diff --git a/.github/workflows/collectivex-sweep.yml b/.github/workflows/collectivex-sweep.yml
index cb9cc5b1cb..702ad39a48 100644
--- a/.github/workflows/collectivex-sweep.yml
+++ b/.github/workflows/collectivex-sweep.yml
@@ -224,6 +224,11 @@ jobs:
             || { echo "CollectiveX allocation cleanup failed; results may still be changing" >&2; exit 1; }
           python3 experimental/CollectiveX/summarize.py \
             --results-dir experimental/CollectiveX/results >> "$GITHUB_STEP_SUMMARY" || true
+          # Second view over the same documents: latency answers how long a point took,
+          # this answers how fast it moved its payload. Both are renderers over the staged
+          # JSON and neither gates the leg, so both are `|| true`.
+          python3 experimental/CollectiveX/bandwidth.py \
+            --results-dir experimental/CollectiveX/results >> "$GITHUB_STEP_SUMMARY" || true
           shopt -s nullglob
           results=(experimental/CollectiveX/results/*.json)
           if [ "${#results[@]}" -eq 0 ]; then
diff --git a/experimental/CollectiveX/bandwidth.py b/experimental/CollectiveX/bandwidth.py
new file mode 100644
index 0000000000..5d6833d810
--- /dev/null
+++ b/experimental/CollectiveX/bandwidth.py
@@ -0,0 +1,219 @@
+#!/usr/bin/env python3
+"""Per-component EP bandwidth, computed purely as a consumer of case-attempt artifacts.
+
+Two views over the bytes + latency every record already emits — no harness or GPU change,
+and this gates nothing (see docs/methodology.md):
+
+  A. effective bandwidth -- per-GPU GB/s at each ladder point: bytes / component latency.
+  B. alpha/beta fit -- OLS of p50 latency vs bytes, latency ~= alpha + bytes/beta, which
+     separates the bandwidth term from the per-call floor that dominates small T.
+
+Bytes are LOGICAL payload (one copy per unique (token, dest-rank) pair), excluding
+protocol/padding. Reading them against a link's peak needs two adjustments: logical bytes
+include the copies a rank routes to itself, which never cross the interconnect, so wire
+traffic is (1 - routing.locality.local_rank_fraction) x these figures; and beta is a
+MARGINAL rate, so it legitimately sits above every measured point.
+
+A line through a ladder is a MODEL and a positive slope alone is not evidence, so every Fit
+carries its own quality and beta is withheld unless it clears both gates below. Deliberately
+NOT added: a per-SKU peak table to check beta against; that hardware knowledge belongs in
+platform_config.json, ages badly, and would mislead in its own right.
+"""
+from __future__ import annotations
+
+import argparse
+import os
+import sys
+from typing import NamedTuple
+
+sys.path.insert(0, os.path.dirname(os.path.abspath(__file__)))
+from summarize import load_results  # noqa: E402  (sibling consumer, stdlib-only)
+
+# stage moves nothing (bytes always 0); isolated_sum is a derived percentile sum, not a real
+# chained rate. Only these three carry a measured latency + bytes.
+COMPONENTS = ("dispatch", "combine", "roundtrip")
+
+# Below this the line is not describing the ladder: two rungs of noise once yielded 1018 GB/s
+# per GPU on a B200 -- above the link's per-direction peak -- at R2 = 0.29, printed
+# indistinguishably from an R2 = 0.99 fit. A judgement call; good fits observed at 0.99+.
+FIT_MIN_R2 = 0.9
+# R2 alone is NOT enough -- a latency-bound dispatch ladder came back at 3763 GB/s with
+# R2 = 0.92, because alpha explained nearly everything and the slope was not identifiable.
+# So also require the transfer term to be a real share of the top rung's latency; below this
+# the honest reading is "not bandwidth-limited here", not "very fast".
+FIT_MIN_BANDWIDTH_SHARE = 0.25
+# If the smallest rung is not within this fraction of the largest, alpha (the intercept at
+# zero bytes) is an extrapolation, not a measured floor -- prefill ladders start at T=1024.
+ALPHA_NEAR_ZERO_FRACTION = 0.1
+MIN_FIT_POINTS = 3
+
+
+class Fit(NamedTuple):
+    """A fitted ladder. `r2`, `points` and `bandwidth_share` are part of the result, not
+    diagnostics: a caller that ignores them can print an impossible bandwidth."""
+    alpha_us: float
+    beta_gbps: float
+    r2: float
+    points: int
+    alpha_extrapolated: bool
+    excluded_rows: int
+    bandwidth_share: float  # transfer term / largest measured latency, at the top rung
+
+    @property
+    def beta_is_reliable(self) -> bool:
+        return self.r2 >= FIT_MIN_R2 and self.bandwidth_share >= FIT_MIN_BANDWIDTH_SHARE
+
+
+def _ep(document: dict) -> int:
+    return int(document["identity"]["case_factors"]["case"]["ep"])
+
+
+def _algbw_per_gpu(total_logical_bytes: int, latency_us: float, ep: int) -> float | None:
+    """Per-GPU effective GB/s, or None when the latency cannot yield a rate. Bytes are the
+    AGGREGATE world payload (routed_copies, routing.py), hence the divide by EP size."""
+    if latency_us <= 0:
+        return None
+    return total_logical_bytes / (latency_us * 1e-6) / 1e9 / ep
+
+
+def _fit_line(xs: list[float], ys: list[float]) -> tuple[float, float, float]:
+    """(intercept, slope, r2) of a 2-parameter least-squares fit (stdlib only)."""
+    n = len(xs)
+    mean_x, mean_y = sum(xs) / n, sum(ys) / n
+    slope = (sum((x - mean_x) * (y - mean_y) for x, y in zip(xs, ys))
+             / sum((x - mean_x) ** 2 for x in xs))
+    intercept = mean_y - slope * mean_x
+    ss_tot = sum((y - mean_y) ** 2 for y in ys)
+    ss_res = sum((y - (intercept + slope * x)) ** 2 for x, y in zip(xs, ys))
+    # ss_tot == 0: no variance to explain, so the fit cannot be credited with anything.
+    return intercept, slope, (1.0 - ss_res / ss_tot) if ss_tot else 0.0
+
+
+def fit_alpha_beta(document: dict, component: str, pct: str = "p50") -> Fit | None:
+    """Fit one component's ladder, or None when no fit is defensible at all.
+
+    Regress against the emitted bytes, never T (routed_copies is sub-linear in T at high
+    fan-out); p50 by default, since p99 carries tail noise. Rungs that failed the correctness
+    gate are excluded — a corrupt payload is not a measurement of how fast correct data moves.
+    """
+    usable, excluded = [], 0
+    for row in document["measurement"]["rows"]:
+        percentiles = row["components"][component]["percentiles_us"]
+        if not percentiles:
+            continue
+        if not row.get("correctness", {}).get("passed", True):
+            excluded += 1
+            continue
+        usable.append((row["byte_provenance"][component]["total_logical_bytes"],
+                       percentiles[pct]))
+    if len(usable) < MIN_FIT_POINTS or len({x for x, _ in usable}) < 2:
+        return None  # too few points, or zero-variance x (e.g. a degenerate ladder)
+    xs, ys = [x for x, _ in usable], [y for _, y in usable]
+    alpha_us, slope, r2 = _fit_line(xs, ys)
+    if slope <= 0:
+        return None  # anti-correlated / noise-dominated: no meaningful bandwidth term
+    return Fit(  # slope is us per aggregate byte; 1e-3/slope = aggregate GB/s; /ep = per-GPU
+        alpha_us=alpha_us,
+        beta_gbps=1e-3 / slope / _ep(document),
+        r2=r2,
+        points=len(usable),
+        alpha_extrapolated=min(xs) > ALPHA_NEAR_ZERO_FRACTION * max(xs),
+        excluded_rows=excluded,
+        bandwidth_share=(slope * max(xs) / max(ys)) if max(ys) > 0 else 0.0,
+    )
+
+
+def _format_fit(component: str, fit: Fit) -> str:
+    """One component's fit, withholding what the data does not support."""
+    if not fit.beta_is_reliable:
+        why = "poor fit" if fit.r2 < FIT_MIN_R2 else "not bandwidth-bound"
+        return (f"{component}: beta=unreliable [{why}] "
+                f"(R2={fit.r2:.2f}, transfer share={fit.bandwidth_share:.2f}, n={fit.points})")
+    alpha = f"alpha={fit.alpha_us:6.1f}us" + ("*" if fit.alpha_extrapolated else " ")
+    return (f"{component}: beta={fit.beta_gbps:6.1f} GB/s {alpha} "
+            f"(R2={fit.r2:.3f}, n={fit.points})")
+
+
+def _cell(row: dict, component: str, ep: int) -> str:
+    percentiles = row["components"][component]["percentiles_us"]
+    if not percentiles:
+        return f"{component}=n/a"
+    nbytes = row["byte_provenance"][component]["total_logical_bytes"]
+    p50 = _algbw_per_gpu(nbytes, percentiles["p50"], ep)
+    p99 = _algbw_per_gpu(nbytes, percentiles["p99"], ep)
+    return f"{component}=n/a" if p50 is None or p99 is None \
+        else f"{component}={p50:6.1f}/{p99:<6.1f}"
+
+
+def _sort_key(document: dict):
+    case = document["identity"]["case_factors"]["case"]
+    return (document["identity"]["case_factors"]["sku"], case["backend"],
+            case["precision"], case["phase"], case["ep"],
+            document.get("generated_at", ""))
+
+
+def _provenance(document: dict) -> str:
+    """Enough identity to tell two attempts of one case apart — CI result directories
+    routinely hold both, and they can disagree."""
+    run_id = document["identity"].get("allocation_factors", {}).get("run_id")
+    attempt = document["identity"].get("attempt_ordinal")
+    parts = ((document.get("generated_at") or "")[:19],
+             f"run {run_id}" if run_id else "", f"attempt {attempt}" if attempt else "")
+    return " · ".join(p for p in parts if p)
+
+
+def render(documents: list[dict]) -> str:
+    lines = [
+        "## CollectiveX EP bandwidth (per-GPU, logical payload)",
+        "",
+        "GB/s at p50/p99 *latency* -- the p99-latency figure is worst-case bandwidth, not a "
+        f"'p99 bandwidth'. `fit`: latency ~= alpha + bytes/beta over the ladder; beta is "
+        f"withheld below R2 {FIT_MIN_R2} or when the ladder never became bandwidth-bound, `*` "
+        "marks an extrapolated alpha, and rungs failing the correctness gate are excluded.",
+        "",
+    ]
+    for document in sorted(documents, key=_sort_key):
+        case = document["identity"]["case_factors"]["case"]
+        ep = _ep(document)
+        lines.append(
+            f"### {document['identity']['case_factors']['sku']} `{case['backend']}` "
+            f"{case['precision']} {case['phase']} ep{ep} — {document['outcome']['status']}"
+        )
+        if provenance := _provenance(document):
+            lines.append(f"  ({provenance})")
+        for row in document["measurement"]["rows"]:
+            cross_node = (row["routing"].get("locality") or {}).get("cross_node_fraction")
+            suffix = f"  xnode={cross_node * 100:4.0f}%" if cross_node is not None else ""
+            if not row.get("correctness", {}).get("passed", True):
+                suffix += "  [correctness FAILED]"
+            cells = "  ".join(_cell(row, component, ep) for component in COMPONENTS)
+            lines.append(f"  T={row['tokens_per_rank']:<5} {cells}{suffix}")
+        fits = [(c, f) for c in COMPONENTS if (f := fit_alpha_beta(document, c))]
+        dropped = max((f.excluded_rows for _, f in fits), default=0)
+        note = f"  (excluded {dropped} gate-failed rung(s))" if dropped else ""
+        lines.append("  fit " + ("   ".join(_format_fit(c, f) for c, f in fits)
+                                 if fits else "insufficient points") + note)
+        lines.append("")
+    if not documents:
+        lines.append("> No case-attempt documents found.")
+    return "\n".join(lines)
+
+
+def main() -> int:
+    parser = argparse.ArgumentParser(
+        description="CollectiveX per-component EP bandwidth (pure consumer; gates nothing)")
+    parser.add_argument("--results-dir", default="results")
+    parser.add_argument("--runner")
+    parser.add_argument("--ts")
+    parser.add_argument("--all", action="store_true",
+                        help="include non-success outcomes (skipped by default)")
+    args = parser.parse_args()
+    documents = load_results(args.results_dir, args.runner, args.ts)
+    if not args.all:
+        documents = [d for d in documents if d["outcome"]["status"] == "success"]
+    print(render(documents))
+    return 0
+
+
+if __name__ == "__main__":
+    raise SystemExit(main())
diff --git a/experimental/CollectiveX/tests/test_bandwidth.py b/experimental/CollectiveX/tests/test_bandwidth.py
new file mode 100644
index 0000000000..3b7380f154
--- /dev/null
+++ b/experimental/CollectiveX/tests/test_bandwidth.py
@@ -0,0 +1,129 @@
+#!/usr/bin/env python3
+"""Math tests for the bandwidth consumer (unit conversion, alpha/beta fit, fit gating)."""
+from __future__ import annotations
+
+import sys
+import unittest
+from pathlib import Path
+
+sys.path[:0] = [str(Path(__file__).resolve().parents[1])]
+
+import bandwidth  # noqa: E402
+
+COMPONENTS = bandwidth.COMPONENTS
+
+
+def _row(tokens, nbytes, latency, passed=True):
+    """A measurement row. `latency` is a scalar, or a per-component dict whose None marks
+    that component unavailable."""
+    lat = latency if isinstance(latency, dict) else dict.fromkeys(COMPONENTS, latency)
+    return {
+        "tokens_per_rank": tokens,
+        "components": {c: {"percentiles_us": None if lat[c] is None else {
+            "p50": lat[c], "p90": lat[c], "p95": lat[c], "p99": lat[c] * 2.0}}
+            for c in COMPONENTS},
+        "byte_provenance": {c: {"total_logical_bytes": nbytes} for c in COMPONENTS},
+        "correctness": {"passed": passed},
+        "routing": {"locality": {"cross_node_fraction": 0.5}},
+    }
+
+
+def _doc(rows, ep=2):
+    case = {"ep": ep, "backend": "deepep-v2", "precision": "bf16", "phase": "decode",
+            "suite": "s", "routing": "uniform"}
+    return {
+        "generated_at": "2026-07-25T22:12:55.760511+00:00",
+        "identity": {"attempt_ordinal": 1, "allocation_factors": {"run_id": "30177021271"},
+                     "case_factors": {"sku": "h100", "case": case}},
+        "measurement": {"rows": rows},
+        "outcome": {"status": "success"},
+    }
+
+
+def _linear(pairs, passed=True):
+    """Rows exactly on latency = 10us + bytes * 2e-6, i.e. alpha=10, beta_agg=500 GB/s."""
+    return [_row(t, b, 10.0 + b * 2e-6, passed) for t, b in pairs]
+
+
+LADDER = ((8, 1e6), (16, 2e6), (32, 3e6))
+
+
+class BandwidthMath(unittest.TestCase):
+    def test_algbw_per_gpu(self):
+        # 1e9 bytes in 1000us = 1e12 B/s = 1000 GB/s aggregate; /ep(2) = 500 per GPU.
+        self.assertAlmostEqual(bandwidth._algbw_per_gpu(1e9, 1000.0, 2), 500.0)
+        self.assertIsNone(bandwidth._algbw_per_gpu(1e9, 0.0, 2))
+
+    def test_fit_recovers_alpha_beta(self):
+        fit = bandwidth.fit_alpha_beta(_doc(_linear(LADDER)), "dispatch")
+        self.assertAlmostEqual(fit.alpha_us, 10.0, places=4)
+        self.assertAlmostEqual(fit.beta_gbps, 250.0, places=4)  # 500 aggregate / ep(2)
+        self.assertAlmostEqual(fit.r2, 1.0, places=6)
+        self.assertEqual(fit.points, 3)
+        self.assertTrue(fit.beta_is_reliable)
+
+    def test_fit_is_none_when_undefensible(self):
+        flat = [_row(t, b, 12.0) for t, b in LADDER]            # slope <= 0
+        self.assertIsNone(bandwidth.fit_alpha_beta(_doc(_linear(LADDER[:2])), "dispatch"))
+        self.assertIsNone(bandwidth.fit_alpha_beta(_doc(flat), "dispatch"))
+
+    def test_noisy_ladder_withholds_beta(self):
+        # A positive slope through noise still fits; printing its beta once produced a
+        # physically impossible 1018 GB/s per GPU on a B200 at R2 = 0.29.
+        rows = [_row(t, b, lat) for t, b, lat in (
+            (8, 1e6, 300.0), (16, 2e6, 40.0), (32, 3e6, 260.0),
+            (64, 4e6, 60.0), (128, 5e6, 320.0))]
+        fit = bandwidth.fit_alpha_beta(_doc(rows), "dispatch")
+        self.assertLess(fit.r2, bandwidth.FIT_MIN_R2)
+        self.assertFalse(fit.beta_is_reliable)
+        out = bandwidth.render([_doc(rows)])
+        self.assertIn("beta=unreliable", out)
+        self.assertNotIn("GB/s alpha", out)  # no number presented as measured
+
+    def test_latency_bound_ladder_withholds_beta_despite_high_r2(self):
+        # Real data gave beta = 3763 GB/s at R2 = 0.92: a near-zero slope explodes beta while
+        # the line still fits, so R2 cannot catch it — the transfer-share gate must.
+        rows = [_row(t, b, 500.0 + b * 1e-9)
+                for t, b in LADDER + ((64, 4e6), (128, 5e6))]
+        fit = bandwidth.fit_alpha_beta(_doc(rows), "dispatch")
+        self.assertGreater(fit.r2, bandwidth.FIT_MIN_R2)
+        self.assertLess(fit.bandwidth_share, bandwidth.FIT_MIN_BANDWIDTH_SHARE)
+        self.assertFalse(fit.beta_is_reliable)
+        self.assertIn("not bandwidth-bound", bandwidth._format_fit("dispatch", fit))
+
+    def test_alpha_marked_only_when_extrapolated(self):
+        prefill = bandwidth.fit_alpha_beta(  # starts at T=1024: intercept is extrapolated
+            _doc(_linear(((1024, 1e9), (2048, 2e9), (4096, 4e9), (8192, 8e9)))), "dispatch")
+        decode = bandwidth.fit_alpha_beta(   # reaches near zero bytes: alpha stands
+            _doc(_linear(((1, 1e5), (64, 6.4e6), (512, 5.12e7)))), "dispatch")
+        self.assertTrue(prefill.alpha_extrapolated)
+        self.assertFalse(decode.alpha_extrapolated)
+        self.assertIn("*", bandwidth._format_fit("dispatch", prefill))
+        self.assertNotIn("*", bandwidth._format_fit("dispatch", decode))
+
+    def test_gate_failed_rung_excluded_from_fit_and_marked(self):
+        rows = _linear(LADDER) + [_row(64, 4e6, 999.0, passed=False)]
+        fit = bandwidth.fit_alpha_beta(_doc(rows), "dispatch")
+        self.assertEqual((fit.points, fit.excluded_rows), (3, 1))
+        self.assertAlmostEqual(fit.beta_gbps, 250.0, places=4)  # the corrupt rung didn't steer it
+        out = bandwidth.render([_doc(rows)])
+        self.assertIn("[correctness FAILED]", out)
+        self.assertIn("excluded 1 gate-failed rung", out)
+
+    def test_render_marks_unavailable_and_separates_attempts(self):
+        unavailable = [_row(t, b, {"dispatch": None, "combine": 5.0, "roundtrip": 6.0})
+                       for t, b in LADDER]
+        out = bandwidth.render([_doc(unavailable)])
+        self.assertIn("dispatch=n/a", out)
+        self.assertIn("xnode=  50%", out)
+        second = _doc(_linear(LADDER))
+        second["identity"]["attempt_ordinal"] = 2
+        second["generated_at"] = "2026-07-25T23:00:00.000000+00:00"
+        out = bandwidth.render([_doc(_linear(LADDER)), second])
+        self.assertIn("attempt 1", out)
+        self.assertIn("attempt 2", out)
+        self.assertIn("run 30177021271", out)
+
+
+if __name__ == "__main__":
+    unittest.main()