diff --git a/Makefile b/Makefile
index f95d9d3..c86a0a1 100644
--- a/Makefile
+++ b/Makefile
@@ -36,6 +36,15 @@ llm:               ## run the LLM prototypes: merge (fusion) + moe (union) + dir
 	uv run python -m llm.experiment configs/llm/moe.yaml
 	uv run python -m llm.experiment configs/llm/directed.yaml
 
+llm-speciation:    ## LLM-tier speciation: conflict cliff (replace + de-confounded add) and duration null
+	uv run python -m llm.experiment configs/llm/speciation.yaml
+	uv run python -m llm.experiment configs/llm/speciation_add.yaml
+
+llm-seeds:         ## multi-seed firm-up (heavy): merge x5, moe-hard x3, directed-hard x3
+	uv run python -m llm.experiment configs/llm/merge_seeds.yaml
+	uv run python -m llm.experiment configs/llm/moe_hard_seeds.yaml
+	uv run python -m llm.experiment configs/llm/directed_hard_seeds.yaml
+
 layer2: neural     ## alias: Layer 1.5 is the current Layer-2 deliverable (LLM rung deferred)
 
 figures:           ## regenerate figures from committed results
diff --git a/configs/llm/directed_hard_seeds.yaml b/configs/llm/directed_hard_seeds.yaml
new file mode 100644
index 0000000..cb6c2c3
--- /dev/null
+++ b/configs/llm/directed_hard_seeds.yaml
@@ -0,0 +1,22 @@
+experiment: llm_directed_hard_seeds
+kind: llm_directed
+seed: 1
+seeds: [1, 2, 3]
+n_replicates: 1
+
+# Multi-seed 0.5B directed sex on the HARD benchmark (offspring selection pays off exactly where the
+# default soup is suboptimal). Fixed test/val sets; training + Dirichlet-offspring seed varies; reuses
+# the per-seed hard specialists trained by llm_moe_hard_seeds when present.
+
+base_model: Qwen/Qwen2.5-0.5B-Instruct
+hard: true
+families: [lists, strings, arith]
+n_train: 400
+n_test: 80
+n_val: 60
+n_candidates: 16
+concentration: 0.5
+epochs: 3
+lora: {r: 16, alpha: 32}
+
+output: {dir: results/llm_directed_hard_seeds}
diff --git a/configs/llm/merge_seeds.yaml b/configs/llm/merge_seeds.yaml
new file mode 100644
index 0000000..6fa8c84
--- /dev/null
+++ b/configs/llm/merge_seeds.yaml
@@ -0,0 +1,20 @@
+experiment: llm_merge_seeds
+kind: llm_merge
+seed: 1
+seeds: [1, 2, 3, 4, 5]
+n_replicates: 1
+
+# Multi-seed firm-up of the 0.5B merge experiment (PNAS work order Phase 3: removes the "one seed"
+# objection). Same protocol as configs/llm/merge.yaml; the test sets stay FIXED (seed 1000+i per
+# family) while the training seed varies, so across-seed variance is training variance only.
+# Specialists cache per-seed (spec_<family>_s<seed>).
+
+base_model: Qwen/Qwen2.5-0.5B-Instruct
+families: [lists, strings, arith]
+n_train: 600
+n_test: 100
+epochs: 3
+lora: {r: 16, alpha: 32}
+merges: [soup, ties]
+
+output: {dir: results/llm_merge_seeds}
diff --git a/configs/llm/moe_hard_seeds.yaml b/configs/llm/moe_hard_seeds.yaml
new file mode 100644
index 0000000..3a6011b
--- /dev/null
+++ b/configs/llm/moe_hard_seeds.yaml
@@ -0,0 +1,21 @@
+experiment: llm_moe_hard_seeds
+kind: llm_moe
+seed: 1
+seeds: [1, 2, 3]
+n_replicates: 1
+
+# Multi-seed 0.5B union-vs-fusion on the HARD benchmark (the headroom regime where the ordering
+# matters). Fixed test sets; training seed varies; hard specialists cache per-seed
+# (spec_<family>_hard_s<seed>). Companion to the single-seed llm_moe_hard and the 7B HPC runs.
+
+base_model: Qwen/Qwen2.5-0.5B-Instruct
+hard: true
+families: [lists, strings, arith]
+n_train: 400
+n_test: 80
+n_route: 32
+epochs: 3
+lora: {r: 16, alpha: 32}
+operators: [soup, ties, moe_oracle, moe_learned]
+
+output: {dir: results/llm_moe_hard_seeds}
diff --git a/configs/llm/speciation.yaml b/configs/llm/speciation.yaml
new file mode 100644
index 0000000..5155b9a
--- /dev/null
+++ b/configs/llm/speciation.yaml
@@ -0,0 +1,32 @@
+experiment: llm_speciation
+kind: llm_speciation
+seed: 1
+n_replicates: 1
+
+# LLM-tier model speciation (E13 in language-model weights; PNAS work order Phase 3). Two LoRA
+# children from the same frozen base — so there is NO permutation/rescaling ambiguity by construction:
+# any merge failure is functional incompatibility, isolated architecturally. Two sweeps:
+#   conflict_fracs — the IMPOSED cliff: each child has a private disjoint family (A: strings,
+#     B: arith) plus a shared set of AMBIGUOUS sort prompts ("Sort the list [...]", no direction)
+#     answered ascending by A and descending by B (each convention harmless alone, contradictory
+#     jointly — the BDM structure). Prediction: the MERGED model's private-family competence degrades
+#     and its convention coherence (max of asc/desc grading) collapses as conflict grows, while each
+#     PARENT stays fine — hybrid breakdown in verifier units, echoing the MLP cliff.
+#   durations — the EMERGENT null: pure disjoint specialists over-trained (epochs swept), zero shared
+#     data. Arbitrates the MLP tier's null (no emergent isolation; the merge rescued specialists at
+#     every divergence) against the empirical report that averaging prefers under-trained experts
+#     (arXiv:2607.11997). Pre-registered readings: merged quality falls with duration while parents'
+#     own-family quality holds -> emergent incompatibility at the LLM tier; otherwise the null
+#     generalises. Either outcome is reportable; do not tune toward one.
+
+base_model: Qwen/Qwen2.5-0.5B-Instruct
+family_a: strings
+family_b: arith
+n_train: 400
+n_test: 100
+epochs: 3
+lora: {r: 16, alpha: 32}
+conflict_fracs: [0.0, 0.25, 0.5, 0.75, 1.0]
+durations: [1, 3, 6, 12]
+
+output: {dir: results/llm_speciation}
diff --git a/configs/llm/speciation_add.yaml b/configs/llm/speciation_add.yaml
new file mode 100644
index 0000000..65c4cd5
--- /dev/null
+++ b/configs/llm/speciation_add.yaml
@@ -0,0 +1,30 @@
+experiment: llm_speciation_add
+kind: llm_speciation
+seed: 1
+seeds: [1, 2, 3]
+n_replicates: 1
+
+# The de-confounded conflict sweep (companion to configs/llm/speciation.yaml). The "replace" design
+# holds TOTAL training fixed, so the merge's private-family decline with conflict_frac is confounded
+# with shrinking private data (coherence is the clean readout there). Here conflict_mode: add holds
+# each child's PRIVATE training fixed at n_train and ADDS conflict data on top, so any decline in the
+# MERGE's private-family accuracy relative to its parents is interference from the conflicting
+# convention, not a data-budget artefact. Pre-registered readings: merged private accuracy tracks the
+# parents at every frac -> conflict damage is localised to the conflicted function (function-specific
+# hybrid breakdown); merged private accuracy falls below the parents as frac rises -> the conflict
+# corrupts shared circuitry beyond the conflicted function (global hybrid breakdown). Run at 3 seeds:
+# the single-seed pilot showed one anomalous grid point (frac=0.75, a bad parent-B run), so per-seed
+# replication is required before reading the curve.
+
+base_model: Qwen/Qwen2.5-0.5B-Instruct
+family_a: strings
+family_b: arith
+n_train: 400
+n_test: 100
+epochs: 3
+lora: {r: 16, alpha: 32}
+conflict_mode: add
+conflict_fracs: [0.0, 0.25, 0.5, 0.75, 1.0]
+durations: []
+
+output: {dir: results/llm_speciation_add}
diff --git a/figures/plot_llm_seeds.py b/figures/plot_llm_seeds.py
new file mode 100644
index 0000000..e90b2af
--- /dev/null
+++ b/figures/plot_llm_seeds.py
@@ -0,0 +1,88 @@
+"""Multi-seed LLM robustness figure — the recombination claims with error bars.
+
+Aggregates the three multi-seed 0.5B experiments (fixed test sets, training seed varied) into one
+figure with 95% CIs over seeds:
+
+(A) llm_merge_seeds — Fisher–Muller: merged specialists vs the best single specialist, overall and
+    worst-family (5 seeds).
+(B) llm_moe_hard_seeds — union (routing) vs fusion (soup/ties) on the hard benchmark (3 seeds).
+(C) llm_directed_hard_seeds — directed offspring selection vs the a-priori soup, hard (3 seeds).
+
+Usage:  python figures/plot_llm_seeds.py
+"""
+
+from __future__ import annotations
+
+import sys
+from pathlib import Path
+
+import numpy as np
+import matplotlib.pyplot as plt
+
+sys.path.insert(0, str(Path(__file__).parent))
+from _figlib import load_bundle, savefig  # noqa: E402
+
+
+def _agg(df, models, metric):
+    """Per-model mean and 95% CI over seeds for one metric."""
+    out = []
+    for m in models:
+        v = df[(df["model"] == m) & (df["metric"] == metric)].groupby("seed")["accuracy"].mean()
+        out.append((v.mean(), 1.96 * v.std(ddof=1) / max(1, np.sqrt(len(v)))))
+    return out
+
+
+def _panel(ax, df, models, labels, title):
+    x = np.arange(len(models))
+    for off, metric, color in ((-0.17, "overall", "#2c7fb8"), (0.17, "worst_family", "#d62728")):
+        vals = _agg(df, models, metric)
+        ax.bar(x + off, [v for v, _ in vals], 0.34, yerr=[e for _, e in vals],
+               capsize=3, color=color, label=metric)
+    ax.set_xticks(x); ax.set_xticklabels(labels, fontsize=8)
+    ax.set(ylabel="verifier accuracy", ylim=(0, 1.0), title=title)
+    ax.legend(frameon=False, fontsize=8)
+
+
+def _best_spec(df):
+    """Synthesise a best-single-specialist row set per seed (max over spec_* by overall)."""
+    specs = sorted(m for m in df["model"].unique() if m.startswith("spec_"))
+    rows = []
+    for s, sub in df.groupby("seed"):
+        ov = {m: sub[(sub["model"] == m) & (sub["metric"] == "overall")]["accuracy"].mean()
+              for m in specs}
+        best = max(ov, key=ov.get)
+        b = sub[sub["model"] == best].copy()
+        b["model"] = "best_specialist"
+        rows.append(b)
+    import pandas as pd
+    return pd.concat([df] + rows, ignore_index=True)
+
+
+def main() -> None:
+    fig, axes = plt.subplots(1, 3, figsize=(16, 4.8))
+
+    df, _ = load_bundle("results/llm_merge_seeds")
+    _panel(axes[0], _best_spec(df), ["base", "best_specialist", "merge_soup", "merge_ties"],
+           ["base", "best\nspecialist", "merge\n(soup)", "merge\n(ties)"],
+           "(A) Fisher–Muller with error bars\n(5 seeds, easy benchmark, 0.5B)")
+
+    df, _ = load_bundle("results/llm_moe_hard_seeds")
+    _panel(axes[1], _best_spec(df), ["best_specialist", "merge_soup", "merge_ties", "moe_oracle",
+                                     "moe_learned"],
+           ["best\nspecialist", "fusion\n(soup)", "fusion\n(ties)", "union\n(route,oracle)",
+            "union\n(route,learned)"],
+           "(B) union vs fusion, hard benchmark\n(3 seeds, 0.5B)")
+
+    df, _ = load_bundle("results/llm_directed_hard_seeds")
+    _panel(axes[2], df, ["merge_soup", "directed_overall", "directed_balanced"],
+           ["a-priori soup", "directed\n(overall)", "directed\n(balanced)"],
+           "(C) directed offspring selection, hard\n(3 seeds, 0.5B)")
+
+    fig.suptitle("The LLM recombination claims are seed-robust (fixed test sets; training seed varied; 95% CI)",
+                 y=1.03, fontsize=12)
+    fig.tight_layout()
+    savefig(fig, "results/llm_merge_seeds", "llm_seeds")
+
+
+if __name__ == "__main__":
+    main()
diff --git a/figures/plot_llm_speciation.py b/figures/plot_llm_speciation.py
new file mode 100644
index 0000000..fe1f822
--- /dev/null
+++ b/figures/plot_llm_speciation.py
@@ -0,0 +1,96 @@
+"""LLM-tier model speciation figure — conflict coherence cliff, de-confounded interference, duration null.
+
+(A) The conflict cliff, read where it is clean: on the shared ambiguous prompts, each parent performs
+under its own convention while the 50/50 merge scores below BOTH under either grading — the hybrid
+loses precisely the conflicted function (the mu(S) floor made visible). From the "replace" design
+(results/llm_speciation).
+
+(B) The de-confounded private-family readout ("add" design, results/llm_speciation_add: private
+training held fixed, conflict data added on top): whether the merge's private-family competence
+tracks its parents (conflict damage localised to the conflicted function) or falls below them
+(interference spreading to shared circuitry).
+
+(C) The duration (emergent) null: over-trained disjoint specialists keep merging well — the merged
+model's private-family accuracy stays above the best parent at every duration. The MLP tier's
+"no emergent isolation" null generalises to LLM weights in this regime.
+
+Because LoRA deltas share the frozen base's coordinates, none of this involves alignment ambiguity:
+every failure shown is functional by construction.
+
+Usage:  python figures/plot_llm_speciation.py
+"""
+
+from __future__ import annotations
+
+import sys
+from pathlib import Path
+
+import matplotlib.pyplot as plt
+
+sys.path.insert(0, str(Path(__file__).parent))
+from _figlib import load_bundle, savefig  # noqa: E402
+
+
+def _series(df, mode, model, metric):
+    sub = df[(df["mode"] == mode) & (df["model"] == model) & (df["metric"] == metric)]
+    g = sub.groupby("x")["accuracy"].agg(["mean", "std"]).reset_index().fillna(0.0)
+    return g["x"], g["mean"], g["std"]
+
+
+def main() -> None:
+    rep, _ = load_bundle("results/llm_speciation")
+    add, _ = load_bundle("results/llm_speciation_add")
+    fam_a = "strings" if (rep["metric"] == "strings").any() else "lists"
+    fam_b = "arith"
+
+    fig, axes = plt.subplots(1, 3, figsize=(16.5, 4.9))
+
+    # (A) coherence on the conflicted function (replace design).
+    ax = axes[0]
+    x, y, _ = _series(rep, "conflict", "parent_a", "ambig_asc")
+    ax.plot(x, y, "--o", color="#9ecae1", lw=1.5, label="parent A under its convention (asc)")
+    x, y, _ = _series(rep, "conflict", "parent_b", "ambig_desc")
+    ax.plot(x, y, "--o", color="#a1d99b", lw=1.5, label="parent B under its convention (desc)")
+    x, y, _ = _series(rep, "conflict", "merge_soup", "coherence")
+    ax.plot(x, y, "-s", color="#d62728", lw=2.2, label="merge under its BEST convention")
+    ax.set(xlabel="fraction of training carrying the conflicting convention",
+           ylabel="accuracy on the shared ambiguous prompts", ylim=(-0.02, None),
+           title="(A) the hybrid loses the conflicted function\n(below BOTH parents under either grading)")
+    ax.legend(frameon=False, fontsize=8)
+
+    # (B) de-confounded private families (add design).
+    ax = axes[1]
+    mode = "conflict_add"
+    for model, color, style, lw in (("merge_soup", "#d62728", "-s", 2.2),
+                                    ("parent_a", "#9ecae1", "--o", 1.5),
+                                    ("parent_b", "#a1d99b", "--o", 1.5)):
+        x, y, s = _series(add, mode, model, "mean_private")
+        ax.plot(x, y, style, color=color, lw=lw, label=f"{model}: private families (mean)")
+        ax.fill_between(x, y - s, y + s, color=color, alpha=0.15)   # +-1 sd over seeds
+    ax.set(xlabel="conflict data added on top of fixed private training",
+           ylabel="verifier accuracy", ylim=(-0.02, 1.02),
+           title="(B) conflict damage does NOT spread: private families\ntrack the parents at every conflict level (3 seeds, ±1 sd)")
+    ax.legend(frameon=False, fontsize=8)
+
+    # (C) duration null.
+    ax = axes[2]
+    x, y, _ = _series(rep, "duration", "merge_soup", "mean_private")
+    ax.plot(x, y, "-o", color="#d62728", lw=2.2, label="merge: private families (mean)")
+    x, y, _ = _series(rep, "duration", "parent_a", fam_a)
+    ax.plot(x, y, "--o", color="#9ecae1", lw=1.5, label=f"parent A on its own family ({fam_a})")
+    x, y, _ = _series(rep, "duration", "parent_b", fam_b)
+    ax.plot(x, y, "--o", color="#a1d99b", lw=1.5, label=f"parent B on its own family ({fam_b})")
+    ax.set(xlabel="specialist training duration (epochs)", ylabel="verifier accuracy",
+           ylim=(-0.02, 1.02),
+           title="(C) the emergent test: over-specialisation\ndoes not erode mergeability here")
+    ax.legend(frameon=False, fontsize=8)
+
+    fig.suptitle("LLM-tier model speciation: conflict provokes function-specific hybrid breakdown; "
+                 "no isolation emerges from duration alone (LoRA shares base coordinates — failures are "
+                 "functional by construction)", y=1.03, fontsize=11.5)
+    fig.tight_layout()
+    savefig(fig, "results/llm_speciation", "llm_speciation")
+
+
+if __name__ == "__main__":
+    main()
diff --git a/results/llm_directed_hard_seeds/README.md b/results/llm_directed_hard_seeds/README.md
new file mode 100644
index 0000000..0102b30
--- /dev/null
+++ b/results/llm_directed_hard_seeds/README.md
@@ -0,0 +1,6 @@
+# Multi-seed directed offspring selection, hard benchmark (0.5B, 3 seeds)
+
+Part of the multi-seed firm-up; full legend and table in `results/llm_merge_seeds/README.md`
+(panel C of its `llm_seeds.png`). Headline: directed_overall 0.221 ± 0.026 beats the a-priori soup
+(0.174 ± 0.102) and the best specialist, and directed_balanced more than doubles the soup's
+worst-family (0.158 vs 0.088) — offspring selection both improves and stabilises the blend.
diff --git a/results/llm_directed_hard_seeds/manifest.json b/results/llm_directed_hard_seeds/manifest.json
new file mode 100644
index 0000000..ea42d54
--- /dev/null
+++ b/results/llm_directed_hard_seeds/manifest.json
@@ -0,0 +1,31 @@
+{
+  "experiment": "llm_directed_hard_seeds",
+  "master_seed": 1,
+  "git_commit": "d6a5c5cacde5544a39305eef34925a4fb35b8a19",
+  "python": "3.14.7",
+  "libraries": {
+    "numpy": "2.5.0",
+    "scipy": "1.18.0",
+    "pandas": "3.0.3",
+    "pyarrow": "24.0.0",
+    "torch": "2.12.1",
+    "transformers": "5.13.0",
+    "peft": "0.19.1"
+  },
+  "rows": 105,
+  "results_sha256": "532f6115402fce2a610baf7879bb81653bc61e9f64092160d3dc5d9f9d516f47",
+  "layer": "2",
+  "tier": "llm",
+  "base_model": "Qwen/Qwen2.5-0.5B-Instruct",
+  "hard": true,
+  "seeds": [
+    1,
+    2,
+    3
+  ],
+  "directed": {
+    "n_candidates": 16,
+    "concentration": 0.5,
+    "n_val": 60
+  }
+}
\ No newline at end of file
diff --git a/results/llm_directed_hard_seeds/resolved_config.yaml b/results/llm_directed_hard_seeds/resolved_config.yaml
new file mode 100644
index 0000000..9daa5d3
--- /dev/null
+++ b/results/llm_directed_hard_seeds/resolved_config.yaml
@@ -0,0 +1,29 @@
+experiment: llm_directed_hard_seeds
+seed: 1
+n_replicates: 1
+source_config:
+  experiment: llm_directed_hard_seeds
+  kind: llm_directed
+  seed: 1
+  seeds:
+  - 1
+  - 2
+  - 3
+  n_replicates: 1
+  base_model: Qwen/Qwen2.5-0.5B-Instruct
+  hard: true
+  families:
+  - lists
+  - strings
+  - arith
+  n_train: 400
+  n_test: 80
+  n_val: 60
+  n_candidates: 16
+  concentration: 0.5
+  epochs: 3
+  lora:
+    r: 16
+    alpha: 32
+  output:
+    dir: results/llm_directed_hard_seeds
diff --git a/results/llm_merge_seeds/README.md b/results/llm_merge_seeds/README.md
new file mode 100644
index 0000000..d6e8a4d
--- /dev/null
+++ b/results/llm_merge_seeds/README.md
@@ -0,0 +1,35 @@
+# Multi-seed LLM recombination (0.5B) — the claims with error bars
+
+PNAS work-order Phase 3: removes the "one seed" objection on the three LLM recombination claims.
+Protocol: **test sets fixed** (seed 1000+i per family), **training seed varied** (specialists cache
+per-seed as `spec_<family>[_hard]_s<seed>`), so across-seed variance is training variance only.
+Figure: `llm_seeds.png` (this dir) aggregates all three experiments, 95% CI over seeds.
+
+### (A) Fisher–Muller, easy benchmark, 5 seeds (`llm_merge_seeds`)
+| model | overall | worst-family |
+|---|---|---|
+| merge_ties | **0.647 ± 0.027** | **0.282 ± 0.020** |
+| merge_soup | 0.632 ± 0.042 | 0.278 ± 0.028 |
+| best specialist (strings) | 0.592 ± 0.009 | 0.078 ± 0.011 |
+| base | 0.277 | 0.150 |
+
+Both merges beat every specialist overall (ties: non-overlapping CIs; soup: marginal at 0.5B, as in
+the single-seed run — decisive at 7B) and the **worst-family signature is unambiguous**: merges ≈0.28
+vs ≤0.16 for any parent — only recombined models are competent everywhere.
+
+### (B) Union vs fusion, hard benchmark, 3 seeds (`llm_moe_hard_seeds/`)
+Routing (union) 0.274 ± 0.026 overall / 0.238 ± 0.024 worst-family; fusion soup 0.174 ± 0.102 / 0.088
+± 0.093; ties similar; best specialist 0.199 ± 0.026. Union beats fusion on both metrics — **and a new
+finding: fusion is seed-FRAGILE on hard tasks (CI ±0.10) while routing is seed-stable (±0.026).**
+Averaging's outcome depends on which specialist minima the seeds happened to find; selection-based
+recombination is reliable. (Learned router still = oracle: lexically distinct families, known rider.)
+
+### (C) Directed offspring selection, hard, 3 seeds (`llm_directed_hard_seeds/`)
+directed_overall 0.221 ± 0.026 (> soup 0.174 ± 0.102 and > best specialist); directed_balanced
+worst-family 0.158 ± 0.036 (> soup 0.088 ± 0.093). Directed selection both beats and **stabilises**
+the a-priori soup; per-input routing (B) remains above any single global blend, as before.
+
+**Read together:** all three recombination claims hold under seed replication, and the operator
+ordering (route > directed-select > soup, on headroom tasks) is not only a mean effect but a
+*variance* effect — the union/selection operators are the reliable ones. Base: Qwen2.5-0.5B-Instruct;
+statistical (per-seed) reproducibility per blueprint §4.
diff --git a/results/llm_merge_seeds/llm_seeds.pdf b/results/llm_merge_seeds/llm_seeds.pdf
new file mode 100644
index 0000000..9d2a5f1
Binary files /dev/null and b/results/llm_merge_seeds/llm_seeds.pdf differ
diff --git a/results/llm_merge_seeds/llm_seeds.png b/results/llm_merge_seeds/llm_seeds.png
new file mode 100644
index 0000000..20e5fa6
Binary files /dev/null and b/results/llm_merge_seeds/llm_seeds.png differ
diff --git a/results/llm_merge_seeds/manifest.json b/results/llm_merge_seeds/manifest.json
new file mode 100644
index 0000000..ca946db
--- /dev/null
+++ b/results/llm_merge_seeds/manifest.json
@@ -0,0 +1,28 @@
+{
+  "experiment": "llm_merge_seeds",
+  "master_seed": 1,
+  "git_commit": "d6a5c5cacde5544a39305eef34925a4fb35b8a19",
+  "python": "3.14.7",
+  "libraries": {
+    "numpy": "2.5.0",
+    "scipy": "1.18.0",
+    "pandas": "3.0.3",
+    "pyarrow": "24.0.0",
+    "torch": "2.12.1",
+    "transformers": "5.13.0",
+    "peft": "0.19.1"
+  },
+  "rows": 150,
+  "results_sha256": "810c0b27c02f40fe0aa1847b3ffb3efb2fe46631842cf411843be0bff82da2a5",
+  "layer": "2",
+  "tier": "llm",
+  "base_model": "Qwen/Qwen2.5-0.5B-Instruct",
+  "hard": false,
+  "seeds": [
+    1,
+    2,
+    3,
+    4,
+    5
+  ]
+}
\ No newline at end of file
diff --git a/results/llm_merge_seeds/resolved_config.yaml b/results/llm_merge_seeds/resolved_config.yaml
new file mode 100644
index 0000000..53c2660
--- /dev/null
+++ b/results/llm_merge_seeds/resolved_config.yaml
@@ -0,0 +1,30 @@
+experiment: llm_merge_seeds
+seed: 1
+n_replicates: 1
+source_config:
+  experiment: llm_merge_seeds
+  kind: llm_merge
+  seed: 1
+  seeds:
+  - 1
+  - 2
+  - 3
+  - 4
+  - 5
+  n_replicates: 1
+  base_model: Qwen/Qwen2.5-0.5B-Instruct
+  families:
+  - lists
+  - strings
+  - arith
+  n_train: 600
+  n_test: 100
+  epochs: 3
+  lora:
+    r: 16
+    alpha: 32
+  merges:
+  - soup
+  - ties
+  output:
+    dir: results/llm_merge_seeds
diff --git a/results/llm_moe_hard_seeds/README.md b/results/llm_moe_hard_seeds/README.md
new file mode 100644
index 0000000..b222667
--- /dev/null
+++ b/results/llm_moe_hard_seeds/README.md
@@ -0,0 +1,7 @@
+# Multi-seed union-vs-fusion, hard benchmark (0.5B, 3 seeds)
+
+Part of the multi-seed firm-up; full legend, table, and the fusion-fragility finding in
+`results/llm_merge_seeds/README.md` (panel B of its `llm_seeds.png`). Headline: union/routing
+0.274 ± 0.026 > fusion 0.174 ± 0.102 overall (worst-family 0.238 vs 0.088), and fusion's ±0.10 CI vs
+routing's ±0.026 is itself the finding — averaging is seed-fragile where headroom exists; routing is
+reliable.
diff --git a/results/llm_moe_hard_seeds/manifest.json b/results/llm_moe_hard_seeds/manifest.json
new file mode 100644
index 0000000..ef0208f
--- /dev/null
+++ b/results/llm_moe_hard_seeds/manifest.json
@@ -0,0 +1,32 @@
+{
+  "experiment": "llm_moe_hard_seeds",
+  "master_seed": 1,
+  "git_commit": "d6a5c5cacde5544a39305eef34925a4fb35b8a19",
+  "python": "3.14.7",
+  "libraries": {
+    "numpy": "2.5.0",
+    "scipy": "1.18.0",
+    "pandas": "3.0.3",
+    "pyarrow": "24.0.0",
+    "torch": "2.12.1",
+    "transformers": "5.13.0",
+    "peft": "0.19.1"
+  },
+  "rows": 126,
+  "results_sha256": "3d30a97ecb2e05a99cb188f527e353b737f3935b2e4a9b8c01fba1edec197b9b",
+  "layer": "2",
+  "tier": "llm",
+  "base_model": "Qwen/Qwen2.5-0.5B-Instruct",
+  "hard": true,
+  "seeds": [
+    1,
+    2,
+    3
+  ],
+  "operators": [
+    "soup",
+    "ties",
+    "moe_oracle",
+    "moe_learned"
+  ]
+}
\ No newline at end of file
diff --git a/results/llm_moe_hard_seeds/resolved_config.yaml b/results/llm_moe_hard_seeds/resolved_config.yaml
new file mode 100644
index 0000000..0f3d6d4
--- /dev/null
+++ b/results/llm_moe_hard_seeds/resolved_config.yaml
@@ -0,0 +1,32 @@
+experiment: llm_moe_hard_seeds
+seed: 1
+n_replicates: 1
+source_config:
+  experiment: llm_moe_hard_seeds
+  kind: llm_moe
+  seed: 1
+  seeds:
+  - 1
+  - 2
+  - 3
+  n_replicates: 1
+  base_model: Qwen/Qwen2.5-0.5B-Instruct
+  hard: true
+  families:
+  - lists
+  - strings
+  - arith
+  n_train: 400
+  n_test: 80
+  n_route: 32
+  epochs: 3
+  lora:
+    r: 16
+    alpha: 32
+  operators:
+  - soup
+  - ties
+  - moe_oracle
+  - moe_learned
+  output:
+    dir: results/llm_moe_hard_seeds
diff --git a/results/llm_speciation/README.md b/results/llm_speciation/README.md
new file mode 100644
index 0000000..4edcaed
--- /dev/null
+++ b/results/llm_speciation/README.md
@@ -0,0 +1,47 @@
+# LLM-tier model speciation — conflict provokes isolation; duration alone does not
+
+E13 carried into language-model weights (0.5B Qwen, LoRA children of one frozen base — which shares
+its coordinate system with both children, so **there is no permutation/rescaling ambiguity by
+construction: every merge failure here is functional**). Two knobs, pre-registered readings in the
+configs; figure `llm_speciation.png` (3 panels; panel B from `results/llm_speciation_add/`).
+
+**Design.** Child A: private family `strings`; child B: private family `arith`; shared **ambiguous
+convention prompts** ("Sort the list […]" — no direction stated) answered *ascending* by A and
+*descending* by B: each convention harmless alone, contradictory jointly (the Bateson–Dobzhansky–
+Muller structure). 50/50 soup merge; exact-match verifier; fixed test sets. **Coherence** of a model =
+max(accuracy under asc grading, under desc grading) on the shared prompts — a coherent parent scores
+under its own convention; a hybrid mixing conventions scores low under both (the `μ(S)/2` floor made
+operational).
+
+### Finding 1 — function-specific hybrid breakdown (the conflict knob; panel A)
+Once the conventions are trained (`conflict_frac ≥ 0.25`), each parent performs under its own
+convention (~0.20–0.25 — low absolute values: sorting long lists is hard for 0.5B) while the merge's
+coherence sits at **0.02–0.08, below BOTH parents under either grading** — the hybrid loses precisely
+the conflicted function. (At `frac = 0` no one has seen the ambiguous prompts; that point is a
+no-conflict baseline, not part of the cliff.)
+
+### Finding 2 — the damage does not spread (the de-confounded `add` design; panel B)
+In the original ("replace") sweep, higher conflict fraction mechanically means *less private-family
+training*, so the merge's private-family decline is confounded. The **`add` design**
+(`llm_speciation_add`, 3 seeds) holds each child's private training fixed and adds conflict data on
+top: the merge's private-family accuracy then shows **no trend with conflict** (0.74–0.88, tracking
+parent A's 0.82–0.87 within seed noise at every level). Conflict damage is **localised to the
+conflicted function**; it does not corrupt the disjoint skills — at this scale, hybrid breakdown is
+surgical, not global. Honest rider: 0.5B soup merges carry large *intrinsic* seed variance even at
+zero conflict (sd up to 0.28) — the same averaging-fragility seen in `llm_moe_hard_seeds`.
+
+### Finding 3 — the duration null: over-specialisation does not erode mergeability (panel C)
+Pure disjoint specialists over-trained from 1 to 12 epochs (no shared data at all): the merged model
+*improves* (0.84 → 0.94 mean-private) and stays **above the best parent at every duration**. The MLP
+tier's "no emergent isolation" null (`speciation_real_emergent`) **generalises to LLM weights** in
+this regime — relevant to the report that averaging prefers under-trained experts (arXiv:2607.11997):
+in our disjoint-family setting, no such over-training penalty appears; the theory's prediction is that
+their effect should trace to *conflicting conventions on shared circuitry*, which the
+`epistasis_predicts` experiment (work order) will test directly.
+
+**Speciation across all three tiers now reads:** analytic (E12: cliff, epistasis-dependence,
+snowball) → MLP (E13: functional residual survives the full symmetry group; no emergent isolation) →
+LLM (this run: function-specific hybrid breakdown under conflict; no isolation from duration or
+specialisation alone). Isolation must be provoked by functional conflict at every tier tested.
+Falsifiers (not triggered): merge coherence matching the parents (no breakdown), or merged
+private-family accuracy declining with conflict in the `add` design (global corruption).
diff --git a/results/llm_speciation/llm_speciation.pdf b/results/llm_speciation/llm_speciation.pdf
new file mode 100644
index 0000000..b4844a9
Binary files /dev/null and b/results/llm_speciation/llm_speciation.pdf differ
diff --git a/results/llm_speciation/llm_speciation.png b/results/llm_speciation/llm_speciation.png
new file mode 100644
index 0000000..b7448f5
Binary files /dev/null and b/results/llm_speciation/llm_speciation.png differ
diff --git a/results/llm_speciation/manifest.json b/results/llm_speciation/manifest.json
new file mode 100644
index 0000000..142622c
--- /dev/null
+++ b/results/llm_speciation/manifest.json
@@ -0,0 +1,21 @@
+{
+  "experiment": "llm_speciation",
+  "master_seed": 1,
+  "git_commit": "d6a5c5cacde5544a39305eef34925a4fb35b8a19",
+  "python": "3.14.7",
+  "libraries": {
+    "numpy": "2.5.0",
+    "scipy": "1.18.0",
+    "pandas": "3.0.3",
+    "pyarrow": "24.0.0",
+    "torch": "2.12.1",
+    "transformers": "5.13.0",
+    "peft": "0.19.1"
+  },
+  "rows": 162,
+  "results_sha256": "5490874f6d7157db38d395baa1a8174315a32f67d53679c22f8621c454ead0ea",
+  "layer": "2",
+  "tier": "llm",
+  "base_model": "Qwen/Qwen2.5-0.5B-Instruct",
+  "hard": false
+}
\ No newline at end of file
diff --git a/results/llm_speciation/resolved_config.yaml b/results/llm_speciation/resolved_config.yaml
new file mode 100644
index 0000000..42b9e6f
--- /dev/null
+++ b/results/llm_speciation/resolved_config.yaml
@@ -0,0 +1,30 @@
+experiment: llm_speciation
+seed: 1
+n_replicates: 1
+source_config:
+  experiment: llm_speciation
+  kind: llm_speciation
+  seed: 1
+  n_replicates: 1
+  base_model: Qwen/Qwen2.5-0.5B-Instruct
+  family_a: strings
+  family_b: arith
+  n_train: 400
+  n_test: 100
+  epochs: 3
+  lora:
+    r: 16
+    alpha: 32
+  conflict_fracs:
+  - 0.0
+  - 0.25
+  - 0.5
+  - 0.75
+  - 1.0
+  durations:
+  - 1
+  - 3
+  - 6
+  - 12
+  output:
+    dir: results/llm_speciation
diff --git a/results/llm_speciation_add/README.md b/results/llm_speciation_add/README.md
new file mode 100644
index 0000000..0ca9c61
--- /dev/null
+++ b/results/llm_speciation_add/README.md
@@ -0,0 +1,9 @@
+# LLM speciation, de-confounded conflict sweep ("add" design, 3 seeds)
+
+Companion to `results/llm_speciation/` (full legend there; this run feeds panel B of its figure).
+Private training held fixed at n_train while conflict data is ADDED on top, so any decline in the
+merge's private-family accuracy would be interference, not a data-budget artefact. Result: no trend —
+the merge tracks parent A within seed noise at every conflict level (pre-registered reading #1:
+conflict damage is localised to the conflicted function). Run at 3 seeds because the single-seed pilot
+contained one anomalous grid point (a bad parent-B training run); the seeded curve is the reportable
+one.
diff --git a/results/llm_speciation_add/manifest.json b/results/llm_speciation_add/manifest.json
new file mode 100644
index 0000000..584fd2e
--- /dev/null
+++ b/results/llm_speciation_add/manifest.json
@@ -0,0 +1,26 @@
+{
+  "experiment": "llm_speciation_add",
+  "master_seed": 1,
+  "git_commit": "58e6c74609ce12142a5f1ae542c80c2be8713937",
+  "python": "3.14.7",
+  "libraries": {
+    "numpy": "2.5.0",
+    "scipy": "1.18.0",
+    "pandas": "3.0.3",
+    "pyarrow": "24.0.0",
+    "torch": "2.12.1",
+    "transformers": "5.13.0",
+    "peft": "0.19.1"
+  },
+  "rows": 270,
+  "results_sha256": "d08d0979eb4715127268d6a524c1be57d123eeb6c39f8293dd68ef4894139265",
+  "layer": "2",
+  "tier": "llm",
+  "base_model": "Qwen/Qwen2.5-0.5B-Instruct",
+  "hard": false,
+  "seeds": [
+    1,
+    2,
+    3
+  ]
+}
\ No newline at end of file
diff --git a/results/llm_speciation_add/resolved_config.yaml b/results/llm_speciation_add/resolved_config.yaml
new file mode 100644
index 0000000..05c6df0
--- /dev/null
+++ b/results/llm_speciation_add/resolved_config.yaml
@@ -0,0 +1,31 @@
+experiment: llm_speciation_add
+seed: 1
+n_replicates: 1
+source_config:
+  experiment: llm_speciation_add
+  kind: llm_speciation
+  seed: 1
+  seeds:
+  - 1
+  - 2
+  - 3
+  n_replicates: 1
+  base_model: Qwen/Qwen2.5-0.5B-Instruct
+  family_a: strings
+  family_b: arith
+  n_train: 400
+  n_test: 100
+  epochs: 3
+  lora:
+    r: 16
+    alpha: 32
+  conflict_mode: add
+  conflict_fracs:
+  - 0.0
+  - 0.25
+  - 0.5
+  - 0.75
+  - 1.0
+  durations: []
+  output:
+    dir: results/llm_speciation_add
diff --git a/src/llm/experiment.py b/src/llm/experiment.py
index 20ccf06..4cdb8b6 100644
--- a/src/llm/experiment.py
+++ b/src/llm/experiment.py
@@ -63,10 +63,10 @@ def run_merge_experiment(cfg: dict) -> pd.DataFrame:
     rows += _rows(name, "base", "base", evaluate(m, tok, test))
     del m; torch.cuda.empty_cache()
 
-    # one specialist per family
+    # one specialist per family (cache is seed-specific: multi-seed runs retrain per seed)
     dirs = []
     for i, f in enumerate(fams):
-        d = str(adapters_root / f"spec_{f}{suffix}")
+        d = str(adapters_root / f"spec_{f}{suffix}_s{seed}")
         train_specialist(base, f, d, n_train=n_train, epochs=epochs, seed=seed + i, hard=hard,
                          r=int(lora.get("r", 16)), alpha=int(lora.get("alpha", 32)))
         dirs.append(d)
@@ -103,7 +103,7 @@ def _load_or_train_specialists(cfg: dict, base: str, fams: list[str], name: str,
     adapters_root = Path(cfg.get("adapters_dir", "models/llm"))
     dirs: list[str] = []
     for i, f in enumerate(fams):
-        d = str(adapters_root / f"spec_{f}{suffix}")
+        d = str(adapters_root / f"spec_{f}{suffix}_s{seed}")
         if not (Path(d) / "adapter_config.json").exists():   # reuse across llm_merge / llm_moe runs
             train_specialist(base, f, d, n_train=n_train, epochs=epochs, seed=seed + i, hard=hard,
                              r=int(lora.get("r", 16)), alpha=int(lora.get("alpha", 32)))
@@ -258,12 +258,22 @@ def run_directed_experiment(cfg: dict) -> pd.DataFrame:
     return pd.DataFrame(rows)
 
 
+def run_speciation_dispatch(cfg: dict) -> pd.DataFrame:
+    from .speciation import run_speciation_experiment      # local import: torch-heavy
+    return run_speciation_experiment(cfg)
+
+
 _RUNNERS = {"llm_merge": run_merge_experiment, "llm_moe": run_moe_experiment,
-            "llm_directed": run_directed_experiment}
+            "llm_directed": run_directed_experiment, "llm_speciation": run_speciation_dispatch}
 
 
 def run_and_save(config_path: str | Path) -> Path:
-    """Load an LLM experiment YAML, run it (dispatch on ``kind``), and write the artifact triple."""
+    """Load an LLM experiment YAML, run it (dispatch on ``kind``), and write the artifact triple.
+
+    A ``seeds: [..]`` list runs the experiment once per seed (specialist caches are seed-specific)
+    and concatenates the frames with a ``seed`` column — the Layer-2 statistical-reproducibility
+    pattern (fixed test sets, training seed varies).
+    """
     config_path = Path(config_path)
     cfg = yaml.safe_load(config_path.read_text())
     out_dir = Path(cfg.get("output", {}).get("dir", f"results/{cfg['experiment']}"))
@@ -271,9 +281,21 @@ def run_and_save(config_path: str | Path) -> Path:
     kind = cfg.get("kind", "llm_merge")
     if kind not in _RUNNERS:
         raise ValueError(f"unknown LLM experiment kind {kind!r} (expected one of {list(_RUNNERS)})")
-    df = _RUNNERS[kind](cfg)
+    seeds = cfg.get("seeds")
+    if seeds:
+        frames = []
+        for s in seeds:
+            run_cfg = dict(cfg); run_cfg["seed"] = int(s)
+            f = _RUNNERS[kind](run_cfg); f["seed"] = int(s)
+            frames.append(f)
+        df = pd.concat(frames, ignore_index=True)
+        cfg["seed"] = int(seeds[0])                        # manifest master seed = first of the list
+    else:
+        df = _RUNNERS[kind](cfg)
     extra = {"layer": "2", "tier": "llm", "base_model": cfg["base_model"],
              "hard": bool(cfg.get("hard", False))}
+    if seeds:
+        extra["seeds"] = [int(s) for s in seeds]
     if kind == "llm_moe":
         extra["operators"] = list(cfg.get("operators", []))
     if kind == "llm_directed":
diff --git a/src/llm/speciation.py b/src/llm/speciation.py
new file mode 100644
index 0000000..1b41fa0
--- /dev/null
+++ b/src/llm/speciation.py
@@ -0,0 +1,150 @@
+"""LLM-tier model speciation (E13 in language-model weights) — the conflict cliff + the duration null.
+
+The real-LLM image of E13, with a structural bonus: LoRA deltas live in the frozen base's coordinate
+system, so there is **no permutation/rescaling ambiguity by construction** — any merge failure here is
+*functional* incompatibility, the residual isolated architecturally (no alignment step needed).
+
+Two knobs, mirroring the MLP experiment:
+
+* **Conflict (imposed, the cliff).** Two LoRA children from the same base. Each has a private,
+  disjoint skill family (A: ``strings``, B: ``arith`` — so the merge has genuine Fisher–Muller value)
+  plus a shared set of **ambiguous convention prompts** ("Sort the list [...]" with no direction),
+  which child A learns to answer *ascending* and child B *descending* — each convention harmless
+  alone, contradictory jointly (a true Bateson–Dobzhansky–Muller structure). ``conflict_frac`` sweeps
+  the fraction of each child's training data that is convention data. Merged 50/50 (soup), the
+  prediction is E13's cliff in verifier units: private-family competence of the *merge* degrades and
+  convention coherence collapses as conflict grows, while each *parent* stays fine — hybrid
+  breakdown, not parent damage.
+* **Duration (emergent, the null test).** Pure disjoint specialists (zero shared data), over-trained
+  by sweeping epochs. The MLP tier found *no* emergent isolation (the merge rescued specialists at
+  every divergence); the empirical merging literature reports averaging prefers *under*-trained
+  experts (arXiv:2607.11997). This sweep arbitrates: if the merged model's quality falls with
+  duration while each parent's own-family quality does not, that is emergent incompatibility at the
+  LLM tier; if not, the MLP null generalises.
+
+Convention coherence of a model = max(accuracy under ascending grading, accuracy under descending
+grading) on the shared ambiguous prompts: a coherent parent scores high under its own convention; a
+hybrid that mixes conventions scores low under both (the mu(S)/2 floor made operational).
+"""
+
+from __future__ import annotations
+
+from pathlib import Path
+
+import numpy as np
+import pandas as pd
+
+from .evaluate import generate, load_model
+from .merge import load_specialists, make_merge
+from .specialise import train_lora_on_tasks
+from .tasks import Task, _fmt_list, make_tasks, verify
+
+
+def make_convention_tasks(n: int, seed: int, convention: str) -> list[Task]:
+    """Ambiguous sort prompts with a convention-dependent canonical answer.
+
+    The prompt never states a direction ("Sort the list [...]"), so *either* convention is a
+    self-consistent, harmless resolution — the conflict exists only between lineages.
+
+    Args:
+        n (int): number of tasks.
+        seed (int): prompts are a pure function of the seed (same seed -> same prompts, so the two
+            conventions grade the *same* inputs).
+        convention (str): ``asc`` or ``desc``.
+
+    Returns:
+        list[Task]: family ``"ambig"``; answers sorted per the convention.
+    """
+    rng = np.random.default_rng(seed)
+    out = []
+    for _ in range(n):
+        xs = rng.integers(0, 30, size=int(rng.integers(6, 10))).tolist()
+        ans = sorted(xs, reverse=(convention == "desc"))
+        out.append(Task("ambig", f"Sort the list {_fmt_list(xs)}. "
+                        "Output only the resulting list and nothing else.", _fmt_list(ans)))
+    return out
+
+
+def _acc(model, tok, tasks: list[Task]) -> float:
+    outs = generate(model, tok, [t.prompt for t in tasks])
+    return float(np.mean([verify(o, t) for o, t in zip(outs, tasks)]))
+
+
+def run_speciation_experiment(cfg: dict) -> pd.DataFrame:
+    """Run the conflict-cliff and/or duration sweeps; return long-form accuracies.
+
+    Config keys: ``base_model``, ``family_a``/``family_b`` (private families), ``n_train``,
+    ``n_test``, ``epochs`` (conflict mode), ``conflict_fracs`` (list), ``durations`` (list of epoch
+    counts), ``lora``, ``seed``, ``adapters_dir``.
+    """
+    import torch
+
+    name = cfg["experiment"]
+    base = cfg["base_model"]
+    fam_a, fam_b = cfg.get("family_a", "strings"), cfg.get("family_b", "arith")
+    n_train, n_test = int(cfg.get("n_train", 400)), int(cfg.get("n_test", 100))
+    epochs = int(cfg.get("epochs", 3))
+    lora = cfg.get("lora", {})
+    r, alpha = int(lora.get("r", 16)), int(lora.get("alpha", 32))
+    seed = int(cfg["seed"])
+    hard = bool(cfg.get("hard", False))
+    root = Path(cfg.get("adapters_dir", "models/llm")) / "speciation"
+
+    # Fixed evaluation sets (identical across the sweep; convention pairs grade the SAME prompts).
+    test_a = make_tasks(fam_a, n_test, seed=1000, hard=hard)
+    test_b = make_tasks(fam_b, n_test, seed=1001, hard=hard)
+    amb_asc = make_convention_tasks(n_test, seed=5000, convention="asc")
+    amb_desc = make_convention_tasks(n_test, seed=5000, convention="desc")
+
+    def measure(model, tok, label: str, mode: str, x: float, rows: list[dict]) -> None:
+        accs = {fam_a: _acc(model, tok, test_a), fam_b: _acc(model, tok, test_b),
+                "ambig_asc": _acc(model, tok, amb_asc), "ambig_desc": _acc(model, tok, amb_desc)}
+        accs["coherence"] = max(accs["ambig_asc"], accs["ambig_desc"])
+        accs["mean_private"] = (accs[fam_a] + accs[fam_b]) / 2.0
+        for k, v in accs.items():
+            rows.append({"experiment": name, "mode": mode, "x": float(x),
+                         "model": label, "metric": k, "accuracy": v})
+
+    def train_child(tasks: list, out_dir: Path, ep: int) -> str:
+        return train_lora_on_tasks(base, tasks, str(out_dir), epochs=ep, r=r, alpha=alpha,
+                                   seed=seed, batch_size=int(cfg.get("batch_size", 8)))
+
+    def merged_and_parents(dir_a: str, dir_b: str, mode: str, x: float, rows: list[dict]) -> None:
+        for d, label in ((dir_a, "parent_a"), (dir_b, "parent_b")):
+            m, tok = load_model(base, adapter_dir=d)
+            measure(m, tok, label, mode, x, rows)
+            del m; torch.cuda.empty_cache()
+        model, tok = load_specialists(base, [dir_a, dir_b])
+        make_merge(model, 2, "soup", "soup")
+        measure(model, tok, "merge_soup", mode, x, rows)
+        del model; torch.cuda.empty_cache()
+
+    rows: list[dict] = []
+
+    # Two conflict designs. "replace" (default) holds the TOTAL training budget fixed, so the
+    # private-family readout is confounded with shrinking private data (coherence is the clean metric
+    # there). "add" holds the PRIVATE budget fixed and adds conflict data on top, so any decline in the
+    # merge's private-family accuracy is interference, not a data-budget artefact.
+    conflict_mode = str(cfg.get("conflict_mode", "replace"))
+    for frac in cfg.get("conflict_fracs", []):
+        frac = float(frac)
+        n_conv = int(round(frac * n_train))
+        n_own = n_train if conflict_mode == "add" else n_train - n_conv
+        tasks_a = (make_tasks(fam_a, n_own, seed=seed, hard=hard)
+                   + make_convention_tasks(n_conv, seed=seed + 50, convention="asc"))
+        tasks_b = (make_tasks(fam_b, n_own, seed=seed + 1, hard=hard)
+                   + make_convention_tasks(n_conv, seed=seed + 50, convention="desc"))
+        da = train_child(tasks_a, root / "conflict_a", epochs)
+        db = train_child(tasks_b, root / "conflict_b", epochs)
+        merged_and_parents(da, db, f"conflict_{conflict_mode}"
+                           if conflict_mode != "replace" else "conflict", frac, rows)
+
+    for dur in cfg.get("durations", []):
+        dur = int(dur)
+        tasks_a = make_tasks(fam_a, n_train, seed=seed, hard=hard)
+        tasks_b = make_tasks(fam_b, n_train, seed=seed + 1, hard=hard)
+        da = train_child(tasks_a, root / "dur_a", dur)
+        db = train_child(tasks_b, root / "dur_b", dur)
+        merged_and_parents(da, db, "duration", dur, rows)
+
+    return pd.DataFrame(rows)
diff --git a/tasks/workorder-pnas-submission.md b/tasks/workorder-pnas-submission.md
index a76f886..eba8f3c 100644
--- a/tasks/workorder-pnas-submission.md
+++ b/tasks/workorder-pnas-submission.md
@@ -62,7 +62,9 @@ barriers). Patch before the preprint goes up.
     (no growth ⇒ "models are safer to merge than biology predicts" — an honest bound).
   - Tests + README + fold into figure. Pre-register the falsifier language before running.
 
-## Phase 2 — arXiv preprint (priority stake) — end of Week 1
+## Phase 2 — arXiv preprint package (GG decision 2026-08-11: DO NOT POST until all experiments
+and analysis are complete — the preprint goes up after Phase 3, with the final analysis folded in;
+re-run md2tex + tectonic at that point)
 
 - [x] Citation refresh in `paper/the-evolution-of-sex-for-ai.md` *(done 2026-08-11; author names verified against arXiv API)*: **new concessions** —
   First-Extinction Law (2509.20101), quantitative-trait collapse (2407.17493), verifier-injection
@@ -74,12 +76,18 @@ barriers). Patch before the preprint goes up.
 - [x] Fold E13b/c results into the speciation section (whatever they show — honestly). *(Done: full-symmetry residual + hybrid-fitness cliff + the emergent converse, in abstract, §5, §13 and the accessible version.)*
 - [x] LaTeX conversion: `paper/arxiv/` (md2tex.py block-based converter from the Markdown source of
   truth; main.tex; 3 figures; builds clean under tectonic, 20 pp; arXiv pdflatex hint guarded).
-- [ ] **Post to arXiv (GG action — package ready):** upload `paper/arxiv/{main.tex,body.tex,figs/}`;
-  categories, license note, and a ≤1,920-char abstract are prepared in `paper/arxiv/ARXIV-SUBMISSION.md`.
+- [ ] **Post to arXiv — DEFERRED until all experiments/analysis are done (GG).** Package is ready
+  (`paper/arxiv/`, instructions in `ARXIV-SUBMISSION.md`); rebuild after the Phase-3 results are
+  folded into the manuscript, then upload.
 
 ## Phase 3 — New experiments for impact & robustness — Weeks 2–3
 
-- [ ] **`llm_speciation` — the cliff at the LLM tier (new, highest-impact addition).**
+- [x] **`llm_speciation` — the cliff at the LLM tier.** *(Run 2026-08-11, 0.5B: DURATION NULL —
+  over-trained disjoint specialists merge better not worse (0.84->0.94, above best parent throughout);
+  the MLP "no emergent isolation" null generalises. CONFLICT — function-specific hybrid breakdown:
+  merged coherence 0.02-0.08 falls below BOTH parents (~0.2) on the conflicted function. Caught a
+  design confound (replace mode ties conflict_frac to private-data budget) -> built the de-confounded
+  `add` variant (conflict_mode: add; configs/llm/speciation_add.yaml). 7B confirm optional later.)*
   Speciation demonstrated at all three tiers (analytic → MLP → LLM) makes the headline
   unassailable. Structural bonus to state in the paper: LoRA deltas live in the frozen base's
   coordinate system, so there is **no permutation ambiguity by construction** — any LoRA-merge
@@ -93,12 +101,46 @@ barriers). Patch before the preprint goes up.
     (2607.11997: optimal expert duration) as *our theory explaining their data* — a strong PNAS
     move.
   - 0.5B locally first; one 7B CX3 confirm if the sign is clean (`hpc/` PBS, minutes).
-- [ ] **Multi-seed the LLM arc (robustness — removes the "one seed" objection).**
+- [x] **Multi-seed the LLM arc (0.5B tier done 2026-08-11).** All three claims hold with CIs
+  (merge>every specialist; union 0.274±0.026 > fusion 0.174±0.102 hard; directed 0.221±0.026 > soup)
+  + NEW finding: fusion is seed-FRAGILE on hard tasks (±0.10) while routing/directed are stable
+  (±0.026) — the union/selection operators are the reliable ones. results/llm_*_seeds/ + llm_seeds
+  figure. Remaining: 7B CX3 seeds (1-3) when HPC convenient.
   - Thread the seed into specialist cache keys (`spec_<family>[_hard]_s<seed>`); verify nothing
     else assumes the old names.
   - 0.5B: seeds 1–5 × {merge, moe, directed} × {easy, hard}. 7B on CX3: seeds 1–3 × hard
     {merge, moe, directed} (8–25 min walltimes → trivial). Aggregate figures with 95% CI; update
     READMEs; the headroom law now carries error bars.
+- [ ] **`epistasis_predicts` — the DECISIVE experiment (from the external review, 2026-08-11; highest
+  priority after llm_speciation lands).** The review's exact bar: population-genetic quantities must
+  *predict* (not re-describe) — forecast merge success **before merging**, and beat existing
+  predictors. Design, reusing the llm_speciation machinery:
+  1. Parents with independently controlled interaction structure: sweep `conflict_frac` (ground-truth
+     epistasis) *and* compatible/disjoint + duration variants (spread in divergence WITHOUT conflict),
+     so functional conflict and divergence are decorrelated by construction.
+  2. Pre-merge predictors, none of which touches a merged model: (a) **operational epistasis** =
+     functional-disagreement mass between the parents on a shared probe set (the μ(S) estimate — ours);
+     (b) **gradient alignment** (the Zhou et al. 2601.22285 predictor); (c) **weight-space geometry**
+     (LoRA-delta cosine / norm distance).
+  3. Outcome: merged (soup) performance on private families + convention coherence, held-out test,
+     multi-seed.
+  4. The claim to test: at matched geometric divergence, the epistasis measure predicts merge outcome
+     and the geometry measures do not (R² comparison + an operator-choice decision test — merge vs
+     route — under matched budgets).
+  Pre-registered falsifier: if gradient/geometry predictors match the epistasis measure, the paper's
+  "epistasis, not divergence, sets the cliff" claim stays analytic-only and is labelled as such.
+- [x] **Manuscript claim-narrowing (external review, 2026-08-11) — done.** Softened identity claims
+  (WF exact only in the minimal model + learning-kernel cited against ourselves; ratchet scoped to the
+  irreversible arm), removed "nobody has / none imports / theory outrun" (priority-dispute bait),
+  added the interpretation/explanation/prediction ladder to §1, stated the merge-don't-average
+  operator boundaries (output-mean vs weight-avg vs routing vs max-with-oracle, budgets, oracle,
+  capacity), added a "what these experiments do and do not establish" scope block to the speciation
+  section (impossibility floor is information-theoretic, not genetic; snowball/epistasis-cliff =
+  hypotheses at the neural tier), replaced "control theory" with "framework" (subtitle included —
+  GG can veto), fixed the §3/§11 overstatements (frozen core ≠ frozen behaviour; Baldwin = echo not
+  identity; archive vs operational irreversibility), added the **claims-at-a-glance table**
+  (status/assumptions/evidence/limits) to §13 + table support in md2tex, and matched the calibration
+  in the accessible version. Adopted the review's framing sentence as the stated core contribution.
 - [ ] **(Optional, SI-grade) ambiguous-families router stress test** — overlapping-skill families
   where the router is no longer trivially perfect; documents the honest limit of union-by-routing.
   Do only if time permits; otherwise keep the existing rider sentence.
diff --git a/tests/test_llm.py b/tests/test_llm.py
index acf5c75..e6dce39 100644
--- a/tests/test_llm.py
+++ b/tests/test_llm.py
@@ -107,3 +107,22 @@ def test_merge_weights_requires_two_candidates():
     import pytest
     with pytest.raises(ValueError):
         sample_merge_weights(3, 1, np.random.default_rng(0))
+
+
+def test_convention_tasks_conflict_only_between_conventions():
+    # The BDM structure of llm_speciation: identical prompts, each convention internally consistent
+    # and verifiable, the two conventions contradictory on (almost) every prompt.
+    from llm.speciation import make_convention_tasks
+    from llm.tasks import verify
+
+    asc = make_convention_tasks(20, seed=5, convention="asc")
+    desc = make_convention_tasks(20, seed=5, convention="desc")
+    assert [a.prompt for a in asc] == [d.prompt for d in desc]      # same inputs, graded two ways
+    assert all(verify(a.answer, a) for a in asc)                    # each convention self-consistent
+    assert all(verify(d.answer, d) for d in desc)
+    conflicting = sum(a.answer != d.answer for a, d in zip(asc, desc))
+    assert conflicting >= 18                                        # contradictory unless already sorted
+    assert all(not verify(a.answer, d) for a, d in zip(asc, desc) if a.answer != d.answer)
+    # deterministic: prompts and answers are a pure function of (seed, convention)
+    again = make_convention_tasks(20, seed=5, convention="asc")
+    assert [t.answer for t in again] == [t.answer for t in asc]