From e602d41c801a55f547fb3744914e4d102c0e79a4 Mon Sep 17 00:00:00 2001
From: Gahow Wang <gahow.wang@gmail.com>
Date: Sun, 19 Jul 2026 19:16:20 +0800
Subject: [PATCH] Report exact state composition verdict

---
 .../analyze_qwen235_fixed_pd_state.py         | 69 +++++++++++++++++--
 1 file changed, 65 insertions(+), 4 deletions(-)

diff --git a/runs/frontier-fidelity-envelope-v1/analyze_qwen235_fixed_pd_state.py b/runs/frontier-fidelity-envelope-v1/analyze_qwen235_fixed_pd_state.py
index 103d176..236fa4d 100644
--- a/runs/frontier-fidelity-envelope-v1/analyze_qwen235_fixed_pd_state.py
+++ b/runs/frontier-fidelity-envelope-v1/analyze_qwen235_fixed_pd_state.py
@@ -691,6 +691,16 @@ def analyze(
         if min(by_batch[config][batch]["n"] for config in CONFIGS) >= 10
     }
     observed_real_contrast = REAL_TPOT_MS[CONFIGS[1]] - REAL_TPOT_MS[CONFIGS[0]]
+    exact_iteration_contrast = None
+    if iteration_real is not None:
+        exact_iteration_contrast = (
+            iteration_real[CONFIGS[1]][
+                "decode_token_weighted_iteration_elapsed_ms"
+            ]
+            - iteration_real[CONFIGS[0]][
+                "decode_token_weighted_iteration_elapsed_ms"
+            ]
+        )
     decision_contrast = exact_contrast if exact_contrast is not None else proxy_contrast
     if decision_contrast is None:
         verdict = "STOP: real Running proxy has insufficient exact simulator support"
@@ -706,10 +716,33 @@ def analyze(
                 "to the frozen-real Running proxy, it still predicts TP8 faster."
             )
     else:
-        verdict = (
-            "Coarse active-batch state can flip Frontier's ordering, but iteration-level "
-            "batch/context telemetry is required before attributing the real gap to state."
-        )
+        if exact_contrast is not None:
+            verdict = (
+                "Exact real state composition flips Frontier to the correct TP4 winner; "
+                "the ranking error is caused by closed-loop state/composition mismatch, "
+                "not a conditional stage reversal."
+            )
+        else:
+            verdict = (
+                "Coarse active-batch state can flip Frontier's ordering, but "
+                "iteration-level batch/context telemetry is required before attributing "
+                "the real gap to state."
+            )
+    decomposition = None
+    if exact_contrast is not None:
+        decomposition = {
+            "simulator_internal_tp8_minus_tp4_ms": simulator_internal_contrast["total"],
+            "exact_state_matched_tp8_minus_tp4_ms": exact_contrast["total"],
+            "state_composition_shift_ms": (
+                exact_contrast["total"] - simulator_internal_contrast["total"]
+            ),
+            "residual_vs_frozen_real_tpot_ms": (
+                exact_contrast["total"] - observed_real_contrast
+            ),
+            "residual_vs_exact_iteration_ms": (
+                exact_contrast["total"] - exact_iteration_contrast
+            ),
+        }
     return {
         "schema": "qwen235-fixed-pd-state-diagnosis-v1",
         "status": "PASS",
@@ -752,9 +785,11 @@ def analyze(
         if iteration_real is not None
         else None,
         "same_batch_contrasts": shared_contrasts,
+        "contrast_decomposition": decomposition,
         "reference": {
             "real_tpot_ms": REAL_TPOT_MS,
             "observed_real_tp8_minus_tp4_ms": observed_real_contrast,
+            "exact_iteration_tp8_minus_tp4_ms": exact_iteration_contrast,
         },
         "verdict": verdict,
     }
@@ -776,6 +811,22 @@ def markdown(result: dict[str, Any]) -> str:
             f"| {result['proxy_matched']['configs'][config]['coverage']:.1%} |"
         )
     exact = result["exact_state_matched"]
+    iteration = result["real_iteration_state"]
+    if exact is not None:
+        lines.extend(
+            [
+                "",
+                "| Config | Real exact decode batch mean | Real token-weighted iteration ms | Matched simulator ms | Exact coverage |",
+                "|---|---:|---:|---:|---:|",
+            ]
+        )
+        for config in CONFIGS:
+            lines.append(
+                f"| {config} | {iteration[config]['decode_batch_size']['mean']:.3f} "
+                f"| {iteration[config]['decode_token_weighted_iteration_elapsed_ms']:.4f} "
+                f"| {exact['configs'][config]['components_ms']['total']:.4f} "
+                f"| {exact['configs'][config]['coverage']:.1%} |"
+            )
     contrast = (
         exact["tp8_minus_tp4_ms"]
         if exact is not None and exact["tp8_minus_tp4_ms"] is not None
@@ -824,6 +875,16 @@ def markdown(result: dict[str, Any]) -> str:
             "## Interpretation boundary",
             "",
             f"- Real observed TPOT contrast: {result['reference']['observed_real_tp8_minus_tp4_ms']:+.4f} ms/token.",
+        ]
+    )
+    if result["reference"]["exact_iteration_tp8_minus_tp4_ms"] is not None:
+        lines.append(
+            "- Real exact-iteration contrast: "
+            f"{result['reference']['exact_iteration_tp8_minus_tp4_ms']:+.4f} "
+            "ms/decode-token-weighted step."
+        )
+    lines.extend(
+        [
             f"- {result['scope']['real_proxy_limitation']}.",
             f"- {result['scope']['graph_observability']}.",
             f"- {result['scope']['op_trace_accounting']}.",