha256": "38c1b13c1f5fbdabd7f79b82ccb1fb37fce6840bcf03614c988d7e4882814506",
+          "score": {
+            "feasible": true,
+            "pass_rate": 1.0,
+            "passed": 64,
+            "throughput_requests_per_second": 13.174042777348715,
+            "ttft_max_ms": 1048.2973400621365,
+            "ttft_p50_ms": 643.9195307403329,
+            "ttft_p95_ms": 962.8278877316862
+          },
+          "status": "completed",
+          "trace_sha256": "2504b6340f764e726110e48251f8017a2f9ba67b44b8af7d3a838bf674823d33"
+        },
+        {
+          "config": {
+            "mns": 32,
+            "name": "tp2_mns32",
+            "tp": 2
+          },
+          "elapsed_seconds": 9.694743156433105,
+          "offered_request_rate": 32.0,
+          "offered_request_rate_per_gpu": 16.0,
+          "request_count": 64,
+          "request_metrics_sha256": "cb3b7b5a2fcf63cd01304b3b530bb21c0a9bb1c8f5e06d489e3863afd0b4d2f3",
+          "score": {
+            "feasible": false,
+            "pass_rate": 0.390625,
+            "passed": 25,
+            "throughput_requests_per_second": 13.539443740959392,
+            "ttft_max_ms": 2851.9297930156335,
+            "ttft_p50_ms": 1510.0519836938302,
+            "ttft_p95_ms": 2758.1797930156335
+          },
+          "status": "completed",
+          "trace_sha256": "5508a3ce590f7d1b0b3c211a2d4f2131886a7ae2d85cafa0c436a8e3029675f1"
+        },
+        {
+          "config": {
+            "mns": 32,
+            "name": "tp2_mns32",
+            "tp": 2
+          },
+          "elapsed_seconds": 9.594672203063965,
+          "offered_request_rate": 64.0,
+          "offered_request_rate_per_gpu": 32.0,
+          "request_count": 64,
+          "request_metrics_sha256": "a4474c98a68171ada0fb007c36fcf69e9e2dc13f06c74889591e23f93418bd5a",
+          "score": {
+            "feasible": false,
+            "pass_rate": 0.265625,
+            "passed": 17,
+            "throughput_requests_per_second": 13.539443740959392,
+            "ttft_max_ms": 3773.8047930156335,
+            "ttft_p50_ms": 1995.3684369244447,
+            "ttft_p95_ms": 3622.636520081023
+          },
+          "status": "completed",
+          "trace_sha256": "8987c3ebd4a7db983dc0e99e44710cd44a9d5d21cdb8e6b58878b160918593d9"
+        }
+      ]
+    },
+    {
+      "config": {
+        "mns": 64,
+        "name": "tp2_mns64",
+        "tp": 2
+      },
+      "loads": [
+        {
+          "config": {
+            "mns": 64,
+            "name": "tp2_mns64",
+            "tp": 2
+          },
+          "elapsed_seconds": 9.398636817932129,
+          "offered_request_rate": 4.0,
+          "offered_request_rate_per_gpu": 2.0,
+          "request_count": 64,
+          "request_metrics_sha256": "e5f6527532682ee5aca3f423367d14eba063f30e63f46d2ba0ed602db4f6eb17",
+          "score": {
+            "feasible": true,
+            "pass_rate": 1.0,
+            "passed": 64,
+            "throughput_requests_per_second": 4.032196934357252,
+            "ttft_max_ms": 122.24062760264154,
+            "ttft_p50_ms": 122.24062760264154,
+            "ttft_p95_ms": 122.24062760264154
+          },
+          "status": "completed",
+          "trace_sha256": "0cef631f351f5afd880172d0467d62d9813de5ba36fd3795f91afe3256450c05"
+        },
+        {
+          "config": {
+            "mns": 64,
+            "name": "tp2_mns64",
+            "tp": 2
+          },
+          "elapsed_seconds": 9.495105981826782,
+          "offered_request_rate": 8.0,
+          "offered_request_rate_per_gpu": 4.0,
+          "request_count": 64,
+          "request_metrics_sha256": "8dd3bedcbf76d30bf4285e10b886f95bef9ae8e5d41f4bb50a7650b322b8eb83",
+          "score": {
+            "feasible": true,
+            "pass_rate": 1.0,
+            "passed": 64,
+            "throughput_requests_per_second": 8.00276032449276,
+            "ttft_max_ms": 122.24062760264154,
+            "ttft_p50_ms": 122.24062760264154,
+            "ttft_p95_ms": 122.24062760264154
+          },
+          "status": "completed",
+          "trace_sha256": "c49314cbe4e4b07a9335b57e861ec5b6541ea4cc061ca9abf923453276c419fd"
+        },
+        {
+          "config": {
+            "mns": 64,
+            "name": "tp2_mns64",
+            "tp": 2
+          },
+          "elapsed_seconds": 9.54387378692627,
+          "offered_request_rate": 16.0,
+          "offered_request_rate_per_gpu": 8.0,
+          "request_count": 64,
+          "request_metrics_sha256": "38c1b13c1f5fbdabd7f79b82ccb1fb37fce6840bcf03614c988d7e4882814506",
+          "score": {
+            "feasible": true,
+            "pass_rate": 1.0,
+            "passed": 64,
+            "throughput_requests_per_second": 13.174042777348715,
+            "ttft_max_ms": 1048.2973400621365,
+            "ttft_p50_ms": 643.9195307403329,
+            "ttft_p95_ms": 962.8278877316862
+          },
+          "status": "completed",
+          "trace_sha256": "2504b6340f764e726110e48251f8017a2f9ba67b44b8af7d3a838bf674823d33"
+        },
+        {
+          "config": {
+            "mns": 64,
+            "name": "tp2_mns64",
+            "tp": 2
+          },
+          "elapsed_seconds": 9.192227840423584,
+          "offered_request_rate": 32.0,
+          "offered_request_rate_per_gpu": 16.0,
+          "request_count": 64,
+          "request_metrics_sha256": "cb3b7b5a2fcf63cd01304b3b530bb21c0a9bb1c8f5e06d489e3863afd0b4d2f3",
+          "score": {
+            "feasible": false,
+            "pass_rate": 0.390625,
+            "passed": 25,
+            "throughput_requests_per_second": 13.539443740959392,
+            "ttft_max_ms": 2851.9297930156335,
+            "ttft_p50_ms": 1510.0519836938302,
+            "ttft_p95_ms": 2758.1797930156335
+          },
+          "status": "completed",
+          "trace_sha256": "5508a3ce590f7d1b0b3c211a2d4f2131886a7ae2d85cafa0c436a8e3029675f1"
+        },
+        {
+          "config": {
+            "mns": 64,
+            "name": "tp2_mns64",
+            "tp": 2
+          },
+          "elapsed_seconds": 9.294084072113037,
+          "offered_request_rate": 64.0,
+          "offered_request_rate_per_gpu": 32.0,
+          "request_count": 64,
+          "request_metrics_sha256": "a4474c98a68171ada0fb007c36fcf69e9e2dc13f06c74889591e23f93418bd5a",
+          "score": {
+            "feasible": false,
+            "pass_rate": 0.265625,
+            "passed": 17,
+            "throughput_requests_per_second": 13.539443740959392,
+            "ttft_max_ms": 3773.8047930156335,
+            "ttft_p50_ms": 1995.3684369244447,
+            "ttft_p95_ms": 3622.636520081023
+          },
+          "status": "completed",
+          "trace_sha256": "8987c3ebd4a7db983dc0e99e44710cd44a9d5d21cdb8e6b58878b160918593d9"
+        }
+      ]
+    }
+  ],
+  "contract": {
+    "arrival": "open_loop_uniform",
+    "input_tokens": 2048,
+    "output_tokens": 1,
+    "prefix_caching": false,
+    "rates": [
+      4.0,
+      8.0,
+      16.0,
+      32.0,
+      64.0
+    ],
+    "requests_per_anchor": 64,
+    "target_pass_rate": 0.95,
+    "ttft_slo_ms": 1256.0
+  },
+  "frontier": {
+    "git_head": "d9cfeb6d8791fbf2f295dd9744c56a666171776e",
+    "git_status_short": " M frontier/config/config.py\n M frontier/entities/request.py\n M frontier/events/cluster_schedule_event.py\n M frontier/execution_time_predictor/sklearn_execution_time_predictor.py\n M frontier/metrics/constants.py\n M frontier/metrics/metrics_store.py\n M frontier/profiling/common/layers/rotary_embedding.py\n M frontier/profiling/moe/moe_impl.py\n M frontier/profiling/moe/moe_vllm_kernel.py\n M frontier/scheduler/cluster_scheduler/__init__.py\n M frontier/scheduler/cluster_scheduler/base_cluster_scheduler.py\n M frontier/scheduler/cluster_scheduler/cluster_scheduler_registry.py\n M frontier/scheduler/cluster_scheduler/sticky_lor_cluster_scheduler.py\n M frontier/scheduler/replica_scheduler/base_replica_scheduler.py\n M frontier/scheduler/replica_scheduler/vllm_v1_engine_replica_scheduler.py\n M frontier/scheduler/replica_stage_scheduler/replica_stage_schduler.py\n M frontier/simulator.py\n M frontier/types/cluster_scheduler_type.py\n?? data/profiling/compute/h20/\n?? frontier/scheduler/cluster_scheduler/prefix_lor_cluster_scheduler.py\n?? runs/\n?? tests/unit/test_attn_prefill_prediction_fallback.py\n",
+    "source": "/tmp/replayserve-frontier-rs1b"
+  },
+  "profiles": {
+    "coverage": {
+      "attention": {
+        "1": {
+          "exact_prefill_2048_rows": 1,
+          "profile_batch_size": 1
+        },
+        "2": {
+          "exact_prefill_2048_rows": 1,
+          "profile_batch_size": 1
+        },
+        "4": {
+          "exact_prefill_2048_rows": 1,
+          "profile_batch_size": 1
+        }
+      },
+      "manifest": {
+        "attention_tp_coverage": [
+          1,
+          2,
+          4
+        ],
+        "environment_contract": {
+          "dtype": "bfloat16",
+          "frontier_commit": "d9cfeb6d8791fbf2f295dd9744c56a666171776e",
+          "hardware": "NVIDIA H20",
+          "model": "Qwen3-30B-A3B",
+          "tensor_parallel_sizes": [
+            1,
+            2,
+            4
+          ],
+          "vllm_source_commit": "88d34c6409e9fb3c7b8ca0c04756f061d2099eb1",
+          "vllm_version": "0.20.0"
+        },
+        "inputs": {
+          "/home/gahow/phd/aituner/runs/frontier-qwen30-vllm020-profile-v1/fleet-artifacts/qwen30-vllm020-allreduce-full-tp2-20260716-v1-dispatch-aware-20260716T140743025781Z/artifacts/artifacts/allreduce-full-tp2-v1/raw/allreduce-tp2.json": "97c3c76b5a04e95bd9192423c2b891667c668f39cc0dfecbd097d749939f2d0a",
+          "/home/gahow/phd/aituner/runs/frontier-qwen30-vllm020-profile-v1/fleet-artifacts/qwen30-vllm020-allreduce-full-tp4-20260716-v1-dispatch-aware-20260716T141106009788Z/artifacts/artifacts/allreduce-full-tp4-v1/raw/allreduce-tp4.json": "809df9baa6f468cf12bf0c99827475acc67894dd9f3f948976590b665fac0e76",
+          "/home/gahow/phd/aituner/runs/frontier-qwen30-vllm020-profile-v1/fleet-artifacts/qwen30-vllm020-flashattn-kv-full-tp1-20260716-v2-20260716T135132587012Z/artifacts/artifacts/flashattn-kv-full-v2-tp1/raw/flashattn-tp1.json": "dcb4c1bf7e76b9c765f78ddd2b8a734f2d7ba2adac13ce017689a8a77fe69a27",
+          "/home/gahow/phd/aituner/runs/frontier-qwen30-vllm020-profile-v1/fleet-artifacts/qwen30-vllm020-flashattn-kv-full-tp2-20260716-v2-20260716T135134194295Z/artifacts/artifacts/flashattn-kv-full-v2-tp2/raw/flashattn-tp2.json": "43ce042556ba887c8860614b43ccf0f564e5cebc1a0cffbce299d0acb9fa8d07",
+          "/home/gahow/phd/aituner/runs/frontier-qwen30-vllm020-profile-v1/fleet-artifacts/qwen30-vllm020-flashattn-kv-full-tp4-20260716-v2-20260716T135135197200Z/artifacts/artifacts/flashattn-kv-full-v2-tp4/raw/flashattn-tp4.json": "84eef31bcad0f556907a093318a420959d14fdc94474823d11f659704bdfec73",
+          "/home/gahow/phd/aituner/runs/frontier-qwen30-vllm020-profile-v1/fleet-artifacts/qwen30-vllm020-frontier-linear-full-20260716-v2-max-tokens-20260716T144444676943Z/artifacts/artifacts/frontier-linear-full-v2/profiles/compute/h20/qwen3-a3b-30b-moe/linear_op.csv": "67666cb0a4901b74599d468df2e31bcaa2a11a7842cc0cefba24ffce62508e0c",
+          "/home/gahow/phd/aituner/runs/frontier-qwen30-vllm020-profile-v1/fleet-artifacts/qwen30-vllm020-moe-full-20260716-v1-local-shard-20260716T141334565164Z/artifacts/artifacts/moe-full-v1/raw/moe-full.json": "588f6ad0d69c9636d1b852e3df0a12d13cfe731f050ea7ec7aea457cceefbde8",
+          "/home/gahow/phd/aituner/runs/frontier-qwen30-vllm020-profile-v1/fleet-artifacts/qwen30-vllm020-router-full-20260716-v3-tp-context-20260716T145446098505Z/artifacts/artifacts/router-full-v3/raw/router.json": "1962972e983bff3e06a721ef4ae4ec65728ff669681497a4a7e7f769b88b4931"
+        },
+        "outputs": {
+          "allreduce.json": "b38d14f990578d668523d25b107aceed433da5020d8ada3b6e44d3562261a3b3",
+          "attention.csv": "76dcb767cebb4ec1c4e24bd04d93ddd48b5d271986ebfb51a197ab33e1b3d87d",
+          "attention_true_mixed_fused.csv": "43ef4be90bddc9aeac6dbbe339feec24162cd1f2129a08fbd959e6ee4eaf5f60",
+          "linear_op.csv": "67666cb0a4901b74599d468df2e31bcaa2a11a7842cc0cefba24ffce62508e0c",
+          "moe.csv": "0e4dcba72918a1c4cf4e96ced31ee3829248a19ad54553cebef14417725808b0"
+        },
+        "profile_id": "qwen3-30b-a3b-bf16-vllm020-h20-tp1-2-4-fused-mixed-total-conserving",
+        "projection_contract": {
+          "allreduce": "Frozen exact runtime measurements; base profile-only comparison keeps the historical Frontier CC backend fixed to isolate compute profile fidelity",
+          "attention": "Pure prefill/extend/decode FA3 core plus separately measured KV update; input/output reshape assumed zero; exported mean is used as median target; true mixed rows use a total-conserving compatibility projection",
+          "attention_true_mixed": "The directly measured fused total is preserved in diagnostics. Frontier's two targets are projected by the same-TP pure prefill/decode reference ratio, with projected prefill + decode exactly equal to the fused total; the split is a schema compatibility attribution, not an observation",
+          "linear": "Frontier profiler using vLLM 0.20 CUDA operators",
+          "moe": "Replicated gate and fused top-k plus TP-local modular expert kernel; expert measurement already includes prepare/finalize so shuffling is zero"
+        },
+        "row_counts": {
+          "allreduce": 24,
+          "attention_frontier_compatible": 132,
+          "attention_true_mixed_fused_diagnostic": 30,
+          "linear": 36,
+          "moe": 72
+        },
+        "schema_version": "frontier_qwen30_vllm020_frozen_profile.v2"
+      }
+    },
+    "root": "/home/gahow/phd/aituner/runs/frontier-qwen30-vllm020-profile-v1/frozen/profile-v2",
+    "sha256": {
+      "attention": "76dcb767cebb4ec1c4e24bd04d93ddd48b5d271986ebfb51a197ab33e1b3d87d",
+      "linear": "67666cb0a4901b74599d468df2e31bcaa2a11a7842cc0cefba24ffce62508e0c",
+      "manifest": "af40545e75aff55c6333cd2d5379ccf042a5a0b7d7fc7df4f745ce256cb290eb",
+      "moe": "0e4dcba72918a1c4cf4e96ced31ee3829248a19ad54553cebef14417725808b0"
+    }
+  },
+  "schema": "frontier-qwen30-prefill-surface-v1",
+  "status": "partial_not_decision_bearing"
+}
diff --git a/runs/frontier-phase-factorial-v0/simulator-tp4/frontier_surface_frozen.json b/runs/frontier-phase-factorial-v0/simulator-tp4/frontier_surface_frozen.json
new file mode 100644
index 0000000..82a1227
--- /dev/null
+++ b/runs/frontier-phase-factorial-v0/simulator-tp4/frontier_surface_frozen.json
@@ -0,0 +1,647 @@
+{
+  "capacity": [
+    {
+      "config": {
+        "mns": 16,
+        "name": "tp4_mns16",
+        "tp": 4
+      },
+      "lower_censored": false,
+      "maximum_tested_feasible_request_rate": 16.0,
+      "maximum_tested_feasible_request_rate_per_gpu": 4.0,
+      "upper_censored": false
+    },
+    {
+      "config": {
+        "mns": 32,
+        "name": "tp4_mns32",
+        "tp": 4
+      },
+      "lower_censored": false,
+      "maximum_tested_feasible_request_rate": 16.0,
+      "maximum_tested_feasible_request_rate_per_gpu": 4.0,
+      "upper_censored": false
+    },
+    {
+      "config": {
+        "mns": 64,
+        "name": "tp4_mns64",
+        "tp": 4
+      },
+      "lower_censored": false,
+      "maximum_tested_feasible_request_rate": 16.0,
+      "maximum_tested_feasible_request_rate_per_gpu": 4.0,
+      "upper_censored": false
+    },
+    {
+      "config": {
+        "mns": 8,
+        "name": "tp4_mns8",
+        "tp": 4
+      },
+      "lower_censored": false,
+      "maximum_tested_feasible_request_rate": 16.0,
+      "maximum_tested_feasible_request_rate_per_gpu": 4.0,
+      "upper_censored": false
+    }
+  ],
+  "config_results": [
+    {
+      "config": {
+        "mns": 8,
+        "name": "tp4_mns8",
+        "tp": 4
+      },
+      "loads": [
+        {
+          "config": {
+            "mns": 8,
+            "name": "tp4_mns8",
+            "tp": 4
+          },
+          "elapsed_seconds": 244.95408463478088,
+          "offered_request_rate": 4.0,
+          "offered_request_rate_per_gpu": 1.0,
+          "request_count": 64,
+          "request_metrics_sha256": "c0c3c34bea43dae2477b10152b07632e5d03968f47cdfd70ad5b676e2502af4c",
+          "score": {
+            "feasible": true,
+            "pass_rate": 1.0,
+            "passed": 64,
+            "throughput_requests_per_second": 4.039607508408214,
+            "ttft_max_ms": 93.12333977685495,
+            "ttft_p50_ms": 93.12333977685405,
+            "ttft_p95_ms": 93.12333977685495
+          },
+          "status": "completed",
+          "trace_sha256": "0cef631f351f5afd880172d0467d62d9813de5ba36fd3795f91afe3256450c05"
+        },
+        {
+          "config": {
+            "mns": 8,
+            "name": "tp4_mns8",
+            "tp": 4
+          },
+          "elapsed_seconds": 9.85124945640564,
+          "offered_request_rate": 8.0,
+          "offered_request_rate_per_gpu": 2.0,
+          "request_count": 64,
+          "request_metrics_sha256": "b5c6ba0cd3b2a81a12fb93b2cad6f80341f9352c110385046490f59d7933a5d5",
+          "score": {
+            "feasible": true,
+            "pass_rate": 1.0,
+            "passed": 64,
+            "throughput_requests_per_second": 8.032004183533672,
+            "ttft_max_ms": 93.12333977685495,
+            "ttft_p50_ms": 93.12333977685495,
+            "ttft_p95_ms": 93.12333977685495
+          },
+          "status": "completed",
+          "trace_sha256": "c49314cbe4e4b07a9335b57e861ec5b6541ea4cc061ca9abf923453276c419fd"
+        },
+        {
+          "config": {
+            "mns": 8,
+            "name": "tp4_mns8",
+            "tp": 4
+          },
+          "elapsed_seconds": 9.70260500907898,
+          "offered_request_rate": 16.0,
+          "offered_request_rate_per_gpu": 4.0,
+          "request_count": 64,
+          "request_metrics_sha256": "c40f10c895f2fc45671f7ff7c6c08f30c20854f0ff4fe3e6719934a7b08b0112",
+          "score": {
+            "feasible": true,
+            "pass_rate": 1.0,
+            "passed": 64,
+            "throughput_requests_per_second": 15.412041282697286,
+            "ttft_max_ms": 314.4210260695206,
+            "ttft_p50_ms": 239.37774488651664,
+            "ttft_p95_ms": 309.403713596319
+          },
+          "status": "completed",
+          "trace_sha256": "2504b6340f764e726110e48251f8017a2f9ba67b44b8af7d3a838bf674823d33"
+        },
+        {
+          "config": {
+            "mns": 8,
+            "name": "tp4_mns8",
+            "tp": 4
+          },
+          "elapsed_seconds": 9.855336904525757,
+          "offered_request_rate": 32.0,
+          "offered_request_rate_per_gpu": 8.0,
+          "request_count": 64,
+          "request_metrics_sha256": "9f77ee40d102bdf14d213132ef10b4a7f4175908ffee411ef30ad3a5f1e4b16f",
+          "score": {
+            "feasible": false,
+            "pass_rate": 0.625,
+            "passed": 40,
+            "throughput_requests_per_second": 16.898601280793688,
+            "ttft_max_ms": 1850.42236218991,
+            "ttft_p50_ms": 1017.6783445024347,
+            "ttft_p95_ms": 1787.92236218991
+          },
+          "status": "completed",
+          "trace_sha256": "5508a3ce590f7d1b0b3c211a2d4f2131886a7ae2d85cafa0c436a8e3029675f1"
+        },
+        {
+          "config": {
+            "mns": 8,
+            "name": "tp4_mns8",
+            "tp": 4
+          },
+          "elapsed_seconds": 9.947782754898071,
+          "offered_request_rate": 64.0,
+          "offered_request_rate_per_gpu": 16.0,
+          "request_count": 64,
+          "request_metrics_sha256": "efb43cc32f2f7129082e47d89239f08b10c6e7e6dd4a4f0b299d2364cb781af0",
+          "score": {
+            "feasible": false,
+            "pass_rate": 0.390625,
+            "passed": 25,
+            "throughput_requests_per_second": 17.184897883039092,
+            "ttft_max_ms": 2771.0751922610093,
+            "ttft_p50_ms": 1468.937931419684,
+            "ttft_p95_ms": 2661.056949107159
+          },
+          "status": "completed",
+          "trace_sha256": "8987c3ebd4a7db983dc0e99e44710cd44a9d5d21cdb8e6b58878b160918593d9"
+        }
+      ]
+    },
+    {
+      "config": {
+        "mns": 16,
+        "name": "tp4_mns16",
+        "tp": 4
+      },
+      "loads": [
+        {
+          "config": {
+            "mns": 16,
+            "name": "tp4_mns16",
+            "tp": 4
+          },
+          "elapsed_seconds": 9.648890733718872,
+          "offered_request_rate": 4.0,
+          "offered_request_rate_per_gpu": 1.0,
+          "request_count": 64,
+          "request_metrics_sha256": "c0c3c34bea43dae2477b10152b07632e5d03968f47cdfd70ad5b676e2502af4c",
+          "score": {
+            "feasible": true,
+            "pass_rate": 1.0,
+            "passed": 64,
+            "throughput_requests_per_second": 4.039607508408214,
+            "ttft_max_ms": 93.12333977685495,
+            "ttft_p50_ms": 93.12333977685405,
+            "ttft_p95_ms": 93.12333977685495
+          },
+          "status": "completed",
+          "trace_sha256": "0cef631f351f5afd880172d0467d62d9813de5ba36fd3795f91afe3256450c05"
+        },
+        {
+          "config": {
+            "mns": 16,
+            "name": "tp4_mns16",
+            "tp": 4
+          },
+          "elapsed_seconds": 9.353798151016235,
+          "offered_request_rate": 8.0,
+          "offered_request_rate_per_gpu": 2.0,
+          "request_count": 64,
+          "request_metrics_sha256": "b5c6ba0cd3b2a81a12fb93b2cad6f80341f9352c110385046490f59d7933a5d5",
+          "score": {
+            "feasible": true,
+            "pass_rate": 1.0,
+            "passed": 64,
+            "throughput_requests_per_second": 8.032004183533672,
+            "ttft_max_ms": 93.12333977685495,
+            "ttft_p50_ms": 93.12333977685495,
+            "ttft_p95_ms": 93.12333977685495
+          },
+          "status": "completed",
+          "trace_sha256": "c49314cbe4e4b07a9335b57e861ec5b6541ea4cc061ca9abf923453276c419fd"
+        },
+        {
+          "config": {
+            "mns": 16,
+            "name": "tp4_mns16",
+            "tp": 4
+          },
+          "elapsed_seconds": 9.39468502998352,
+          "offered_request_rate": 16.0,
+          "offered_request_rate_per_gpu": 4.0,
+          "request_count": 64,
+          "request_metrics_sha256": "c40f10c895f2fc45671f7ff7c6c08f30c20854f0ff4fe3e6719934a7b08b0112",
+          "score": {
+            "feasible": true,
+            "pass_rate": 1.0,
+            "passed": 64,
+            "throughput_requests_per_second": 15.412041282697286,
+            "ttft_max_ms": 314.4210260695206,
+            "ttft_p50_ms": 239.37774488651664,
+            "ttft_p95_ms": 309.403713596319
+          },
+          "status": "completed",
+          "trace_sha256": "2504b6340f764e726110e48251f8017a2f9ba67b44b8af7d3a838bf674823d33"
+        },
+        {
+          "config": {
+            "mns": 16,
+            "name": "tp4_mns16",
+            "tp": 4
+          },
+          "elapsed_seconds": 9.29756498336792,
+          "offered_request_rate": 32.0,
+          "offered_request_rate_per_gpu": 8.0,
+          "request_count": 64,
+          "request_metrics_sha256": "9f77ee40d102bdf14d213132ef10b4a7f4175908ffee411ef30ad3a5f1e4b16f",
+          "score": {
+            "feasible": false,
+            "pass_rate": 0.625,
+            "passed": 40,
+            "throughput_requests_per_second": 16.898601280793688,
+            "ttft_max_ms": 1850.42236218991,
+            "ttft_p50_ms": 1017.6783445024347,
+            "ttft_p95_ms": 1787.92236218991
+          },
+          "status": "completed",
+          "trace_sha256": "5508a3ce590f7d1b0b3c211a2d4f2131886a7ae2d85cafa0c436a8e3029675f1"
+        },
+        {
+          "config": {
+            "mns": 16,
+            "name": "tp4_mns16",
+            "tp": 4
+          },
+          "elapsed_seconds": 9.34343934059143,
+          "offered_request_rate": 64.0,
+          "offered_request_rate_per_gpu": 16.0,
+          "request_count": 64,
+          "request_metrics_sha256": "efb43cc32f2f7129082e47d89239f08b10c6e7e6dd4a4f0b299d2364cb781af0",
+          "score": {
+            "feasible": false,
+            "pass_rate": 0.390625,
+            "passed": 25,
+            "throughput_requests_per_second": 17.184897883039092,
+            "ttft_max_ms": 2771.0751922610093,
+            "ttft_p50_ms": 1468.937931419684,
+            "ttft_p95_ms": 2661.056949107159
+          },
+          "status": "completed",
+          "trace_sha256": "8987c3ebd4a7db983dc0e99e44710cd44a9d5d21cdb8e6b58878b160918593d9"
+        }
+      ]
+    },
+    {
+      "config": {
+        "mns": 32,
+        "name": "tp4_mns32",
+        "tp": 4
+      },
+      "loads": [
+        {
+          "config": {
+            "mns": 32,
+            "name": "tp4_mns32",
+            "tp": 4
+          },
+          "elapsed_seconds": 9.296925783157349,
+          "offered_request_rate": 4.0,
+          "offered_request_rate_per_gpu": 1.0,
+          "request_count": 64,
+          "request_metrics_sha256": "c0c3c34bea43dae2477b10152b07632e5d03968f47cdfd70ad5b676e2502af4c",
+          "score": {
+            "feasible": true,
+            "pass_rate": 1.0,
+            "passed": 64,
+            "throughput_requests_per_second": 4.039607508408214,
+            "ttft_max_ms": 93.12333977685495,
+            "ttft_p50_ms": 93.12333977685405,
+            "ttft_p95_ms": 93.12333977685495
+          },
+          "status": "completed",
+          "trace_sha256": "0cef631f351f5afd880172d0467d62d9813de5ba36fd3795f91afe3256450c05"
+        },
+        {
+          "config": {
+            "mns": 32,
+            "name": "tp4_mns32",
+            "tp": 4
+          },
+          "elapsed_seconds": 9.345365285873413,
+          "offered_request_rate": 8.0,
+          "offered_request_rate_per_gpu": 2.0,
+          "request_count": 64,
+          "request_metrics_sha256": "b5c6ba0cd3b2a81a12fb93b2cad6f80341f9352c110385046490f59d7933a5d5",
+          "score": {
+            "feasible": true,
+            "pass_rate": 1.0,
+            "passed": 64,
+            "throughput_requests_per_second": 8.032004183533672,
+            "ttft_max_ms": 93.12333977685495,
+            "ttft_p50_ms": 93.12333977685495,
+            "ttft_p95_ms": 93.12333977685495
+          },
+          "status": "completed",
+          "trace_sha256": "c49314cbe4e4b07a9335b57e861ec5b6541ea4cc061ca9abf923453276c419fd"
+        },
+        {
+          "config": {
+            "mns": 32,
+            "name": "tp4_mns32",
+            "tp": 4
+          },
+          "elapsed_seconds": 9.350512504577637,
+          "offered_request_rate": 16.0,
+          "offered_request_rate_per_gpu": 4.0,
+          "request_count": 64,
+          "request_metrics_sha256": "c40f10c895f2fc45671f7ff7c6c08f30c20854f0ff4fe3e6719934a7b08b0112",
+          "score": {
+            "feasible": true,
+            "pass_rate": 1.0,
+            "passed": 64,
+            "throughput_requests_per_second": 15.412041282697286,
+            "ttft_max_ms": 314.4210260695206,
+            "ttft_p50_ms": 239.37774488651664,
+            "ttft_p95_ms": 309.403713596319
+          },
+          "status": "completed",
+          "trace_sha256": "2504b6340f764e726110e48251f8017a2f9ba67b44b8af7d3a838bf674823d33"
+        },
+        {
+          "config": {
+            "mns": 32,
+            "name": "tp4_mns32",
+            "tp": 4
+          },
+          "elapsed_seconds": 9.300485372543335,
+          "offered_request_rate": 32.0,
+          "offered_request_rate_per_gpu": 8.0,
+          "request_count": 64,
+          "request_metrics_sha256": "9f77ee40d102bdf14d213132ef10b4a7f4175908ffee411ef30ad3a5f1e4b16f",
+          "score": {
+            "feasible": false,
+            "pass_rate": 0.625,
+            "passed": 40,
+            "throughput_requests_per_second": 16.898601280793688,
+            "ttft_max_ms": 1850.42236218991,
+            "ttft_p50_ms": 1017.6783445024347,
+            "ttft_p95_ms": 1787.92236218991
+          },
+          "status": "completed",
+          "trace_sha256": "5508a3ce590f7d1b0b3c211a2d4f2131886a7ae2d85cafa0c436a8e3029675f1"
+        },
+        {
+          "config": {
+            "mns": 32,
+            "name": "tp4_mns32",
+            "tp": 4
+          },
+          "elapsed_seconds": 9.292456150054932,
+          "offered_request_rate": 64.0,
+          "offered_request_rate_per_gpu": 16.0,
+          "request_count": 64,
+          "request_metrics_sha256": "efb43cc32f2f7129082e47d89239f08b10c6e7e6dd4a4f0b299d2364cb781af0",
+          "score": {
+            "feasible": false,
+            "pass_rate": 0.390625,
+            "passed": 25,
+            "throughput_requests_per_second": 17.184897883039092,
+            "ttft_max_ms": 2771.0751922610093,
+            "ttft_p50_ms": 1468.937931419684,
+            "ttft_p95_ms": 2661.056949107159
+          },
+          "status": "completed",
+          "trace_sha256": "8987c3ebd4a7db983dc0e99e44710cd44a9d5d21cdb8e6b58878b160918593d9"
+        }
+      ]
+    },
+    {
+      "config": {
+        "mns": 64,
+        "name": "tp4_mns64",
+        "tp": 4
+      },
+      "loads": [
+        {
+          "config": {
+            "mns": 64,
+            "name": "tp4_mns64",
+            "tp": 4
+          },
+          "elapsed_seconds": 9.294773817062378,
+          "offered_request_rate": 4.0,
+          "offered_request_rate_per_gpu": 1.0,
+          "request_count": 64,
+          "request_metrics_sha256": "c0c3c34bea43dae2477b10152b07632e5d03968f47cdfd70ad5b676e2502af4c",
+          "score": {
+            "feasible": true,
+            "pass_rate": 1.0,
+            "passed": 64,
+            "throughput_requests_per_second": 4.039607508408214,
+            "ttft_max_ms": 93.12333977685495,
+            "ttft_p50_ms": 93.12333977685405,
+            "ttft_p95_ms": 93.12333977685495
+          },
+          "status": "completed",
+          "trace_sha256": "0cef631f351f5afd880172d0467d62d9813de5ba36fd3795f91afe3256450c05"
+        },
+        {
+          "config": {
+            "mns": 64,
+            "name": "tp4_mns64",
+            "tp": 4
+          },
+          "elapsed_seconds": 9.34990382194519,
+          "offered_request_rate": 8.0,
+          "offered_request_rate_per_gpu": 2.0,
+          "request_count": 64,
+          "request_metrics_sha256": "b5c6ba0cd3b2a81a12fb93b2cad6f80341f9352c110385046490f59d7933a5d5",
+          "score": {
+            "feasible": true,
+            "pass_rate": 1.0,
+            "passed": 64,
+            "throughput_requests_per_second": 8.032004183533672,
+            "ttft_max_ms": 93.12333977685495,
+            "ttft_p50_ms": 93.12333977685495,
+            "ttft_p95_ms": 93.12333977685495
+          },
+          "status": "completed",
+          "trace_sha256": "c49314cbe4e4b07a9335b57e861ec5b6541ea4cc061ca9abf923453276c419fd"
+        },
+        {
+          "config": {
+            "mns": 64,
+            "name": "tp4_mns64",
+            "tp": 4
+          },
+          "elapsed_seconds": 9.346213579177856,
+          "offered_request_rate": 16.0,
+          "offered_request_rate_per_gpu": 4.0,
+          "request_count": 64,
+          "request_metrics_sha256": "c40f10c895f2fc45671f7ff7c6c08f30c20854f0ff4fe3e6719934a7b08b0112",
+          "score": {
+            "feasible": true,
+            "pass_rate": 1.0,
+            "passed": 64,
+            "throughput_requests_per_second": 15.412041282697286,
+            "ttft_max_ms": 314.4210260695206,
+            "ttft_p50_ms": 239.37774488651664,
+            "ttft_p95_ms": 309.403713596319
+          },
+          "status": "completed",
+          "trace_sha256": "2504b6340f764e726110e48251f8017a2f9ba67b44b8af7d3a838bf674823d33"
+        },
+        {
+          "config": {
+            "mns": 64,
+            "name": "tp4_mns64",
+            "tp": 4
+          },
+          "elapsed_seconds": 9.498731136322021,
+          "offered_request_rate": 32.0,
+          "offered_request_rate_per_gpu": 8.0,
+          "request_count": 64,
+          "request_metrics_sha256": "9f77ee40d102bdf14d213132ef10b4a7f4175908ffee411ef30ad3a5f1e4b16f",
+          "score": {
+            "feasible": false,
+            "pass_rate": 0.625,
+            "passed": 40,
+            "throughput_requests_per_second": 16.898601280793688,
+            "ttft_max_ms": 1850.42236218991,
+            "ttft_p50_ms": 1017.6783445024347,
+            "ttft_p95_ms": 1787.92236218991
+          },
+          "status": "completed",
+          "trace_sha256": "5508a3ce590f7d1b0b3c211a2d4f2131886a7ae2d85cafa0c436a8e3029675f1"
+        },
+        {
+          "config": {
+            "mns": 64,
+            "name": "tp4_mns64",
+            "tp": 4
+          },
+          "elapsed_seconds": 9.300755977630615,
+          "offered_request_rate": 64.0,
+          "offered_request_rate_per_gpu": 16.0,
+          "request_count": 64,
+          "request_metrics_sha256": "efb43cc32f2f7129082e47d89239f08b10c6e7e6dd4a4f0b299d2364cb781af0",
+          "score": {
+            "feasible": false,
+            "pass_rate": 0.390625,
+            "passed": 25,
+            "throughput_requests_per_second": 17.184897883039092,
+            "ttft_max_ms": 2771.0751922610093,
+            "ttft_p50_ms": 1468.937931419684,
+            "ttft_p95_ms": 2661.056949107159
+          },
+          "status": "completed",
+          "trace_sha256": "8987c3ebd4a7db983dc0e99e44710cd44a9d5d21cdb8e6b58878b160918593d9"
+        }
+      ]
+    }
+  ],
+  "contract": {
+    "arrival": "open_loop_uniform",
+    "input_tokens": 2048,
+    "output_tokens": 1,
+    "prefix_caching": false,
+    "rates": [
+      4.0,
+      8.0,
+      16.0,
+      32.0,
+      64.0
+    ],
+    "requests_per_anchor": 64,
+    "target_pass_rate": 0.95,
+    "ttft_slo_ms": 1256.0
+  },
+  "frontier": {
+    "git_head": "d9cfeb6d8791fbf2f295dd9744c56a666171776e",
+    "git_status_short": " M frontier/config/config.py\n M frontier/entities/request.py\n M frontier/events/cluster_schedule_event.py\n M frontier/execution_time_predictor/sklearn_execution_time_predictor.py\n M frontier/metrics/constants.py\n M frontier/metrics/metrics_store.py\n M frontier/profiling/common/layers/rotary_embedding.py\n M frontier/profiling/moe/moe_impl.py\n M frontier/profiling/moe/moe_vllm_kernel.py\n M frontier/scheduler/cluster_scheduler/__init__.py\n M frontier/scheduler/cluster_scheduler/base_cluster_scheduler.py\n M frontier/scheduler/cluster_scheduler/cluster_scheduler_registry.py\n M frontier/scheduler/cluster_scheduler/sticky_lor_cluster_scheduler.py\n M frontier/scheduler/replica_scheduler/base_replica_scheduler.py\n M frontier/scheduler/replica_scheduler/vllm_v1_engine_replica_scheduler.py\n M frontier/scheduler/replica_stage_scheduler/replica_stage_schduler.py\n M frontier/simulator.py\n M frontier/types/cluster_scheduler_type.py\n?? data/profiling/compute/h20/\n?? frontier/scheduler/cluster_scheduler/prefix_lor_cluster_scheduler.py\n?? runs/\n?? tests/unit/test_attn_prefill_prediction_fallback.py\n",
+    "source": "/tmp/replayserve-frontier-rs1b"
+  },
+  "profiles": {
+    "coverage": {
+      "attention": {
+        "1": {
+          "exact_prefill_2048_rows": 1,
+          "profile_batch_size": 1
+        },
+        "2": {
+          "exact_prefill_2048_rows": 1,
+          "profile_batch_size": 1
+        },
+        "4": {
+          "exact_prefill_2048_rows": 1,
+          "profile_batch_size": 1
+        }
+      },
+      "manifest": {
+        "attention_tp_coverage": [
+          1,
+          2,
+          4
+        ],
+        "environment_contract": {
+          "dtype": "bfloat16",
+          "frontier_commit": "d9cfeb6d8791fbf2f295dd9744c56a666171776e",
+          "hardware": "NVIDIA H20",
+          "model": "Qwen3-30B-A3B",
+          "tensor_parallel_sizes": [
+            1,
+            2,
+            4
+          ],
+          "vllm_source_commit": "88d34c6409e9fb3c7b8ca0c04756f061d2099eb1",
+          "vllm_version": "0.20.0"
+        },
+        "inputs": {
+          "/home/gahow/phd/aituner/runs/frontier-qwen30-vllm020-profile-v1/fleet-artifacts/qwen30-vllm020-allreduce-full-tp2-20260716-v1-dispatch-aware-20260716T140743025781Z/artifacts/artifacts/allreduce-full-tp2-v1/raw/allreduce-tp2.json": "97c3c76b5a04e95bd9192423c2b891667c668f39cc0dfecbd097d749939f2d0a",
+          "/home/gahow/phd/aituner/runs/frontier-qwen30-vllm020-profile-v1/fleet-artifacts/qwen30-vllm020-allreduce-full-tp4-20260716-v1-dispatch-aware-20260716T141106009788Z/artifacts/artifacts/allreduce-full-tp4-v1/raw/allreduce-tp4.json": "809df9baa6f468cf12bf0c99827475acc67894dd9f3f948976590b665fac0e76",
+          "/home/gahow/phd/aituner/runs/frontier-qwen30-vllm020-profile-v1/fleet-artifacts/qwen30-vllm020-flashattn-kv-full-tp1-20260716-v2-20260716T135132587012Z/artifacts/artifacts/flashattn-kv-full-v2-tp1/raw/flashattn-tp1.json": "dcb4c1bf7e76b9c765f78ddd2b8a734f2d7ba2adac13ce017689a8a77fe69a27",
+          "/home/gahow/phd/aituner/runs/frontier-qwen30-vllm020-profile-v1/fleet-artifacts/qwen30-vllm020-flashattn-kv-full-tp2-20260716-v2-20260716T135134194295Z/artifacts/artifacts/flashattn-kv-full-v2-tp2/raw/flashattn-tp2.json": "43ce042556ba887c8860614b43ccf0f564e5cebc1a0cffbce299d0acb9fa8d07",
+          "/home/gahow/phd/aituner/runs/frontier-qwen30-vllm020-profile-v1/fleet-artifacts/qwen30-vllm020-flashattn-kv-full-tp4-20260716-v2-20260716T135135197200Z/artifacts/artifacts/flashattn-kv-full-v2-tp4/raw/flashattn-tp4.json": "84eef31bcad0f556907a093318a420959d14fdc94474823d11f659704bdfec73",
+          "/home/gahow/phd/aituner/runs/frontier-qwen30-vllm020-profile-v1/fleet-artifacts/qwen30-vllm020-frontier-linear-full-20260716-v2-max-tokens-20260716T144444676943Z/artifacts/artifacts/frontier-linear-full-v2/profiles/compute/h20/qwen3-a3b-30b-moe/linear_op.csv": "67666cb0a4901b74599d468df2e31bcaa2a11a7842cc0cefba24ffce62508e0c",
+          "/home/gahow/phd/aituner/runs/frontier-qwen30-vllm020-profile-v1/fleet-artifacts/qwen30-vllm020-moe-full-20260716-v1-local-shard-20260716T141334565164Z/artifacts/artifacts/moe-full-v1/raw/moe-full.json": "588f6ad0d69c9636d1b852e3df0a12d13cfe731f050ea7ec7aea457cceefbde8",
+          "/home/gahow/phd/aituner/runs/frontier-qwen30-vllm020-profile-v1/fleet-artifacts/qwen30-vllm020-router-full-20260716-v3-tp-context-20260716T145446098505Z/artifacts/artifacts/router-full-v3/raw/router.json": "1962972e983bff3e06a721ef4ae4ec65728ff669681497a4a7e7f769b88b4931"
+        },
+        "outputs": {
+          "allreduce.json": "b38d14f990578d668523d25b107aceed433da5020d8ada3b6e44d3562261a3b3",
+          "attention.csv": "76dcb767cebb4ec1c4e24bd04d93ddd48b5d271986ebfb51a197ab33e1b3d87d",
+          "attention_true_mixed_fused.csv": "43ef4be90bddc9aeac6dbbe339feec24162cd1f2129a08fbd959e6ee4eaf5f60",
+          "linear_op.csv": "67666cb0a4901b74599d468df2e31bcaa2a11a7842cc0cefba24ffce62508e0c",
+          "moe.csv": "0e4dcba72918a1c4cf4e96ced31ee3829248a19ad54553cebef14417725808b0"
+        },
+        "profile_id": "qwen3-30b-a3b-bf16-vllm020-h20-tp1-2-4-fused-mixed-total-conserving",
+        "projection_contract": {
+          "allreduce": "Frozen exact runtime measurements; base profile-only comparison keeps the historical Frontier CC backend fixed to isolate compute profile fidelity",
+          "attention": "Pure prefill/extend/decode FA3 core plus separately measured KV update; input/output reshape assumed zero; exported mean is used as median target; true mixed rows use a total-conserving compatibility projection",
+          "attention_true_mixed": "The directly measured fused total is preserved in diagnostics. Frontier's two targets are projected by the same-TP pure prefill/decode reference ratio, with projected prefill + decode exactly equal to the fused total; the split is a schema compatibility attribution, not an observation",
+          "linear": "Frontier profiler using vLLM 0.20 CUDA operators",
+          "moe": "Replicated gate and fused top-k plus TP-local modular expert kernel; expert measurement already includes prepare/finalize so shuffling is zero"
+        },
+        "row_counts": {
+          "allreduce": 24,
+          "attention_frontier_compatible": 132,
+          "attention_true_mixed_fused_diagnostic": 30,
+          "linear": 36,
+          "moe": 72
+        },
+        "schema_version": "frontier_qwen30_vllm020_frozen_profile.v2"
+      }
+    },
+    "root": "/home/gahow/phd/aituner/runs/frontier-qwen30-vllm020-profile-v1/frozen/profile-v2",
+    "sha256": {
+      "attention": "76dcb767cebb4ec1c4e24bd04d93ddd48b5d271986ebfb51a197ab33e1b3d87d",
+      "linear": "67666cb0a4901b74599d468df2e31bcaa2a11a7842cc0cefba24ffce62508e0c",
+      "manifest": "af40545e75aff55c6333cd2d5379ccf042a5a0b7d7fc7df4f745ce256cb290eb",
+      "moe": "0e4dcba72918a1c4cf4e96ced31ee3829248a19ad54553cebef14417725808b0"
+    }
+  },
+  "schema": "frontier-qwen30-prefill-surface-v1",
+  "status": "partial_not_decision_bearing"
+}
diff --git a/runs/frontier-phase-factorial-v0/test_phase_factorial.py b/runs/frontier-phase-factorial-v0/test_phase_factorial.py
new file mode 100644
index 0000000..6e30aa5
--- /dev/null
+++ b/runs/frontier-phase-factorial-v0/test_phase_factorial.py
@@ -0,0 +1,39 @@
+#!/usr/bin/env python3
+
+from __future__ import annotations
+
+import importlib.util
+import sys
+from pathlib import Path
+
+
+ROOT = Path(__file__).parent
+
+
+def load(name: str):
+    path = ROOT / name
+    spec = importlib.util.spec_from_file_location(path.stem, path)
+    assert spec and spec.loader
+    module = importlib.util.module_from_spec(spec)
+    sys.modules[spec.name] = module
+    spec.loader.exec_module(module)
+    return module
+
+
+def test_percentile() -> None:
+    client = load("qwen30_prefill_client.py")
+    assert client.percentile([4.0, 1.0, 3.0, 2.0], 0.50) == 2.0
+    assert client.percentile([4.0, 1.0, 3.0, 2.0], 0.95) == 4.0
+    assert client.percentile([], 0.95) is None
+
+
+def test_grid_and_trace(tmp_path: Path) -> None:
+    surface = load("run_frontier_qwen30_prefill_surface.py")
+    assert len(surface.GRID) == 12
+    assert {config.tp for config in surface.GRID} == {1, 2, 4}
+    trace = tmp_path / "trace.csv"
+    surface.write_trace(trace, requests=3, rate=4.0)
+    lines = trace.read_text().splitlines()
+    assert len(lines) == 4
+    assert lines[1].split(",")[:3] == ["0.000000000000", "2048", "1"]
+    assert lines[3].split(",")[:3] == ["0.500000000000", "2048", "1"]