epistasis_predicts: functional conflict, not weight geometry, predicts merge failure pre-merge

The decisive experiment from the external review. 39 LoRA parent pairs
(0.5B, 3 seeds) on three axes decorrelated by construction: conflict
(contradictory conventions on shared prompts, private budgets fixed),
compat (same prompts, SAME convention — overlap without conflict), and
duration (weight divergence, zero conflict). Six pre-merge predictors;
primary outcome = merge penalty (parent potential − merged achieved).

League table (Spearman vs penalty, n=39): functional measures predict
(dis_raw +0.460, epi_conf +0.446, p<0.005); geometry collapses
(delta_cos +0.03, delta_l2 +0.17 n.s.); gradient alignment weak (−0.35);
performance ~0. The first grid's apparent geometry win (+0.60) was an
overlap/volume artifact — the compat control axis (added for exactly
this) exposed and killed it: same overlap and data volume, zero penalty.
Honest riders in the README: confidence weighting does not beat raw
disagreement as a rank predictor (pre-registered internal prediction not
confirmed; it does double the conflict/compat level contrast), and
|rho|~0.45 is bounded by 0.5B merge-outcome noise (7B is the firm-up).

Also: micro-batched gradient accumulation (OOM fix on the shared 16GB
GPU), exact r-space LoRA-delta geometry (brute-force-verified test,
151 green), systemd-run runbook lesson (tmux dies with the SSH session
scope on this box).

Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01BkRLcc18rwT2Lysu6PbG7v

configs/llm/epistasis.yaml

@@ -0,0 +1,43 @@
+experiment: llm_epistasis
+kind: llm_epistasis
+seed: 1
+seeds: [1, 2, 3]
+n_replicates: 1
+
+# THE DECISIVE EXPERIMENT (external review, 2026-08-11; PNAS work order Phase 3): do
+# population-genetic quantities PREDICT merge success BEFORE merging, beyond existing predictors?
+#
+# Grid: functional conflict (conflict_fracs, add-design: private budget fixed, contradictory
+# convention data added) and weight divergence (durations, zero conflict) are DECORRELATED BY
+# CONSTRUCTION. Pre-merge predictors, none touching a merged model:
+#   epi_conf   - operational epistasis: confidence-weighted bilateral disagreement on a broad probe
+#                mix (ambiguous + both private families, drawn blind to where conflict lives). OURS.
+#   dis_raw    - raw disagreement rate (the internal ablation: our theory PREDICTS this mispredicts,
+#                because it counts harmless complementation - one parent ignorant - as conflict).
+#   grad_cos   - gradient alignment at the shared base (the ML-literature predictor, cf. 2601.22285).
+#   delta_cos / delta_l2 - LoRA-delta weight geometry (computed exactly in r-space).
+#   cross_perf - performance-based predictor (cross-family accuracy).
+# Outcome (PRIMARY, pre-registered): merge_penalty = parent_potential - merged_overall (the
+# hybrid-load analogue: what the merge loses relative to what its parents could jointly deliver);
+# secondary: merged_overall, and the soup-vs-route gap (route_private - merged_private).
+#
+# PRE-REGISTERED READINGS. Success for the framework: |Spearman rho(epi_conf, merge_penalty)| high,
+# AND geometry/gradient predictors fail at matched divergence (their apparent correlation, if any,
+# collapses within the conflict axis where divergence is near-constant), AND epi_conf > dis_raw
+# (the confidence-weighting prediction). FALSIFIER: gradient/geometry/raw-disagreement match or beat
+# epi_conf -> the "epistasis, not divergence, sets the cliff" claim stays analytic-only and the paper
+# says so. 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: 80
+epochs: 3
+n_probe_each: 30
+grad_k: 32
+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_epistasis}

configs/llm/epistasis_compat.yaml

@@ -0,0 +1,29 @@
+experiment: llm_epistasis_compat
+kind: llm_epistasis
+seed: 1
+seeds: [1, 2, 3]
+n_replicates: 1
+
+# The missing CONTROL axis for the decisive experiment (identified from the first grid's results,
+# 2026-08-11): in the original grid every shared-data pair was a CONFLICTED pair, so the delta-cosine
+# geometry predictor could succeed as a mere task-OVERLAP detector (overlap coincided with conflict by
+# construction). This sweep adds overlap WITHOUT conflict: both children train on the SAME ambiguous
+# prompts with the SAME convention (asc/asc), private budgets fixed, at the same fractions as the
+# conflict sweep. Pre-registered readings: if delta_cos stays high here while merge_penalty stays ~0,
+# geometry was detecting overlap, not incompatibility, and its apparent predictive power collapses
+# once compat pairs enter the pool; functional measures (epi_conf / dis_raw) should correctly stay LOW
+# here (the parents AGREE on the shared prompts). If geometry still predicts across all three axes, the
+# falsifier stands as stated in configs/llm/epistasis.yaml.
+
+base_model: Qwen/Qwen2.5-0.5B-Instruct
+family_a: strings
+family_b: arith
+n_train: 400
+n_test: 80
+epochs: 3
+n_probe_each: 30
+grad_k: 32
+lora: {r: 16, alpha: 32}
+compat_fracs: [0.25, 0.5, 0.75, 1.0]
+
+output: {dir: results/llm_epistasis_compat}

figures/plot_llm_epistasis.py

@@ -0,0 +1,87 @@
+"""The decisive-experiment figure — does pre-merge epistasis predict merge failure?
+
+(A) The theory's predictor: operational epistasis (confidence-weighted bilateral disagreement,
+measured before merging) against the merge penalty (parent potential − merged achieved, the
+hybrid-load analogue). Conflict-axis pairs in red, duration-axis pairs in blue.
+
+(B) The geometry predictor on the same outcome: weight divergence (LoRA-delta L2) — the
+matched-divergence contrast: the duration axis spans large weight divergence at ~zero penalty, while
+the conflict axis generates penalty at modest divergence. Distance is not what breaks merging.
+
+(C) The league table: |Spearman rho| against the merge penalty for every pre-merge predictor,
+including the internal ablation (raw disagreement, which the theory predicts must mislead because it
+counts harmless complementation as conflict).
+
+Usage:  python figures/plot_llm_epistasis.py
+"""
+
+from __future__ import annotations
+
+import sys
+from pathlib import Path
+
+import matplotlib.pyplot as plt
+import numpy as np
+from scipy.stats import spearmanr
+
+sys.path.insert(0, str(Path(__file__).parent))
+from _figlib import load_bundle, savefig  # noqa: E402
+
+PREDICTORS = [("epi_conf", "operational\nepistasis"),
+              ("dis_raw", "raw\ndisagreement"),
+              ("grad_cos", "gradient\nalignment"),
+              ("delta_cos", "delta\ncosine"),
+              ("delta_l2", "delta\nL2"),
+              ("cross_perf", "cross-family\naccuracy")]
+OUTCOME = "merge_penalty"
+
+
+def _scatter(ax, df, xcol, xlabel, title):
+    for mode, color, marker in (("conflict", "#d62728", "o"), ("duration", "#2c7fb8", "s"),
+                                ("compat", "#41ab5d", "^")):
+        sub = df[df["mode"] == mode]
+        ax.scatter(sub[xcol], sub[OUTCOME], c=color, marker=marker, s=42, alpha=0.75,
+                   label=f"{mode} axis")
+    rho, p = spearmanr(df[xcol], df[OUTCOME])
+    ax.set(xlabel=xlabel, ylabel="merge penalty (parent potential − merged)",
+           title=f"{title}\nSpearman ρ = {rho:.2f} (p = {p:.1g}, n = {len(df)})")
+    ax.axhline(0, color="#999", lw=0.6)
+    ax.legend(frameon=False, fontsize=8)
+
+
+def main() -> None:
+    import pandas as pd
+    df, _ = load_bundle("results/llm_epistasis")
+    try:                                             # the overlap-without-conflict control axis
+        compat, _ = load_bundle("results/llm_epistasis_compat")
+        df = pd.concat([df, compat], ignore_index=True)
+    except Exception:
+        pass
+
+    fig, axes = plt.subplots(1, 3, figsize=(16.5, 4.9))
+    _scatter(axes[0], df, "epi_conf", "operational epistasis (pre-merge)",
+             "(A) the theory's predictor")
+    _scatter(axes[1], df, "delta_cos", "LoRA-delta cosine similarity (pre-merge)",
+             "(B) the geometry predictor — does it detect\nincompatibility, or just task overlap?")
+
+    ax = axes[2]
+    rhos, labels = [], []
+    for col, label in PREDICTORS:
+        rho, _ = spearmanr(df[col], df[OUTCOME])
+        rhos.append(abs(rho)); labels.append(label)
+    colors = ["#d62728" if c == "epi_conf" else ("#fc9272" if c == "dis_raw" else "#9ecae1")
+              for c, _ in PREDICTORS]
+    x = np.arange(len(rhos))
+    ax.bar(x, rhos, 0.62, color=colors)
+    ax.set_xticks(x); ax.set_xticklabels(labels, fontsize=7)
+    ax.set(ylabel="|Spearman ρ| vs merge penalty", ylim=(0, 1),
+           title="(C) predictor league table (pre-merge only)")
+
+    fig.suptitle("Predicting merge failure BEFORE merging: functional conflict, not weight divergence "
+                 "(conflict, overlap-without-conflict, and divergence decorrelated by construction; 3 seeds)", y=1.03, fontsize=12)
+    fig.tight_layout()
+    savefig(fig, "results/llm_epistasis", "llm_epistasis")
+
+
+if __name__ == "__main__":
+    main()

results/llm_epistasis/README.md

@@ -0,0 +1,52 @@
+# The decisive experiment — predicting merge failure BEFORE merging
+
+The external review's bar (2026-08-11): population-genetic quantities must *predict*, not
+re-describe — forecast merge success **pre-merge** and beat existing predictors. Design: 39 parent
+pairs (0.5B LoRA children of one frozen base, 3 seeds) on **three axes decorrelated by
+construction** — `conflict` (contradictory conventions on shared ambiguous prompts, private budgets
+fixed), `compat` (the control: **same shared prompts, same convention** — task overlap *without*
+conflict; added after the first grid exposed a confound, see below), and `duration` (weight
+divergence with zero conflict, 1→12 epochs). Primary outcome (pre-registered):
+**merge penalty** = parent potential − merged achieved (the hybrid-load analogue). Figure:
+`llm_epistasis.png`.
+
+### The league table (Spearman ρ vs merge penalty, full three-axis pool, n = 39)
+| pre-merge predictor | ρ | p | reading |
+|---|---|---|---|
+| raw functional disagreement (`dis_raw`) | **+0.460** | 0.003 | predicts |
+| operational epistasis (`epi_conf`, confidence-weighted) | **+0.446** | 0.004 | predicts |
+| gradient alignment at the base (cf. 2601.22285) | −0.347 | 0.03 | weakly informative |
+| LoRA-delta L2 distance (geometry) | +0.165 | 0.32 | uninformative |
+| LoRA-delta cosine (geometry) | +0.030 | 0.86 | uninformative |
+| cross-family accuracy (performance) | −0.005 | 0.98 | uninformative |
+
+**Headline: functional conflict, measured before merging, predicts merge failure; weight geometry
+does not.** The duration axis spans the same weight-divergence range as the conflict axis
+(L2 ≈ 2.4–4.0) at ~zero penalty, and the compat axis adds the same *data volumes* and overlap at
+~zero penalty — so both geometric predictors collapse once overlap and volume are controlled.
+
+### The control that did the work (`llm_epistasis_compat/`)
+In the first grid (conflict + duration only), `delta_cos` scored ρ = +0.60 — apparently the best
+predictor. That was an **artifact**: every shared-data pair in that pool was a conflicted pair, so
+geometry could win as a mere task-overlap/volume detector. The `compat` axis (overlap without
+conflict) exposes it: penalty ≈ 0.005 there, and the geometry correlations collapse (+0.60 → +0.03).
+The functional measures behave correctly on the control — parents trained on the same convention
+*agree* on the shared prompts (epi_conf: 0.46 conflict vs 0.23 compat, a 2× contrast; raw
+disagreement 0.72 vs 0.47, only 1.5× — the confidence weighting removes complementation noise from
+the *level*, giving the cleaner axis separation).
+
+### Honest riders (pre-registered falsifier status)
+1. The internal prediction that confidence weighting would beat raw disagreement **as a rank
+   predictor is not confirmed**: `epi_conf` and `dis_raw` are statistically indistinguishable at
+   n = 39 (the weighting does improve the conflict-vs-compat *contrast* in levels). The paper reports
+   the functional-vs-geometric verdict, not a win for the refinement.
+2. Correlations are moderate (|ρ| ≈ 0.45), bounded by 0.5B merge-outcome noise (soup merges carry
+   large intrinsic seed variance — see `llm_moe_hard_seeds`); read as signs and ordering, not
+   magnitudes. 7B replication is the natural firm-up.
+3. Gradient alignment carries real signal (it differentiates conflicting conventions at the base) but
+   less than the functional measures in this design.
+
+**Bottom line for the paper:** the framework's claim — *epistasis (functional conflict), not
+divergence, sets merge compatibility* — survives its designed falsification test at this tier: the
+operational conflict measures predict, the divergence measures do not, and the case was made honest
+by a control that first *broke our own experiment's* favourite-looking geometric predictor.

results/llm_epistasis/manifest.json

@@ -0,0 +1,26 @@
+{
+  "experiment": "llm_epistasis",
+  "master_seed": 1,
+  "git_commit": "5a23ddaf2a906a14d9aeb2797cf8fef821a519f4",
+  "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": 27,
+  "results_sha256": "2895adb3347c4550fa1bc052a32e86aa17f4891f1163a80ea118618a911e4bd5",
+  "layer": "2",
+  "tier": "llm",
+  "base_model": "Qwen/Qwen2.5-0.5B-Instruct",
+  "hard": false,
+  "seeds": [
+    1,
+    2,
+    3
+  ]
+}

results/llm_epistasis/resolved_config.yaml

@@ -0,0 +1,36 @@
+experiment: llm_epistasis
+seed: 1
+n_replicates: 1
+source_config:
+  experiment: llm_epistasis
+  kind: llm_epistasis
+  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: 80
+  epochs: 3
+  n_probe_each: 30
+  grad_k: 32
+  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_epistasis

results/llm_epistasis_compat/README.md

@@ -0,0 +1,8 @@
+# The overlap-without-conflict control axis
+
+Companion to `results/llm_epistasis/` (full analysis and league table there). Both children train on
+the SAME ambiguous prompts with the SAME convention (asc/asc) — task overlap and added-data volume
+identical to the conflict axis, with zero conflict. Result: merge penalty ≈ 0.005, geometry
+predictors' apparent power collapses (delta_cos ρ +0.60 → +0.03 once these pairs enter the pool),
+functional measures correctly stay low. The control that separated "detects incompatibility" from
+"detects overlap".

results/llm_epistasis_compat/manifest.json

@@ -0,0 +1,26 @@
+{
+  "experiment": "llm_epistasis_compat",
+  "master_seed": 1,
+  "git_commit": "5a23ddaf2a906a14d9aeb2797cf8fef821a519f4",
+  "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": 12,
+  "results_sha256": "e2091aebfdfdeaa1d44c3ca8b9a60d2dfd13bfe546d811f19ce8eb1ccfc14e7b",
+  "layer": "2",
+  "tier": "llm",
+  "base_model": "Qwen/Qwen2.5-0.5B-Instruct",
+  "hard": false,
+  "seeds": [
+    1,
+    2,
+    3
+  ]
+}

results/llm_epistasis_compat/resolved_config.yaml

@@ -0,0 +1,30 @@
+experiment: llm_epistasis_compat
+seed: 1
+n_replicates: 1
+source_config:
+  experiment: llm_epistasis_compat
+  kind: llm_epistasis
+  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: 80
+  epochs: 3
+  n_probe_each: 30
+  grad_k: 32
+  lora:
+    r: 16
+    alpha: 32
+  compat_fracs:
+  - 0.25
+  - 0.5
+  - 0.75
+  - 1.0
+  output:
+    dir: results/llm_epistasis_compat

src/llm/epistasis.py

@@ -0,0 +1,251 @@
+"""The decisive experiment: does an operational EPISTASIS measure predict merge success PRE-merge?
+
+The external review's bar (2026-08-11): population-genetic quantities must *predict*, not re-describe —
+forecast merge success **before merging**, and beat existing predictors. This module builds a grid of
+parent pairs in which functional conflict and weight divergence are **decorrelated by construction**
+(the conflict knob adds contradictory-convention data at fixed private budget; the duration knob grows
+divergence with zero conflict), computes a battery of **pre-merge predictors**, then merges and
+measures the outcome.
+
+The theory-derived predictor — and its built-in ablation. A raw functional-disagreement rate between
+two parents is, by our own theory, the WRONG measure: disjoint specialists disagree hugely (one knows,
+one is ignorant) yet merge perfectly — harmless *complementation*. The Dobzhansky–Muller analogue is
+**bilateral confident contradiction**: both parents confidently produce different answers to the same
+input. So the operational epistasis measure is confidence-weighted disagreement,
+
+    epi_conf = E_probe[ conf_A · conf_B · 1(ans_A != ans_B) ],   conf = exp(mean token logprob),
+
+and the theory makes an internal, falsifiable prediction: ``epi_conf`` should predict merge failure
+where raw disagreement (``dis_raw``) mispredicts. Baseline predictors from the ML literature:
+**gradient alignment** at the shared base (cf. arXiv:2601.22285), **LoRA-delta geometry** (cosine and
+L2 distance, computed exactly in r×r space), and a **performance-based** predictor (cross-family
+accuracy). Primary outcome (pre-registered): ``merge_penalty`` = parent-potential − merged-achieved,
+the hybrid-load analogue. Pre-registered falsifier: if gradient/geometry predictors match or beat
+``epi_conf``, the paper's "epistasis, not divergence, sets the cliff" claim stays analytic-only.
+"""
+
+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 _encode, train_lora_on_tasks
+from .speciation import make_convention_tasks
+from .tasks import _normalise, make_tasks, verify
+
+
+# ---------------------------------------------------------------- pre-merge predictors
+
+def generate_with_confidence(model, tok, prompts, max_new_tokens=24, batch_size=32, device="cuda"):
+    """Greedy batched generation returning (answers, confidences).
+
+    Confidence = exp(mean logprob of the generated tokens up to the first EOS) — a per-answer
+    self-certainty in [0, 1], measurable without any ground truth.
+    """
+    import torch
+
+    outs, confs = [], []
+    for i in range(0, len(prompts), batch_size):
+        chunk = prompts[i:i + batch_size]
+        msgs = [[{"role": "user", "content": p}] for p in chunk]
+        enc = tok.apply_chat_template(msgs, add_generation_prompt=True, return_tensors="pt",
+                                      return_dict=True, padding=True).to(device)
+        with torch.no_grad():
+            gen = model.generate(**enc, max_new_tokens=max_new_tokens, do_sample=False,
+                                 pad_token_id=tok.pad_token_id, output_scores=True,
+                                 return_dict_in_generate=True)
+        L = enc["input_ids"].shape[1]
+        seq = gen.sequences
+        for j in range(len(chunk)):
+            toks = seq[j, L:]
+            lps = []
+            for t, tid in enumerate(toks):
+                if int(tid) in (tok.eos_token_id, tok.pad_token_id):
+                    break
+                logp = torch.log_softmax(gen.scores[t][j].float(), dim=-1)[int(tid)]
+                lps.append(float(logp))
+            outs.append(tok.decode(toks, skip_special_tokens=True).strip())
+            confs.append(float(np.exp(np.mean(lps))) if lps else 0.0)
+    return outs, np.array(confs)
+
+
+def lora_delta_inner(A1, B1, A2, B2) -> float:
+    """Frobenius inner product <B1@A1, B2@A2> computed in r×r space (never materialises the deltas)."""
+    import torch
+    return float(torch.trace((B1.T @ B2) @ (A2 @ A1.T)))
+
+
+def delta_geometry(dir_a: str, dir_b: str) -> dict:
+    """Cosine similarity and L2 distance between two LoRA adapters' weight deltas (CPU, exact)."""
+    import torch
+    from safetensors.torch import load_file
+
+    def pairs(d):
+        sd = load_file(str(Path(d) / "adapter_model.safetensors"))
+        mods = {}
+        for k, v in sd.items():
+            if "lora_A" in k:
+                mods.setdefault(k.replace("lora_A", "@"), {})["A"] = v.float()
+            elif "lora_B" in k:
+                mods.setdefault(k.replace("lora_B", "@"), {})["B"] = v.float()
+        return mods
+
+    ma, mb = pairs(dir_a), pairs(dir_b)
+    keys = sorted(set(ma) & set(mb))
+    i12 = sum(lora_delta_inner(ma[k]["A"], ma[k]["B"], mb[k]["A"], mb[k]["B"]) for k in keys)
+    i11 = sum(lora_delta_inner(ma[k]["A"], ma[k]["B"], ma[k]["A"], ma[k]["B"]) for k in keys)
+    i22 = sum(lora_delta_inner(mb[k]["A"], mb[k]["B"], mb[k]["A"], mb[k]["B"]) for k in keys)
+    n1, n2 = np.sqrt(max(i11, 1e-12)), np.sqrt(max(i22, 1e-12))
+    return {"delta_cos": i12 / (n1 * n2), "delta_l2": float(np.sqrt(max(i11 + i22 - 2 * i12, 0.0)))}
+
+
+def gradient_alignment(base_name: str, tasks_a: list, tasks_b: list, k: int = 32,
+                       device: str = "cuda", seed: int = 0) -> float:
+    """Cosine between the two tasks' SFT gradients at the SHARED BASE (no adapters) — the
+    gradient-alignment predictor of the ML literature, computed on ``k`` examples per side."""
+    import torch
+    from transformers import AutoModelForCausalLM, AutoTokenizer
+
+    torch.cuda.empty_cache()                                # release the eval models' cache first
+    tok = AutoTokenizer.from_pretrained(base_name)
+    if tok.pad_token_id is None:
+        tok.pad_token = tok.eos_token
+    model = AutoModelForCausalLM.from_pretrained(base_name, dtype=torch.bfloat16).to(device)
+    targets = [p for n, p in model.named_parameters()
+               if any(t in n for t in ("q_proj", "k_proj", "v_proj", "o_proj",
+                                       "gate_proj", "up_proj", "down_proj"))]
+    rng = np.random.default_rng(seed)
+
+    def grad_vec(tasks, micro=4):
+        # Micro-batched gradient accumulation: identical total gradient (losses weighted to a mean
+        # over the k examples), 1/8 the activation memory of one big backward — OOM-safe on a shared
+        # 16 GB GPU.
+        model.zero_grad(set_to_none=True)
+        sel = [tasks[i] for i in rng.choice(len(tasks), size=min(k, len(tasks)), replace=False)]
+        for s in range(0, len(sel), micro):
+            enc = [_encode(tok, t, device) for t in sel[s:s + micro]]
+            L = max(len(ids) for ids, _ in enc)
+            ii = torch.full((len(enc), L), tok.pad_token_id, dtype=torch.long)
+            ll = torch.full((len(enc), L), -100, dtype=torch.long)
+            am = torch.zeros((len(enc), L), dtype=torch.long)
+            for b, (ids, lab) in enumerate(enc):
+                ii[b, :len(ids)] = torch.tensor(ids); ll[b, :len(lab)] = torch.tensor(lab)
+                am[b, :len(ids)] = 1
+            out = model(input_ids=ii.to(device), attention_mask=am.to(device), labels=ll.to(device))
+            (out.loss * (len(enc) / len(sel))).backward()
+        return torch.cat([p.grad.detach().float().flatten().cpu() for p in targets])
+
+    ga, gb = grad_vec(tasks_a), grad_vec(tasks_b)
+    cos = float((ga @ gb) / (ga.norm() * gb.norm() + 1e-12))
+    del model
+    torch.cuda.empty_cache()
+    return cos
+
+
+# ---------------------------------------------------------------- the experiment
+
+def run_epistasis_experiment(cfg: dict) -> pd.DataFrame:
+    """Grid of parent pairs (conflict × duration axes) → pre-merge predictors + merged outcomes."""
+    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", 80))
+    epochs = int(cfg.get("epochs", 3))
+    n_probe = int(cfg.get("n_probe_each", 30))
+    grad_k = int(cfg.get("grad_k", 32))
+    lora = cfg.get("lora", {})
+    r, alpha = int(lora.get("r", 16)), int(lora.get("alpha", 32))
+    seed = int(cfg["seed"])
+    root = Path(cfg.get("adapters_dir", "models/llm")) / "epistasis"
+
+    # Fixed evaluation + probe sets (identical across the grid and across seeds).
+    test_a = make_tasks(fam_a, n_test, seed=1000)
+    test_b = make_tasks(fam_b, n_test, seed=1001)
+    amb_asc = make_convention_tasks(n_test, seed=5000, convention="asc")
+    amb_desc = make_convention_tasks(n_test, seed=5000, convention="desc")
+    # Probes: a broad mix drawn WITHOUT knowledge of where the conflict lives (ambiguous + both
+    # private families), so the disagreement measures are not handed the answer.
+    probe_prompts = ([t.prompt for t in make_convention_tasks(n_probe, seed=7000, convention="asc")]
+                     + [t.prompt for t in make_tasks(fam_a, n_probe, seed=7001)]
+                     + [t.prompt for t in make_tasks(fam_b, n_probe, seed=7002)])
+
+    def acc(model, tok, tasks):
+        outs = generate(model, tok, [t.prompt for t in tasks])
+        return float(np.mean([verify(o, t) for o, t in zip(outs, tasks)]))
+
+    grid = ([("conflict", float(f)) for f in cfg.get("conflict_fracs", [])]
+            + [("compat", float(f)) for f in cfg.get("compat_fracs", [])]
+            + [("duration", int(d)) for d in cfg.get("durations", [])])
+
+    rows: list[dict] = []
+    for mode, x in grid:
+        if mode in ("conflict", "compat"):                  # add-design: private budget fixed
+            # "compat" is the control that separates functional from geometric predictors: the SAME
+            # shared ambiguous prompts, but BOTH children learn the SAME convention — task overlap
+            # (which inflates delta similarity) without any conflict. A geometry predictor that is
+            # really an overlap detector mispredicts here; a functional-conflict measure does not.
+            conv_b = "desc" if mode == "conflict" else "asc"
+            n_conv = int(round(x * n_train))
+            tasks_a = (make_tasks(fam_a, n_train, seed=seed)
+                       + make_convention_tasks(n_conv, seed=seed + 50, convention="asc"))
+            tasks_b = (make_tasks(fam_b, n_train, seed=seed + 1)
+                       + make_convention_tasks(n_conv, seed=seed + 50, convention=conv_b))
+            ep = epochs
+        else:                                               # duration: pure private, epochs swept
+            tasks_a = make_tasks(fam_a, n_train, seed=seed)
+            tasks_b = make_tasks(fam_b, n_train, seed=seed + 1)
+            ep = int(x)
+        da = train_lora_on_tasks(base, tasks_a, str(root / "a"), epochs=ep, r=r, alpha=alpha,
+                                 seed=seed)
+        db = train_lora_on_tasks(base, tasks_b, str(root / "b"), epochs=ep, r=r, alpha=alpha,
+                                 seed=seed)
+
+        # --- pre-merge: per-parent probe answers/confidences + parent outcomes ------------------
+        par = {}
+        for d, label in ((da, "a"), (db, "b")):
+            m, tok = load_model(base, adapter_dir=d)
+            ans, conf = generate_with_confidence(m, tok, probe_prompts)
+            par[label] = {
+                "ans": [_normalise(o) for o in ans], "conf": conf,
+                "fam_a": acc(m, tok, test_a), "fam_b": acc(m, tok, test_b),
+                "coh": max(acc(m, tok, amb_asc), acc(m, tok, amb_desc)),
+            }
+            del m; torch.cuda.empty_cache()
+        neq = np.array([a != b for a, b in zip(par["a"]["ans"], par["b"]["ans"])], dtype=float)
+        dis_raw = float(neq.mean())
+        epi_conf = float(np.mean(par["a"]["conf"] * par["b"]["conf"] * neq))
+        geo = delta_geometry(da, db)
+        grad_cos = gradient_alignment(base, tasks_a, tasks_b, k=grad_k, seed=seed)
+        cross_perf = float(np.mean([par["a"]["fam_b"], par["b"]["fam_a"]]))
+
+        # --- merge and measure -------------------------------------------------------------------
+        model, tok = load_specialists(base, [da, db])
+        make_merge(model, 2, "soup", "soup")
+        mo = {"fam_a": acc(model, tok, test_a), "fam_b": acc(model, tok, test_b),
+              "coh": max(acc(model, tok, amb_asc), acc(model, tok, amb_desc))}
+        del model; torch.cuda.empty_cache()
+
+        potential = float(np.mean([par["a"]["fam_a"], par["b"]["fam_b"],
+                                   max(par["a"]["coh"], par["b"]["coh"])]))
+        achieved = float(np.mean([mo["fam_a"], mo["fam_b"], mo["coh"]]))
+        rows.append({
+            "experiment": name, "mode": mode, "x": float(x),
+            "epi_conf": epi_conf, "dis_raw": dis_raw, "grad_cos": grad_cos,
+            "delta_cos": geo["delta_cos"], "delta_l2": geo["delta_l2"], "cross_perf": cross_perf,
+            "merge_fam_a": mo["fam_a"], "merge_fam_b": mo["fam_b"], "merge_coh": mo["coh"],
+            "merged_private": float(np.mean([mo["fam_a"], mo["fam_b"]])),
+            "merged_overall": achieved,
+            "parent_potential": potential,
+            "merge_penalty": potential - achieved,
+            "route_private": float(np.mean([par["a"]["fam_a"], par["b"]["fam_b"]])),
+            "pa_fam_a": par["a"]["fam_a"], "pa_fam_b": par["a"]["fam_b"],
+            "pb_fam_a": par["b"]["fam_a"], "pb_fam_b": par["b"]["fam_b"],
+            "pa_coh": par["a"]["coh"], "pb_coh": par["b"]["coh"],
+        })
+    return pd.DataFrame(rows)

src/llm/experiment.py

@@ -263,8 +263,14 @@ def run_speciation_dispatch(cfg: dict) -> pd.DataFrame:
     return run_speciation_experiment(cfg)
 
 
+def run_epistasis_dispatch(cfg: dict) -> pd.DataFrame:
+    from .epistasis import run_epistasis_experiment        # local import: torch-heavy
+    return run_epistasis_experiment(cfg)
+
+
 _RUNNERS = {"llm_merge": run_merge_experiment, "llm_moe": run_moe_experiment,
-            "llm_directed": run_directed_experiment, "llm_speciation": run_speciation_dispatch}
+            "llm_directed": run_directed_experiment, "llm_speciation": run_speciation_dispatch,
+            "llm_epistasis": run_epistasis_dispatch}
 
 
 def run_and_save(config_path: str | Path) -> Path:

tasks/lessons.md

@@ -13,3 +13,9 @@ is about (output-mean vs weight-average vs max-with-oracle vs routing are differ
 different budgets); (4) don't write "nobody has / none imports" — invite no priority disputes; say
 "to our knowledge" and state the positive contribution; (5) negative results (E13b) are strengths —
 lead with them; (6) "control theory" needs states/controls/dynamics/rule or it's a "framework".
+
+**2026-08-11 — tmux does NOT survive SSH disconnects on this machine.** The tmux server starts inside
+the SSH session's systemd scope and gets reaped on logout (lost ~20 min of the epistasis grid; GG:
+"SSH disconnected"). Reliable pattern here: `systemd-run --user --collect --unit=<name>
+--working-directory="$PWD" bash -c '<cmd>'` — lands in user@.service (kept alive by the desktop
+session), survives disconnects; check with `systemctl --user is-active <name>`, logs via redirect.

tasks/workorder-pnas-submission.md

@@ -111,8 +111,15 @@ re-run md2tex + tectonic at that point)
   - 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
+- [x] **`epistasis_predicts` — the DECISIVE experiment — DONE (2026-08-11, 0.5B, 39 pairs, 3 seeds).**
+  *Verdict: functional conflict measured pre-merge PREDICTS merge failure (dis_raw rho=+0.46,
+  epi_conf +0.45, both p<0.005); weight geometry does NOT (delta_cos +0.03, delta_l2 +0.17 n.s.);
+  gradient alignment weakly informative (-0.35). The first grid's apparent geometry win (+0.60) was an
+  overlap artifact, exposed and killed by the added `compat` control axis (same overlap+volume, no
+  conflict, zero penalty). Honest rider: confidence weighting did NOT beat raw disagreement as a rank
+  predictor (internal prediction not confirmed; it does give a 2x vs 1.5x conflict/compat contrast in
+  levels). |rho|~0.45 bounded by 0.5B merge noise — 7B replication is the firm-up.
+  results/llm_epistasis{,_compat}/ + figure.* 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

tests/test_llm.py

@@ -126,3 +126,15 @@ def test_convention_tasks_conflict_only_between_conventions():
     # 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]
+
+
+def test_lora_delta_inner_matches_brute_force():
+    # The r-space Frobenius inner product <B1@A1, B2@A2> must equal the materialised computation.
+    import torch
+    from llm.epistasis import lora_delta_inner
+
+    g = torch.Generator().manual_seed(0)
+    A1, B1 = torch.randn(4, 20, generator=g), torch.randn(12, 4, generator=g)
+    A2, B2 = torch.randn(4, 20, generator=g), torch.randn(12, 4, generator=g)
+    brute = float(((B1 @ A1) * (B2 @ A2)).sum())
+    assert abs(lora_delta_inner(A1, B1, A2, B2) - brute) < 1e-3