diff --git a/bench/analysis/experimental_gap.py b/bench/analysis/experimental_gap.py
new file mode 100644
index 0000000..b86d7c5
--- /dev/null
+++ b/bench/analysis/experimental_gap.py
@@ -0,0 +1,105 @@
+#!/usr/bin/env python3
+"""Where DeltaSignal loses to MP-BioPath on the experimental axis, case by case.
+
+MP-BioPath's per-case calls are published in
+mp-biopath-pathways/analysis_results/MP_BioPathReactomePathwayAccuracy-Tests-updated.tsv
+(`mp_biopath_state` against `experimental_value`, 849 cases, 643 correct).
+This joins them with an arm's experimental_cases.tsv and splits the cases into
+both right / only MP-BioPath / only DeltaSignal / neither, by pathway,
+perturbation, our failure category and (truth, our call).
+
+A case that does not join, or whose experimental value disagrees with our
+expected label, is an error, not a skip: a silent mis-join would make the gap
+look like whatever the join happened to keep.
+
+Usage: experimental_gap.py ARM_RESULTS_DIR [--mpb TSV] [--out cases.tsv]
+"""
+from __future__ import annotations
+
+import argparse
+import collections
+import csv
+import sys
+from pathlib import Path
+
+DEFAULT_MPB = (Path.home() / "gitroot" / "mp-biopath-pathways" / "analysis_results"
+               / "MP_BioPathReactomePathwayAccuracy-Tests-updated.tsv")
+LABEL = {"0": "DOWN", "1": "NORMAL", "2": "UP"}
+
+
+def load_mpb(path: Path) -> dict[tuple, tuple[str, str]]:
+    """(pathway, gene, direction, key_output) -> (experimental, mp-biopath state)."""
+    out: dict[tuple, tuple[str, str]] = {}
+    with path.open(newline="") as handle:
+        for r in csv.DictReader(handle, delimiter="\t"):
+            if r["experimental_value"] in ("NA", ""):
+                continue
+            gene, direction = r["scenario"].rsplit("_", 1)
+            key = (r["pathway_name"], gene, direction, r["keyoutput_id"])
+            if key in out:
+                raise SystemExit(f"duplicate MP-BioPath row {key}")
+            out[key] = (r["experimental_value"], r["mp_biopath_state"])
+    return out
+
+
+def split(mpb: dict, ours_rows: list[dict]) -> list[dict]:
+    rows = []
+    for o in ours_rows:
+        key = (o["pathway"], o["gene"], o["direction"], o["key_output"])
+        if key not in mpb:
+            raise SystemExit(f"case not in MP-BioPath's table: {key}")
+        truth, them = mpb[key]
+        if truth != o["expected"]:
+            raise SystemExit(f"experimental value {truth} != our expected {o['expected']} for {key}")
+        us = o["predicted"] if o["valid"] == "1" else "invalid"
+        rows.append({"pathway": key[0], "gene": key[1], "direction": key[2], "key_output": key[3],
+                     "truth": truth, "mpbiopath": them, "deltasignal": us,
+                     "quadrant": ("MPB" if them == truth else "-") + "/" + ("DS" if us == truth else "-"),
+                     "category": o.get("category", "")})
+    missing = set(mpb) - {(r["pathway"], r["gene"], r["direction"], r["key_output"]) for r in rows}
+    if missing:
+        raise SystemExit(f"{len(missing)} MP-BioPath cases have no row in the arm, e.g. {sorted(missing)[0]}")
+    return rows
+
+
+def report(rows: list[dict]) -> str:
+    out = []
+    q = collections.Counter(r["quadrant"] for r in rows)
+    mpb_right = q["MPB/DS"] + q["MPB/-"]; ds_right = q["MPB/DS"] + q["-/DS"]
+    out.append(f"cases {len(rows)}: MP-BioPath {mpb_right}, DeltaSignal {ds_right}, gap {mpb_right - ds_right:+d}")
+    out.append(f"  both right {q['MPB/DS']}, only MP-BioPath {q['MPB/-']}, "
+               f"only DeltaSignal {q['-/DS']}, neither {q['-/-']}; either {len(rows) - q['-/-']}")
+    pw = collections.defaultdict(collections.Counter)
+    for r in rows:
+        pw[r["pathway"]][r["quadrant"]] += 1
+    out.append("per pathway: only-MPB | only-DS | net")
+    for p, c in sorted(pw.items(), key=lambda x: -(x[1]["MPB/-"] - x[1]["-/DS"])):
+        out.append(f"  {p[:40]:40s} {c['MPB/-']:4d} | {c['-/DS']:4d} | {c['MPB/-'] - c['-/DS']:+4d}  (n={sum(c.values())})")
+    lost = [r for r in rows if r["quadrant"] == "MPB/-"]
+    out.append("only-MPB by our category: " + str(collections.Counter(r["category"] for r in lost).most_common()))
+    out.append("only-MPB by (truth, our call): " + str(collections.Counter(
+        (LABEL.get(r["truth"]), LABEL.get(r["deltasignal"], r["deltasignal"])) for r in lost).most_common()))
+    pert = collections.Counter((r["pathway"][:24], r["gene"], "KO" if r["direction"] == "0" else "UP") for r in lost)
+    out.append("only-MPB by perturbation: " + str(pert.most_common(15)))
+    return "\n".join(out)
+
+
+def main() -> int:
+    ap = argparse.ArgumentParser(description=__doc__, formatter_class=argparse.RawDescriptionHelpFormatter)
+    ap.add_argument("arm", type=Path)
+    ap.add_argument("--mpb", type=Path, default=DEFAULT_MPB)
+    ap.add_argument("--out", type=Path)
+    a = ap.parse_args()
+    with (a.arm / "experimental_cases.tsv").open(newline="") as handle:
+        ours = list(csv.DictReader(handle, delimiter="\t"))
+    rows = split(load_mpb(a.mpb), ours)
+    print(report(rows))
+    if a.out:
+        with a.out.open("w", newline="") as handle:
+            w = csv.DictWriter(handle, fieldnames=list(rows[0]), delimiter="\t")
+            w.writeheader(); w.writerows(rows)
+    return 0
+
+
+if __name__ == "__main__":
+    sys.exit(main())
diff --git a/bench/analysis/mpbiopath_network_adapter.py b/bench/analysis/mpbiopath_network_adapter.py
index 9027259..9ac2569 100644
--- a/bench/analysis/mpbiopath_network_adapter.py
+++ b/bench/analysis/mpbiopath_network_adapter.py
@@ -44,6 +44,7 @@
 
 import argparse
 import csv
+import re
 import sys
 from pathlib import Path
 
@@ -80,6 +81,10 @@ def catalog_ids(catalog: Path) -> dict[str, str]:
     mapping: dict[str, str] = {}
     for directory in sorted(p for p in catalog.iterdir() if p.is_dir()):
         name = directory.name
+        if re.fullmatch(r"R-HSA-\d+", name):
+            # Current generator naming: the directory IS the stable id.
+            mapping[name] = name
+            continue
         if "_R-HSA-" not in name:
             continue
         prefix, _, numeric = name.rpartition("_R-HSA-")
@@ -100,8 +105,11 @@ def case_pathway_names(cases_tsv: Path) -> dict[str, str]:
     """
     mapping: dict[str, str] = {}
     with cases_tsv.open(newline="") as handle:
-        for row in csv.DictReader(handle, delimiter="\t"):
-            pid = (row.get("pathway_id") or "").strip()
+        lines = (ln for ln in handle if not ln.startswith("#"))
+        for row in csv.DictReader(lines, delimiter="\t"):
+            # A cases dump (pathway_id, pathway_name) or the versioned
+            # pathway list bench/catalog_pathways.tsv (id, pathway_name).
+            pid = (row.get("pathway_id") or row.get("id") or "").strip()
             name = (row.get("pathway_name") or "").strip()
             if pid and name:
                 mapping[pid] = name
@@ -165,7 +173,7 @@ def main() -> int:
             missing_mpb.append(f"{prefix} ({stable})")
             continue
         edges = parse_network(source)
-        target = args.out / f"{prefix}_{stable}"
+        target = args.out / (stable if prefix == stable else f"{prefix}_{stable}")
         write_catalog_dir(target, edges)
         written.append((target.name, len(edges)))
 
diff --git a/bench/analysis/test_experimental_gap.py b/bench/analysis/test_experimental_gap.py
new file mode 100644
index 0000000..af94909
--- /dev/null
+++ b/bench/analysis/test_experimental_gap.py
@@ -0,0 +1,40 @@
+"""experimental_gap.py: the join must be exact and the quadrants right."""
+import sys
+from pathlib import Path
+
+import pytest
+
+sys.path.insert(0, str(Path(__file__).parent))
+from experimental_gap import report, split  # noqa: E402
+
+
+def case(p, g, d, k, pred, exp, valid="1", cat="pass"):
+    return {"pathway": p, "gene": g, "direction": d, "key_output": k, "predicted": pred,
+            "expected": exp, "valid": valid, "category": cat}
+
+
+MPB = {("P", "A", "0", "1"): ("0", "0"), ("P", "A", "0", "2"): ("2", "1"),
+       ("P", "B", "2", "1"): ("2", "2"), ("Q", "C", "0", "9"): ("1", "2")}
+
+
+def test_quadrants_and_invalid_counts_as_wrong():
+    rows = split(MPB, [case("P", "A", "0", "1", "2", "0"),            # only MPB
+                       case("P", "A", "0", "2", "2", "2"),            # only DS
+                       case("P", "B", "2", "1", "1", "2", valid="0"),  # invalid -> only MPB
+                       case("Q", "C", "0", "9", "0", "1")])           # neither
+    assert [r["quadrant"] for r in rows] == ["MPB/-", "-/DS", "MPB/-", "-/-"]
+    text = report(rows)
+    assert "MP-BioPath 2, DeltaSignal 1, gap +1" in text
+    assert "either 3" in text
+
+
+def test_a_label_disagreement_is_an_error():
+    with pytest.raises(SystemExit):
+        split(MPB, [case("P", "A", "0", "1", "0", "2")])
+
+
+def test_a_case_missing_on_either_side_is_an_error():
+    with pytest.raises(SystemExit):
+        split(MPB, [case("P", "A", "0", "1", "0", "0")])               # 3 MPB cases unmatched
+    with pytest.raises(SystemExit):
+        split({}, [case("P", "A", "0", "1", "0", "0")])
diff --git a/bench/analysis/test_mpbiopath_adapter.py b/bench/analysis/test_mpbiopath_adapter.py
index ee68b93..ea286ac 100644
--- a/bench/analysis/test_mpbiopath_adapter.py
+++ b/bench/analysis/test_mpbiopath_adapter.py
@@ -15,7 +15,7 @@
 
 sys.path.insert(0, str(Path(__file__).parent))
 from mpbiopath_network_adapter import (  # noqa: E402
-    CONJUNCTION, POLARITY, UUID_PREFIX, case_pathway_names, parse_network,
+    CONJUNCTION, POLARITY, UUID_PREFIX, case_pathway_names, catalog_ids, parse_network,
     write_catalog_dir,
 )
 
@@ -89,9 +89,22 @@ def check(label, actual, expected):
     check("case name bridge", case_pathway_names(cases),
           {"R-HSA-453279": "Mitotic_G1-G1_S_phases"})
 
+    # Current generator naming (bare stable-id directories) and the versioned
+    # pathway list as the name bridge; both were unreadable until 2026-10-09,
+    # so the adapter silently adapted nothing.
+    cat = tmp / "catalog"
+    (cat / "R-HSA-453279").mkdir(parents=True)
+    (cat / "Old_Name_R-HSA-69620").mkdir()
+    check("bare and legacy catalog dirs", catalog_ids(cat),
+          {"R-HSA-453279": "R-HSA-453279", "Old_Name": "R-HSA-69620"})
+    plist = tmp / "catalog_pathways.tsv"
+    plist.write_text("# comment\nid\tpathway_name\nR-HSA-453279\tMitotic_G1-G1_S_phases\n")
+    check("pathway list as name bridge", case_pathway_names(plist),
+          {"R-HSA-453279": "Mitotic_G1-G1_S_phases"})
+
 if failures:
     print("FAILED:")
     for f in failures:
         print("  -", f)
     sys.exit(1)
-print(f"all adapter assertions passed ({12} checks)")
+print(f"all adapter assertions passed ({14} checks)")
diff --git a/specs/047-experimental-gap/research.md b/specs/047-experimental-gap/research.md
new file mode 100644
index 0000000..a57c04c
--- /dev/null
+++ b/specs/047-experimental-gap/research.md
@@ -0,0 +1,287 @@
+# 047: Where the experimental gap to MP-BioPath is, case by case
+
+**Status:** analysis in progress (2026-10-09). No behaviour change yet.
+
+## Why
+
+On the experimental axis we score 594/849 (69.96%), and MP-BioPath scores
+643/849 (75.74%). Curator-axis work has gained 0.3–0.7pp per feature while
+experimental drifted from 627 to 594 across generator changes. Experimental
+is the axis measured against biology, so it should be the target, not a
+guardrail. This spec decomposes the gap before any lever is designed.
+
+## Data
+
+MP-BioPath's per-case calls are published in
+`mp-biopath-pathways/analysis_results/MP_BioPathReactomePathwayAccuracy-Tests-updated.tsv`
+(`mp_biopath_state` against `experimental_value`).
+`bench/analysis/experimental_gap.py` joins them with an arm's
+`experimental_cases.tsv`:
+- all 849 cases join;
+- the experimental value equals our expected label on all 849;
+- 643 MP-BioPath-correct reproduces.
+
+A missing case or a label disagreement is an error, not a skip.
+
+## Decomposition (canonical: build `20261004-1959_6990015`, solver 5979e48)
+
+| | Cases |
+|---|---|
+| Both right | 540 |
+| **Only MP-BioPath right** | **103** |
+| Only DeltaSignal right | 54 |
+| Neither | 152 |
+
+The two are complementary: either one right covers 697/849 (82.1%).
+
+**Per pathway** (only-MP-BioPath | only-DeltaSignal | net):
+
+| Pathway | Only MP-BioPath | Only DeltaSignal | Net |
+|---|---|---|---|
+| TP53 | 32 | 18 | +14 |
+| Mitotic G1 | 20 | 11 | +9 |
+| PIP3 | 18 | 9 | +9 |
+| Cell Cycle Checkpoints | 8 | 1 | +7 |
+| RAF | 9 | 3 | +6 |
+| HDR | 4 | 0 | +4 |
+| Mitotic Prophase | 2 | 0 | +2 |
+| S Phase | 1 | 1 | 0 |
+| ERBB2 | 3 | 4 | −1 |
+| WNT | 6 | 7 | −1 |
+
+The gap is spread across pathways, not one.
+
+**The 103 only-MP-BioPath cases:**
+- **By our failure category:** propagator_missed 47, no_path 31,
+  false_positive_change 24, readout not in network 1.
+- **By (truth, our call):**
+
+  | Truth → our call | Cases |
+  |---|---|
+  | DOWN → UP | 30 |
+  | UP → NORMAL | 22 |
+  | DOWN → NORMAL | 15 |
+  | NORMAL → UP | 14 |
+  | UP → DOWN | 11 |
+  | NORMAL → DOWN | 10 |
+
+  **41 are sign inversions**: the signal reaches the readout backwards.
+- **By perturbation (top):**
+
+  | Perturbation | Cases | Note |
+  |---|---|---|
+  | PTEN UP (PIP3) | 9 | all inverted |
+  | RBL2 KO (Mitotic G1) | 7 | |
+  | PDPK1 KO (PIP3) | 7 | inverted |
+  | DAXX KO + UP (TP53) | 11 | |
+  | CDKN2A UP + KO (TP53) | 8 | |
+  | RB1 KO | 4 | |
+  | NF1 KO | 4 | |
+  | BRCA1 KO | 4 | |
+
+## Trace 1: PTEN overexpression inverts (9 cases)
+
+The readouts read 100× (UP) after PTEN overexpression; truth is DOWN. They
+also read 100× after PTEN knockout, so they do not depend on PTEN's direction.
+
+1. The bench pins all three PTEN roots at 80×: PTEN gene (R-HSA-5632949),
+   K48polyUb-PTEN (R-HSA-8948777) and PTEN mRNA (R-HSA-2318746).
+2. PTEN translation, R-HSA-8944497 ("PTEN mRNA translation is negatively
+   regulated by microRNAs"), has **12 negative regulators**. Each is a
+   `miR-x RISC:PTEN mRNA` complex, and each reads 64.5×.
+3. Under `divide` they multiply: 80 / 64.5¹² ≈ 1.5e-20. **Overexpressing PTEN
+   shuts off PTEN protein.** The PTEN copy that depletes PI(4,5)P2 reads 0,
+   PIP2 is de-repressed to 6.7×, PIP3 rises, and a positive loop rails AKT to
+   100×.
+4. The specs/022 self-inhibitor rule is on (227 slots flagged in the
+   pathway), and the complexes contain PTEN mRNA. The rule still does not
+   fire, because it requires the shared input node to reach the inhibitor.
+   PTEN mRNA exists as **two silo copies**:
+   - root bb6f1223 feeds only translation;
+   - produced copy 337e82aa (from transcription of the pinned gene) feeds the
+     12 RISC-binding reactions.
+
+   Translation reads one copy, its inhibitors are built from the other.
+5. With only the mRNA pinned, the readout reads 0.00065 (DOWN, correct).
+
+Biologically, RISC binding sequesters a fraction of the mRNA. More mRNA
+means proportionally more RISC:mRNA, not less translation. The failure is
+the specs/012/022 double count, unmasked by copy siloing.
+
+**Candidate levers** (not designed yet):
+- (a) the self-inhibitor rule matches the shared input by entity across copies
+  rather than by node reachability;
+- (b) the generator joins the transcription product to the translation input
+  (the copy silo).
+
+Either is a pre-registered arm, measured with experimental as the primary
+axis.
+
+## Trace 2: PDPK1 knockout inverts through a SARS-CoV-2 inhibitor (7 cases)
+
+PDPK1 knockout should lower phospho-AKT; truth is DOWN, and we read 100×.
+
+1. "PDPK1 phosphorylates AKT at T308" (R-HSA-198270) has a negative
+   regulator, `N:M:PDPK1` (R-HSA-9755778): the SARS-CoV-2 N dimer and M
+   protein bound to PDPK1. Reactome added it after 2019, so MP-BioPath's
+   networks lack it.
+2. Knocking out PDPK1 zeroes that complex, the reaction is de-repressed to the
+   ceiling, and the AKT positive loop
+   (R-HSA-198270 ↔ R-HSA-2317313 ↔ R-HSA-2317314) rails to 100×.
+3. In a human cell without the virus the inhibitor does not exist. Holding it
+   at baseline makes it a constant factor, which cancels in fold-change. That
+   is the specs/032 drug rule's mechanism applied to pathogen-derived nodes.
+
+Across the ten experimental pathways, the containment tables name only three
+non-human entities, all SARS-CoV-2: N dimer, a modified N, and M. A pathogen
+rule would be narrow.
+
+Separately, the catalyst `p-S-AKT:PDPK1:PIP3` (R-HSA-2317313) reads 100× under
+a PDPK1 knockout. Its PDPK1 comes in through the loop, not from the pinned
+copy. That is the same copy-silo pattern as trace 1, not yet traced to a
+line.
+
+**Candidate lever:** `DS_PATHOGEN_MODE=inert`, or the generator listing
+non-human-derived nodes the way `drugs.csv` lists drugs. Pre-registered and
+measured with experimental as the primary axis.
+
+## Network or propagator? The control
+
+**Arms.** Solver 3a8d59d (`src/` identical to 0525296).
+- `mpbnet`: our propagator on MP-BioPath's own 2019 networks (build
+  `20261009-2051_mpbnet2019`, 85 pathways, via
+  `bench/analysis/mpbiopath_network_adapter.py`).
+- `canon_tablefree`: the canonical networks under the same solver settings.
+
+Both run with `DS_CYCLE_MODE=off`, `DS_SELF_INHIBITOR_WEIGHT=off` and
+`DS_SELF_INHIBITOR_LEAVES=0`, because MP-BioPath's networks carry no pool or
+containment tables. The canonical default arm (`5979e48`) is shown too.
+
+| Pathway | n | MP-BioPath | Canonical | Canonical, table-free | **Ours on MPB networks** |
+|---|---|---|---|---|---|
+| TP53 | 257 | 182 | 168 | 167 | 179 |
+| PIP3 | 200 | 177 | 168 | 177 | 142 |
+| Mitotic G1 | 89 | 39 | 30 | 30 | 41 |
+| Cell Cycle Checkpoints | 55 | 46 | 39 | 31 | 47 |
+| WNT | 51 | 36 | 37 | 31 | 36 |
+| RAF | 49 | 45 | 39 | 17 | 49 |
+| ERBB2 | 49 | 41 | 42 | 42 | 41 |
+| HDR | 48 | 40 | 36 | 36 | 36 |
+| Mitotic Prophase | 26 | 21 | 19 | 20 | 22 |
+| S Phase | 25 | 16 | 16 | 16 | 17 |
+| **Total** | 849 | **643** | **594** | **567** | **610** |
+
+**Findings:**
+1. **The gap is network construction, not the propagator, in every
+   pathway except PIP3.** On MP-BioPath's networks our propagator matches or
+   beats MP-BioPath in 8 of 10 pathways: RAF 49 vs 45, Mitotic G1 41 vs 39,
+   Cell Cycle Checkpoints 47 vs 46, Prophase, S Phase, WNT, ERBB2, and TP53
+   179 vs 182.
+2. **Under identical solver settings, our networks cost 43 cases**
+   (567 vs 610). The largest single pathway is RAF, −32 (17 vs 49); the
+   rest is spread over Mitotic G1, Cell Cycle Checkpoints, TP53 and WNT.
+3. **Our network-specific rules recover 27 of the 43** (567 → 594). Cycle
+   balance does most of it in RAF (17 → 39). In PIP3 they cost 9
+   (177 → 168).
+4. **PIP3 is the exception.** Our propagator on their PIP3 network scores 142
+   against their 177; on our own PIP3 network it scores 168–177. So PIP3's
+   shortfall on their network is a propagator matter, irrelevant to our
+   networks.
+5. **Choosing per pathway** between canonical and their networks gives 638,
+   against MP-BioPath's 643. Nothing in the propagator stands between us and
+   MP-BioPath's experimental accuracy; the networks do.
+
+**Where to look:**
+- The network share lives in RAF (39 vs 49), Mitotic G1 (30 vs 41), Cell Cycle
+  Checkpoints (39 vs 47) and TP53 (168 vs 179). That is 41 cases across 4
+  pathways.
+- specs/034 already compared RAF's networks: MP-BioPath's ~1,800 hand edits,
+  mostly reversal-enzyme inhibitions of interconversion cycles, 187 absent
+  from ours.
+- **Next:** for each discordant case in these four pathways, diff the two
+  networks along the gene → readout route (both are keyed by Reactome ids)
+  and classify what MP-BioPath's network has that ours lacks, or vice versa.
+
+## Trace 3: RAF, a recycled scaffold amplifies itself (NRAS KO, and likely NF1 KO)
+
+These are the 10 RAF cases our propagator gets right on MP-BioPath's network
+and wrong on ours:
+
+| Perturbation | Cases | Truth → ours |
+|---|---|---|
+| NRAS KO | 3 | DOWN → UP |
+| NF1 KO | 4 | UP → DOWN or no change |
+| ARAF / BRAF / RAF1 KO → R-HSA-5674341 | 3 | DOWN → no change |
+
+**NRAS KO → p-T,Y MAPK dimers (R-HSA-1268261) reads 63×.** The trace:
+1. "RAF/MAPK scaffolds" (R-HSA-5672717, a CandidateSet of ARRB1, ARRB2,
+   focal-adhesion and others) is an input to "MAP2Ks and MAPKs bind to the
+   activated RAF complex" (R-HSA-5672972).
+2. RAF phosphorylates MAP2K (R-HSA-5672978), MAP2Ks phosphorylate MAPKs
+   (R-HSA-5672973), and "Dissociation of RAS:RAF complex" (R-HSA-5672980)
+   **releases the scaffold**. Reactome curates that: it is an output of
+   R-HSA-5672980 in Neo4j.
+3. The released scaffold re-enters step 1 at the complex's own fold. This
+   positive loop rails the cascade to 63–100× whatever RAS does, so NRAS KO
+   reads UP.
+4. MP-BioPath's RAF network has the same scaffold-member → R-HSA-5672972
+   edges, but its R-HSA-5672980 outputs only R-HSA-5672712, 169289 and
+   5672721. **No scaffold release**, so no recycling loop, and RAS
+   perturbations reach the readouts. Whether 2019 Reactome lacked the
+   release or it was a hand cut, their network is acyclic here.
+
+**The class: a conserved carrier.** The scaffold is bound at one step,
+carried unchanged through two modification steps, and released at a
+fourth. Its total is conserved, so it cannot amplify itself.
+- specs/039 balance pools model exactly that conservation, but the
+  generator's pool detection finds interconversion cycles of a modified
+  species. It does not find a carrier bound and released across a
+  multi-step chain: RAF has one pool of 3 rows, none touching the scaffold.
+- The blanket alternatives were refuted: specs/035 (all conserved loop
+  inputs held at baseline, experimental −122, it held PIP3) and specs/040
+  rule A.
+
+**Candidate lever:** extend pool detection to multi-step carrier cycles (a
+species bound by one reaction and released unchanged by a later reaction in
+the same chain), solved by the existing balance rule. Narrower than
+specs/035, and consistent with D1 identity. Measured with experimental as
+the primary axis.
+
+## Trace 4: Mitotic G1, released E2F copies are not the ones that drive expression (no-path class)
+
+In Mitotic G1, 20 cases are right on MP-BioPath's network and wrong on ours
+(9 the reverse). Most are no-path in ours: CCND1 UP (5), RB1 KO (3+),
+SRC UP (3), CDK4, CDKN1B, CDKN2A, PPP2R1A.
+
+**CCND1 UP → CDT1 (R-HSA-68536):**
+- Forward reach from the CCND1 root is 459 nodes. Backward reach from the
+  readout is 97 nodes. They share **no node**.
+- They do share 7 **entities** as different, unjoined copies: E2F1, E2F4,
+  E2F5, TFDP1, TFDP2, RBL1, RBL2.
+- CDK4/6:CCND phosphorylates RB/RBL and releases E2F:DP. The released copies
+  are not the copies that "CDT1 gene expression is stimulated by E2F1"
+  reads.
+
+This is the dissociation-sink silo of specs/015 and 031: a released subunit
+beside an unfed twin. Blanket bridging of it was measured harmful four times
+on the curator axis (fan-out median 139). MP-BioPath's hand-built network
+connects it.
+
+`route_diff.py` (analysis dir) can't match MP-BioPath's set nodes (Cyclin D,
+CDK4/CDK6, sets of complexes) to our member-decomposed nodes, so its ABSENT
+marks there are artefacts. The reach comparison above is the reliable test.
+
+## Mechanism classes so far
+
+| Class | Example | Cases (approx.) | Prior measurements |
+|---|---|---|---|
+| Copy silo hides a self-inhibitor | PTEN UP | 9 | specs/022 rule needs node reachability |
+| Pathogen-derived regulator, added after 2019 | PDPK1 KO | 7 | specs/032 drug rule is the analogue |
+| Multi-step carrier cycle amplifies itself | RAF NRAS/NF1 KO | ~10 | specs/035 and 040 blanket rules refuted |
+| Released transcription factor in a silo | G1 CCND1/RB1/SRC/CDK4 | ~16 | specs/015/031 bridging refuted on curator |
+| Over-coupling (false change) | TP53 DAXX | 12 | — |
+
+Every class is a network-construction issue. Each was seen before on the
+curator axis, where blanket versions failed. What has never been done is to
+target them with experimental as the primary axis and a rule narrow enough
+to match MP-BioPath's hand edits.