Layer 1.5: architecture-general neural existence proof

Re-scopes Layer 2 into a cheaper, architecture-general neural collapse proof
before the LLM rung. Realises the same Wright–Fisher abstractions in real trained
generative models on a fully-synthetic sandbox with an exact oracle, reusing
knowledge.metrics/truth/seeding and the output contract so neural curves overlay
the Layer-1 analytic curves.

  - src/neural/: synthetic token-grammar sandbox (lossless identity + stochastic
    style), ExactOracle, HistogramModel bridge, generation loop, experiment runner
  - HARD GATE passed: histogram lineage reproduces Layer 1 exactly (neutral decay,
    exact H_eq, tracks run_lineage) — tests/test_neural_validation.py
  - torch models: autoregressive RNN + MLP (VAE implemented, not yet fidelity-
    passing); determinism seeding derived from the SeedSequence stream
  - N0 bridge (neural g*=0.047 ≈ Layer-1 0.048), N1 collapse-in-weights, N2 phase
    boundary, N5 architecture-generality (collapse + grounding-rescue in histogram
    + RNN + MLP). Manifests/configs committed; parquet gitignored, hashes tracked
  - additive backward-compatible save_artifacts extension; Makefile neural targets

Finding: neural smoothing partially resists H-collapse, so forward-KL and tail
survival are the sharp neural collapse metrics (H is smooth, per Layer 1).

92 tests green. Remaining (tasks/todo.md): N4 merge, N2 refine, N3/N6, VAE
fidelity, MNIST tier, figures. LLM/LoRA rung and C3 deferred.

Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>

src/knowledge/experiment.py

@@ -211,11 +211,28 @@ def _content_hash(path: Path) -> str:
     return h.hexdigest()
 
 
-def save_artifacts(cfg: dict, df: pd.DataFrame, out_dir: Path) -> None:
+def save_artifacts(cfg: dict, df: pd.DataFrame, out_dir: Path,
+                   extra_libs: tuple[str, ...] = (),
+                   extra_manifest: dict | None = None,
+                   grid: list | None = None) -> None:
     """Write the reproducibility output contract (blueprint 2.7 / 4).
 
     Writes ``results.parquet``, ``resolved_config.yaml`` (the fully-expanded config), and
     ``manifest.json`` (library versions, master seed, git commit, content hash).
+
+    Args:
+        cfg (dict): The parsed experiment config.
+        df (pd.DataFrame): The long-form results.
+        out_dir (Path): Output directory.
+        extra_libs (tuple[str, ...]): Extra library names to record versions for (e.g.
+            ``torch``, ``torchvision`` for Layer 1.5). Missing libraries are skipped, so a
+            caller can pass optional deps unconditionally.
+        extra_manifest (dict | None): Extra key/value pairs to merge into the manifest
+            (e.g. model architecture, oracle checkpoint hash, determinism flags).
+        grid (list | None): Pre-expanded ``[{label, lineage_cfg}, ...]`` to record in the
+            resolved config. If None, it is computed via ``expand_sweeps`` for the Layer-1
+            ``lineage`` kind (a caller with a different schema, e.g. Layer 1.5, passes its
+            own expanded grid here).
     """
     out_dir.mkdir(parents=True, exist_ok=True)
     results_path = out_dir / "results.parquet"
@@ -227,24 +244,32 @@ def save_artifacts(cfg: dict, df: pd.DataFrame, out_dir: Path) -> None:
         "n_replicates": cfg["n_replicates"],
         "source_config": cfg,
     }
-    if cfg.get("kind", "lineage") == "lineage":
+    if grid is not None:
+        resolved["grid"] = grid
+    elif cfg.get("kind", "lineage") == "lineage":
         resolved["grid"] = [
             {"label": label, "lineage_cfg": lineage_cfg}
             for label, lineage_cfg in expand_sweeps(cfg)
         ]
     (out_dir / "resolved_config.yaml").write_text(yaml.safe_dump(resolved, sort_keys=False))
 
+    libraries: dict[str, str] = {}
+    for lib in ("numpy", "scipy", "pandas", "pyarrow") + tuple(extra_libs):
+        try:
+            libraries[lib] = version(lib)
+        except Exception:  # optional dep not installed -> omit rather than crash
+            pass
     manifest = {
         "experiment": cfg["experiment"],
         "master_seed": cfg["seed"],
         "git_commit": _git_commit(),
         "python": sys.version.split()[0],
-        "libraries": {
-            lib: version(lib) for lib in ("numpy", "scipy", "pandas", "pyarrow")
-        },
+        "libraries": libraries,
         "rows": int(len(df)),
         "results_sha256": _content_hash(results_path),
     }
+    if extra_manifest:
+        manifest.update(extra_manifest)
     (out_dir / "manifest.json").write_text(json.dumps(manifest, indent=2))
 
 

src/neural/__init__.py

@@ -0,0 +1,13 @@
+"""Layer 1.5 — the architecture-general neural existence proof.
+
+Realises the Layer-1 (``knowledge``) Wright-Fisher abstractions in *real trained
+generative models* on a fully-synthetic sandbox whose ground-truth ``p*`` is known
+exactly. A model's knowledge is measured as its output distribution over ``K`` discrete
+*modes* (via an oracle), so the same metrics (``knowledge.metrics``), the same closed
+forms, and the same experiments carry over — a neural collapse curve can be overlaid on
+a Layer-1 analytic curve.
+
+The package is staged by cost: the histogram model (pure NumPy) reduces this layer
+*exactly* to Layer 1 and is the validation bridge; the RNN/VAE/MLP models (torch, added
+from Stage C) show that collapse is architecture-general.
+"""

src/neural/config.py

@@ -0,0 +1,124 @@
+"""Resolved run configuration for a neural (Layer 1.5) lineage.
+
+Mirrors the ``knowledge.config`` idiom exactly: frozen dataclasses with a ``from_dict``
+that fills defaults and rejects unknown keys via ``knowledge.config._sub``. The grounding,
+re-mint, and metrics blocks are *reused verbatim* from ``knowledge.config`` so the neural
+runner speaks the same schema as Layer 1 (grounding ``m``, the ``g -> m`` conversion, the
+re-mint gate, and the KL/support floors are all identical). Only the data source
+(``synthetic``) and the model (``model``) are neural-specific.
+"""
+
+from __future__ import annotations
+
+import math
+from dataclasses import dataclass, field, replace
+from typing import Any, Mapping
+
+from knowledge.config import GroundingCfg, MetricsCfg, RemintCfg, _sub
+
+
+@dataclass(frozen=True)
+class SyntheticCfg:
+    """The fully-synthetic mode-truth and observation grammar.
+
+    The first seven fields are the Layer-1 ``TruthCfg`` knobs (they build ``p*`` over the
+    ``K`` modes via ``knowledge.truth.make_true_distribution``). The remaining fields
+    define how a mode is rendered to a categorical token sequence: an *identity* segment
+    that encodes the mode losslessly (read by the exact oracle) followed by a *style*
+    segment of within-mode stochastic tokens (so a real generative model has a
+    distribution to learn, not just a lookup table).
+    """
+
+    K: int
+    R: int = 1
+    tail: str = "zipf"
+    zipf_s: float = 1.1
+    tail_frac: float = 0.5
+    tail_threshold: float = 1e-3
+    init: str = "uniform"        # initial p_0 over modes: {uniform, truth}
+    style_len: int = 4           # style-segment length (within-mode entropy)
+    style_vocab: int = 6         # style token alphabet size
+    id_base: int = 2             # identity segment encodes the mode in this base
+
+    @property
+    def id_len(self) -> int:
+        """Identity-segment length: fewest base-``id_base`` digits to index ``K`` modes."""
+        if self.K <= 1:
+            return 1
+        return max(1, math.ceil(math.log(self.K, self.id_base)))
+
+    @property
+    def vocab(self) -> int:
+        """Token alphabet size (shared by identity and style segments)."""
+        return max(self.id_base, self.style_vocab)
+
+    @property
+    def seq_len(self) -> int:
+        """Total observation length in tokens."""
+        return self.id_len + self.style_len
+
+
+@dataclass(frozen=True)
+class ModelCfg:
+    """The generative learner. ``kind`` selects the architecture behind a thin adapter.
+
+    Neural hyperparameters are ignored by the ``histogram`` bridge model.
+    """
+
+    kind: str = "histogram"      # {histogram, rnn, vae, mlp}
+    hidden: int = 64
+    embed: int = 16
+    epochs: int = 30
+    lr: float = 1.0e-3
+    batch_size: int = 256
+    device: str = "auto"         # {auto, cpu, cuda}
+    latent: int = 16             # VAE latent dimension (VAE only)
+    beta: float = 1.0            # VAE KL weight (VAE only)
+    # Samples used to estimate a neural model's mode distribution by generate-and-classify
+    # (ignored by the exact histogram bridge). Larger -> less measurement noise on p_hat.
+    n_eval: int = 8000
+
+
+@dataclass(frozen=True)
+class NeuralDynamicsCfg:
+    """Generational dynamics: drift strength ``n`` + reused grounding/re-mint blocks."""
+
+    n: int = 4000                # pupil training-sample size (drift strength ~ 1/n)
+    grounding: GroundingCfg = field(default_factory=GroundingCfg)
+    remint: RemintCfg = field(default_factory=RemintCfg)
+
+
+@dataclass(frozen=True)
+class NeuralLineageCfg:
+    """A fully-resolved neural lineage configuration."""
+
+    synthetic: SyntheticCfg
+    model: ModelCfg = field(default_factory=ModelCfg)
+    dynamics: NeuralDynamicsCfg = field(default_factory=NeuralDynamicsCfg)
+    generations: int = 30
+    metrics: MetricsCfg = field(default_factory=MetricsCfg)
+
+    @staticmethod
+    def from_dict(cfg: Mapping[str, Any]) -> "NeuralLineageCfg":
+        """Build a validated NeuralLineageCfg from a nested mapping, filling defaults."""
+        if isinstance(cfg, NeuralLineageCfg):
+            return cfg
+        synthetic = _sub(cfg.get("synthetic", {}), SyntheticCfg)
+        model = _sub(cfg.get("model", {}), ModelCfg)
+        dyn_raw = dict(cfg.get("dynamics", {}))
+        dynamics = NeuralDynamicsCfg(
+            n=dyn_raw.get("n", NeuralDynamicsCfg.n),
+            grounding=_sub(dyn_raw.get("grounding", {}), GroundingCfg),
+            remint=_sub(dyn_raw.get("remint", {}), RemintCfg),
+        )
+        metrics = _sub(cfg.get("metrics", {}), MetricsCfg)
+        return NeuralLineageCfg(
+            synthetic=synthetic,
+            model=model,
+            dynamics=dynamics,
+            generations=int(cfg.get("generations", NeuralLineageCfg.generations)),
+            metrics=metrics,
+        )
+
+    def replace(self, **kw) -> "NeuralLineageCfg":
+        return replace(self, **kw)

src/neural/evaluate.py

@@ -0,0 +1,77 @@
+"""Metrics for a neural lineage — the *same* row schema as ``knowledge.lineage``.
+
+``measure_metrics`` takes a model's oracle-measured mode distribution ``p_hat`` and emits a
+row with exactly the columns Layer 1 logs per generation (``knowledge.lineage.record``),
+computed with the *same* ``knowledge.metrics`` functions. Identical columns are what let a
+neural collapse curve be plotted on top of an analytic one, and let the same figure and
+analysis code (``knowledge.analysis``) run unchanged.
+"""
+
+from __future__ import annotations
+
+import numpy as np
+
+from knowledge.config import MetricsCfg
+from knowledge.lineage import N_BANDS
+from knowledge.metrics import (
+    forward_kl,
+    heterozygosity,
+    per_region,
+    support_size,
+    tail_band_metrics,
+    tail_mass,
+)
+
+
+def measure_metrics(p: np.ndarray, p_star_orig: np.ndarray, tail_mask: np.ndarray,
+                    regions: np.ndarray, R: int, metrics_cfg: MetricsCfg) -> dict:
+    """Compute every per-generation metric for a measured mode distribution.
+
+    Mirrors ``knowledge.lineage.record`` field-for-field. ``forward_kl`` and the tail set
+    are always measured against the *original* truth, so a re-minted lineage that has lost
+    tails is penalised exactly as in Layer 1's E6.
+
+    Args:
+        p (np.ndarray): The model's measured mode distribution ``p_hat`` (length ``K``).
+        p_star_orig (np.ndarray): The original true distribution over modes.
+        tail_mask (np.ndarray): Boolean tail mask on the original truth.
+        regions (np.ndarray): Length-``K`` region index per mode.
+        R (int): Number of regions.
+        metrics_cfg (MetricsCfg): KL floor and support epsilon.
+
+    Returns:
+        dict: One row of metrics (no ``generation``/label columns; the runner adds those).
+    """
+    eps = metrics_cfg.support_eps
+    kl_floor = metrics_cfg.kl_floor
+    head_mask = ~tail_mask
+    n_tail = int(tail_mask.sum())
+    n_head = int(head_mask.sum())
+
+    row = {
+        "heterozygosity": heterozygosity(p),
+        "forward_kl": forward_kl(p_star_orig, p, kl_floor),
+        "tail_mass": tail_mass(p, tail_mask),
+        "support_size": support_size(p, eps),
+        "tail_support": int(np.sum(p[tail_mask] > eps)),
+        "head_support": int(np.sum(p[head_mask] > eps)),
+        "tail_frac_alive": (float(np.mean(p[tail_mask] > eps)) if n_tail else 0.0),
+        "head_frac_alive": (float(np.mean(p[head_mask] > eps)) if n_head else 0.0),
+        "tail_truth_mass_alive": (
+            float(p_star_orig[tail_mask][p[tail_mask] > eps].sum()
+                  / p_star_orig[tail_mask].sum()) if n_tail else 0.0),
+    }
+    if n_tail >= N_BANDS:
+        fa, _ = tail_band_metrics(p, p_star_orig, tail_mask, n_bands=N_BANDS, alive_eps=eps)
+        for b in range(N_BANDS):
+            row[f"band{b}_alive"] = fa[b]
+    if R > 1:
+        for r, v in per_region(heterozygosity, p, regions).items():
+            row[f"H_region_{r}"] = v
+        for r, v in per_region(tail_mass, p, regions, tail_mask).items():
+            row[f"tail_region_{r}"] = v
+        for r in range(R):
+            region_tail = (regions == r) & tail_mask
+            row[f"tailalive_region_{r}"] = (
+                float(np.mean(p[region_tail] > eps)) if region_tail.any() else 0.0)
+    return row

src/neural/experiment.py

@@ -0,0 +1,124 @@
+"""Neural (Layer 1.5) experiment runner: sweep a grid x replicates, write artifacts.
+
+Mirrors ``knowledge.experiment`` and reuses its sweep-expansion primitives
+(``_apply_param`` — including the ``g -> m`` conversion — and ``_set_by_path``), its
+provenance helpers, and its output contract (``save_artifacts``). Only the per-run call and
+the config key set differ: a neural run trains generative models rather than resampling a
+frequency vector, and its config groups are ``synthetic``/``model``/``dynamics``/... .
+
+CLI:  python -m neural.experiment configs/neural/N0.yaml
+"""
+
+from __future__ import annotations
+
+import argparse
+import copy
+import itertools
+from pathlib import Path
+from typing import Any
+
+import pandas as pd
+import yaml
+
+from knowledge.experiment import _apply_param, save_artifacts
+from knowledge.seeding import spawn_seeds
+
+from .generation_loop import run_generative_lineage
+
+# Config groups that make up a single neural lineage (everything else is experiment-level).
+_NEURAL_KEYS = ("synthetic", "model", "dynamics", "generations", "metrics", "n_eval")
+
+# Libraries recorded in the manifest on top of the Layer-1 core set (skipped if absent).
+_EXTRA_LIBS = ("torch", "torchvision")
+
+
+def expand_sweeps(cfg: dict) -> list[tuple[dict, dict]]:
+    """Expand the sweep grid into (label, resolved_neural_cfg) pairs.
+
+    Identical semantics to ``knowledge.experiment.expand_sweeps`` (Cartesian product of the
+    declared ``{param, values}`` entries, reusing ``_apply_param`` for the ``g -> m`` and
+    ``arm`` special cases) but assembling the base from the neural config groups.
+
+    Returns:
+        list[tuple[dict, dict]]: One (label-columns, neural-config) pair per grid point.
+    """
+    base = {k: copy.deepcopy(cfg[k]) for k in _NEURAL_KEYS if k in cfg}
+    sweeps = cfg.get("sweep", [])
+    if isinstance(sweeps, dict):
+        sweeps = [sweeps]
+    if not sweeps:
+        return [({}, base)]
+    params = [s["param"] for s in sweeps]
+    value_lists = [list(s["values"]) for s in sweeps]
+    combos: list[tuple[dict, dict]] = []
+    for values in itertools.product(*value_lists):
+        lin = copy.deepcopy(base)
+        label: dict = {}
+        for param, val in zip(params, values):
+            label.update(_apply_param(lin, param, val))
+        combos.append((label, lin))
+    return combos
+
+
+def run_experiment(cfg: dict) -> pd.DataFrame:
+    """Run every grid point x every replicate; return long-form results.
+
+    Replicate seeds are derived once from the master seed and reused across grid points, so
+    comparisons across sweep values are paired (shared drift noise) — as in Layer 1.
+
+    Args:
+        cfg (dict): Parsed experiment YAML.
+
+    Returns:
+        pd.DataFrame: One row per (combo, replicate, generation).
+    """
+    name = cfg["experiment"]
+    master = int(cfg["seed"])
+    n_rep = int(cfg["n_replicates"])
+    combos = expand_sweeps(cfg)
+    seeds = spawn_seeds(master, n_rep)
+
+    frames: list[pd.DataFrame] = []
+    for label, neural_cfg in combos:
+        for rep, ss in enumerate(seeds):
+            df = run_generative_lineage(neural_cfg, int(ss.generate_state(1)[0]))
+            for col, val in label.items():
+                df[col] = val
+            df["replicate"] = rep
+            frames.append(df)
+    out = pd.concat(frames, ignore_index=True)
+    out.insert(0, "experiment", name)
+    return out
+
+
+def run_and_save(config_path: str | Path) -> Path:
+    """Load a neural experiment YAML, run it, and write artifacts. Returns the output dir."""
+    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']}"))
+    kind = cfg.get("kind", "gen_lineage")
+    if kind == "recombination":
+        from .recombine import run_recombination  # Stage C (N4); imported lazily
+        df = run_recombination(cfg)
+        grid = None
+    elif kind == "gen_lineage":
+        df = run_experiment(cfg)
+        grid = [{"label": label, "neural_cfg": c} for label, c in expand_sweeps(cfg)]
+    else:
+        raise ValueError(f"unknown neural experiment kind {kind!r}")
+    model_kind = cfg.get("model", {}).get("kind", "histogram")
+    save_artifacts(cfg, df, out_dir, extra_libs=_EXTRA_LIBS,
+                   extra_manifest={"layer": "1.5", "model_kind": model_kind}, grid=grid)
+    return out_dir
+
+
+def main(argv: list[str] | None = None) -> None:
+    parser = argparse.ArgumentParser(description="Run a Layer-1.5 neural experiment from a YAML config.")
+    parser.add_argument("config", help="Path to configs/neural/NX.yaml")
+    args = parser.parse_args(argv)
+    out_dir = run_and_save(args.config)
+    print(f"wrote artifacts to {out_dir}/")
+
+
+if __name__ == "__main__":
+    main()

src/neural/generation_loop.py

@@ -0,0 +1,113 @@
+"""The neural analogue of ``knowledge.lineage.run_lineage``.
+
+Runs ``T`` generations of *train-a-model-on-the-previous-model's-samples*, the neural
+image of the Wright-Fisher generational step. Each generation the pupil is trained on a
+pool of (i) ``n`` observations drawn from the parent model (drift) and (ii) ``m`` fresh
+observations drawn from the grounding reference (immigration, ``g = m/(n+m)``), then its
+oracle-measured mode distribution is logged with the *same* metric schema Layer 1 uses.
+Grounding structure (proportional / uniform / matched over regions) and the re-mint gate
+reuse ``knowledge.step`` and mirror ``run_lineage`` exactly, so a histogram-model lineage
+reproduces the analytic core and a neural-model lineage tests whether the same signs hold
+in real weights.
+"""
+
+from __future__ import annotations
+
+from typing import Any, Mapping
+
+import numpy as np
+import pandas as pd
+
+from knowledge.metrics import heterozygosity
+from knowledge.step import allocate_m, structured_multinomial
+from knowledge.truth import uniform_init
+
+from .config import NeuralLineageCfg
+from .evaluate import measure_metrics
+from .models import make_model
+from .oracle import ExactOracle
+from .synthetic import id_codewords, make_mode_truth, render_modes
+
+
+def _counts_to_observations(counts: np.ndarray, cfg, rng, codewords) -> np.ndarray:
+    """Expand a per-mode count vector into rendered token sequences."""
+    modes = np.repeat(np.arange(counts.size), counts)
+    return render_modes(modes, cfg, rng, codewords)
+
+
+def run_generative_lineage(cfg: Mapping[str, Any] | NeuralLineageCfg,
+                           seed: int) -> pd.DataFrame:
+    """Run one neural lineage and return per-generation metrics.
+
+    Args:
+        cfg (Mapping | NeuralLineageCfg): Resolved neural-lineage configuration.
+        seed (int): Seed for this replicate; the run is a pure function of (cfg, seed) for
+            the histogram model (statistically reproducible for torch models).
+
+    Returns:
+        pd.DataFrame: One row per generation 0..T with the same metric columns as
+        ``knowledge.lineage.run_lineage``.
+    """
+    cfg = NeuralLineageCfg.from_dict(cfg)
+    syn = cfg.synthetic
+    td = make_mode_truth(syn)
+    p_star_orig = td.p_star            # forward_kl is always vs the original truth
+    regions = td.regions
+    tail_mask = td.tail_mask
+    R = syn.R
+
+    rng = np.random.default_rng(seed)
+    oracle = ExactOracle(syn)
+    codewords = id_codewords(syn)
+
+    # Initial distribution over modes (exact, like Layer 1).
+    if syn.init == "uniform":
+        p0 = uniform_init(syn.K)
+    elif syn.init == "truth":
+        p0 = p_star_orig.copy()
+    else:
+        raise ValueError(f"unknown init {syn.init!r} (expected uniform|truth)")
+
+    # Grounding wiring (reused verbatim from Layer 1).
+    grounding = cfg.dynamics.grounding
+    exercised = np.asarray(grounding.exercised) if grounding.exercised is not None else None
+    m_vector = allocate_m(grounding.m, R, grounding.policy, exercised)
+    p_star_eff = p_star_orig.copy()    # grounding reference; may be re-minted (N6)
+    remint = cfg.dynamics.remint
+    n = cfg.dynamics.n
+
+    model = make_model(cfg.model, syn, oracle)
+    model.initialise(p0, rng)
+
+    rows: list[dict] = []
+
+    def record(t: int, p: np.ndarray) -> None:
+        row = {"generation": t}
+        row.update(measure_metrics(p, p_star_orig, tail_mask, regions, R, cfg.metrics))
+        rows.append(row)
+
+    record(0, model.mode_distribution(rng))
+
+    for t in range(1, cfg.generations + 1):
+        X_syn = model.sample(n, rng)                      # drift: n from the parent
+        if m_vector is not None:                          # immigration: m grounded samples
+            counts_real = structured_multinomial(
+                m_vector, p_star_eff, regions, grounding.policy, rng)
+            X_real = _counts_to_observations(counts_real, syn, rng, codewords)
+            pool = np.concatenate([X_syn, X_real], axis=0)
+        else:
+            pool = X_syn
+
+        pupil = make_model(cfg.model, syn, oracle)
+        pupil.fit(pool, rng)
+        model = pupil
+        p = model.mode_distribution(rng)
+
+        if remint.enabled and remint.period and t % remint.period == 0:
+            # Founder event: current distribution becomes the new grounding reference and
+            # the original truth is discarded for grounding. Gated on diversity (N6).
+            if remint.H_gate is None or heterozygosity(p) >= remint.H_gate:
+                p_star_eff = p.copy()
+        record(t, p)
+
+    return pd.DataFrame(rows)

src/neural/models.py

@@ -0,0 +1,119 @@
+"""Generative models behind a thin adapter, so architecture is a config switch.
+
+Every model implements the same three-method protocol: ``fit`` on a batch of token
+sequences, ``sample`` fresh token sequences, and report its ``mode_distribution`` (the
+model's ``p_t``). Keeping the interface identical is what makes "collapse is
+architecture-general" (experiment N5) a single sweep over ``model.kind``.
+
+``HistogramModel`` is the bridge: its ``fit`` is a maximum-likelihood mode histogram and
+its ``sample`` is a multinomial draw, so a lineage of histogram models is *exactly*
+neutral Wright-Fisher drift with immigration — the analytic core in disguise. The
+torch-backed RNN/VAE/MLP models are added in Stage C and reuse this same protocol.
+"""
+
+from __future__ import annotations
+
+from typing import Protocol, runtime_checkable
+
+import numpy as np
+
+from .config import ModelCfg, SyntheticCfg
+from .oracle import Oracle, measure_distribution
+from .synthetic import id_codewords, render_modes
+
+
+@runtime_checkable
+class GenerativeModel(Protocol):
+    """A learner of ``p(x)`` over the synthetic observation space."""
+
+    def initialise(self, p0: np.ndarray, rng: np.random.Generator) -> None:
+        """Initialise generation 0 to represent the mode distribution ``p0``.
+
+        The histogram bridge sets ``p0`` exactly (matching Layer 1's exact ``p_0`` start);
+        a neural model trains on a sample drawn from ``p0`` (its gen-0 fidelity is checked
+        by the Stage-C fidelity gate).
+        """
+        ...
+
+    def fit(self, X: np.ndarray, rng: np.random.Generator) -> None:
+        """Train (from scratch) on a batch of token sequences ``X``."""
+        ...
+
+    def sample(self, n: int, rng: np.random.Generator) -> np.ndarray:
+        """Draw ``n`` fresh token sequences of shape ``(n, seq_len)``."""
+        ...
+
+    def mode_distribution(self, rng: np.random.Generator) -> np.ndarray:
+        """Return the model's length-``K`` distribution over modes (its ``p_t``)."""
+        ...
+
+
+class HistogramModel:
+    """MLE mode-histogram generator — reduces Layer 1.5 exactly to Layer 1.
+
+    ``fit`` counts oracle-labelled modes in the training pool and stores the empirical
+    distribution; ``sample`` draws modes multinomially and renders them; the stored
+    distribution *is* the model's ``p_t`` (read exactly, no eval-sampling noise). Composed
+    over generations this is neutral Wright-Fisher drift with immigration.
+
+    Args:
+        cfg (SyntheticCfg): The synthetic grammar (for ``K`` and rendering).
+        oracle (Oracle): The mode adjudicator used to label the training pool.
+        model_cfg (ModelCfg): Present for interface symmetry; unused by the histogram.
+    """
+
+    def __init__(self, cfg: SyntheticCfg, oracle: Oracle,
+                 model_cfg: ModelCfg | None = None) -> None:
+        self.cfg = cfg
+        self.oracle = oracle
+        self._codewords = id_codewords(cfg)
+        self._p: np.ndarray | None = None
+
+    def initialise(self, p0: np.ndarray, rng: np.random.Generator) -> None:
+        """Set the stored distribution to ``p0`` exactly (no gen-0 sampling noise)."""
+        p0 = np.asarray(p0, dtype=float)
+        self._p = p0 / p0.sum()
+
+    def fit(self, X: np.ndarray, rng: np.random.Generator) -> None:
+        """Store the empirical mode distribution of the (oracle-labelled) pool ``X``."""
+        self._p = measure_distribution(X, self.oracle, self.cfg.K)
+
+    def sample(self, n: int, rng: np.random.Generator) -> np.ndarray:
+        """Draw ``n`` observations whose modes follow the stored distribution."""
+        if self._p is None:
+            raise RuntimeError("HistogramModel.sample called before fit")
+        counts = rng.multinomial(n, self._p)
+        modes = np.repeat(np.arange(self.cfg.K), counts)
+        return render_modes(modes, self.cfg, rng, self._codewords)
+
+    def mode_distribution(self, rng: np.random.Generator) -> np.ndarray:
+        """Return the stored mode distribution (exact; no eval sampling)."""
+        if self._p is None:
+            raise RuntimeError("HistogramModel.mode_distribution called before fit")
+        return self._p.copy()
+
+
+def make_model(model_cfg: ModelCfg, cfg: SyntheticCfg, oracle: Oracle) -> GenerativeModel:
+    """Construct a generative model of the requested ``kind``.
+
+    Args:
+        model_cfg (ModelCfg): Selects the architecture and its hyperparameters.
+        cfg (SyntheticCfg): The synthetic grammar.
+        oracle (Oracle): The mode adjudicator (needed by the histogram bridge; the neural
+            models estimate their mode distribution by generate-and-classify).
+
+    Returns:
+        GenerativeModel: A fresh, untrained model.
+
+    Raises:
+        ValueError: If ``kind`` is unknown.
+    """
+    kind = model_cfg.kind
+    if kind == "histogram":
+        return HistogramModel(cfg, oracle, model_cfg)
+    if kind in ("rnn", "vae", "mlp"):
+        # Torch-backed models arrive in Stage C; imported lazily so Stages A-B need no GPU.
+        from .torch_models import make_torch_model  # noqa: PLC0415
+
+        return make_torch_model(model_cfg, cfg, oracle)
+    raise ValueError(f"unknown model kind {kind!r} (expected histogram|rnn|vae|mlp)")

src/neural/oracle.py

@@ -0,0 +1,82 @@
+"""The oracle — the neural analogue of Layer 1's "reality's no".
+
+An oracle maps an observation to the mode it belongs to. For the fully-synthetic sandbox
+the oracle is **exact** (it decodes the lossless identity segment), so the measured mode
+distribution ``p_hat`` carries zero measurement noise — this is what lets a trained model's
+collapse be read directly against the known ``p*``. (The MNIST tier will add a
+``ClassifierOracle`` wrapping a frozen network plus its confusion matrix; that arrives with
+Stage C and is confirmation-only.)
+"""
+
+from __future__ import annotations
+
+from typing import Protocol, runtime_checkable
+
+import numpy as np
+
+from .config import SyntheticCfg
+
+
+@runtime_checkable
+class Oracle(Protocol):
+    """Adjudicates which mode an observation belongs to."""
+
+    def classify(self, X: np.ndarray) -> np.ndarray:
+        """Return the length-``n`` mode index for each row of ``X``."""
+        ...
+
+
+class ExactOracle:
+    """Zero-error oracle for the synthetic sandbox: decodes the identity segment.
+
+    Args:
+        cfg (SyntheticCfg): The synthetic configuration whose grammar produced ``X``.
+    """
+
+    def __init__(self, cfg: SyntheticCfg) -> None:
+        self.cfg = cfg
+        self.id_len = cfg.id_len
+        self.base = cfg.id_base
+        self.K = cfg.K
+        # Positional weights for base-`base` decoding, most-significant digit first.
+        self._weights = self.base ** np.arange(self.id_len - 1, -1, -1, dtype=np.int64)
+
+    def classify(self, X: np.ndarray) -> np.ndarray:
+        """Decode mode indices from the identity segment of each observation.
+
+        Args:
+            X (np.ndarray): Token sequences of shape ``(n, seq_len)``.
+
+        Returns:
+            np.ndarray: Length-``n`` decoded indices. Grammar-valid data always decodes into
+            ``[0, K)``; a *neural* model may emit an invalid codeword that decodes to
+            ``>= K`` — such samples are dropped by :func:`measure_distribution` rather than
+            being clipped onto a real mode (which would bias ``p_hat``).
+        """
+        X = np.asarray(X, dtype=np.int64)
+        ident = X[:, : self.id_len]
+        return ident @ self._weights
+
+
+def measure_distribution(X: np.ndarray, oracle: Oracle, K: int) -> np.ndarray:
+    """Measure the empirical mode distribution ``p_hat`` of a sample.
+
+    This is the neural readout of ``p_t``: classify every observation and normalise the
+    mode histogram. Modes absent from ``X`` receive zero mass (support shrinks exactly as
+    in Layer 1). Invalid codewords (decoded index ``>= K``, only producible by a neural
+    model) are dropped, so ``p_hat`` is renormalised over grammar-valid samples.
+
+    Args:
+        X (np.ndarray): Token sequences of shape ``(n, seq_len)``.
+        oracle (Oracle): The mode adjudicator.
+        K (int): Number of modes.
+
+    Returns:
+        np.ndarray: Length-``K`` probability vector summing to 1.
+    """
+    modes = oracle.classify(X)
+    counts = np.bincount(modes, minlength=K)[:K].astype(float)
+    total = counts.sum()
+    if total <= 0:
+        raise ValueError("measure_distribution received an empty sample")
+    return counts / total

src/neural/synthetic.py

@@ -0,0 +1,109 @@
+"""The fully-synthetic sandbox: a known ``p*`` over modes + a lossless observation grammar.
+
+The mode-truth (``p*``, regions, tail mask) comes straight from Layer 1's
+``knowledge.truth.make_true_distribution`` — so "mode", "region", and "tail" are *the same
+objects* as in the analytic core. Each mode is rendered to a categorical token sequence:
+
+* an **identity** segment of ``id_len`` base-``id_base`` digits that encodes the mode
+  index exactly (the exact oracle reads these back with zero error), and
+* a **style** segment of ``style_len`` tokens drawn uniformly at random, giving genuine
+  within-mode entropy so a real generative model must learn a *distribution* ``p(x|mode)``
+  rather than memorise ``K`` fixed strings.
+
+Because the identity segment is lossless, the measured mode distribution ``p_hat`` is a
+noise-free readout of the model's output — the property that lets the histogram model
+reduce this layer exactly to Wright-Fisher drift.
+"""
+
+from __future__ import annotations
+
+import numpy as np
+
+from knowledge.truth import TrueDist, make_true_distribution
+
+from .config import SyntheticCfg
+
+
+def make_mode_truth(cfg: SyntheticCfg) -> TrueDist:
+    """Build the true distribution over modes (thin wrapper over Layer 1's truth).
+
+    Args:
+        cfg (SyntheticCfg): The synthetic configuration (its first seven fields are the
+            Layer-1 ``TruthCfg`` knobs).
+
+    Returns:
+        TrueDist: ``p_star`` (length ``K``), ``regions``, and ``tail_mask`` over modes.
+    """
+    return make_true_distribution(
+        cfg.K, cfg.R, cfg.tail, cfg.tail_frac, cfg.zipf_s, 0,
+        tail_threshold=cfg.tail_threshold,
+    )
+
+
+def id_codewords(cfg: SyntheticCfg) -> np.ndarray:
+    """Return the ``(K, id_len)`` matrix of base-``id_base`` identity codewords.
+
+    Codeword of mode ``k`` is ``k`` written in base ``id_base``, most-significant digit
+    first, zero-padded to ``id_len``. Deterministic and invertible.
+
+    Args:
+        cfg (SyntheticCfg): The synthetic configuration.
+
+    Returns:
+        np.ndarray: Integer array of shape ``(K, id_len)`` with tokens in
+        ``[0, id_base)``.
+    """
+    k = np.arange(cfg.K, dtype=np.int64)
+    id_len, base = cfg.id_len, cfg.id_base
+    digits = np.empty((cfg.K, id_len), dtype=np.int64)
+    for pos in range(id_len - 1, -1, -1):     # least-significant digit last
+        digits[:, pos] = k % base
+        k //= base
+    return digits
+
+
+def render_modes(modes: np.ndarray, cfg: SyntheticCfg, rng: np.random.Generator,
+                 codewords: np.ndarray | None = None) -> np.ndarray:
+    """Render an array of mode indices to token sequences.
+
+    Args:
+        modes (np.ndarray): Length-``n`` integer array of mode indices in ``[0, K)``.
+        cfg (SyntheticCfg): The synthetic configuration.
+        rng (np.random.Generator): Random source for the style segment.
+        codewords (np.ndarray | None): Optional precomputed identity codewords.
+
+    Returns:
+        np.ndarray: Integer array of shape ``(n, seq_len)`` — identity segment followed by
+        a freshly-sampled style segment.
+    """
+    modes = np.asarray(modes, dtype=np.int64)
+    if codewords is None:
+        codewords = id_codewords(cfg)
+    ident = codewords[modes]                                    # (n, id_len)
+    style = rng.integers(0, cfg.style_vocab, size=(modes.shape[0], cfg.style_len))
+    return np.concatenate([ident, style], axis=1)
+
+
+def sample_synthetic(p_over_modes: np.ndarray, n: int, cfg: SyntheticCfg,
+                     rng: np.random.Generator,
+                     codewords: np.ndarray | None = None) -> tuple[np.ndarray, np.ndarray]:
+    """Draw ``n`` observations whose modes follow ``p_over_modes``.
+
+    This is the *grounding* generator (draw from a fixed distribution and render) and the
+    reference sampler used to seed generation 0.
+
+    Args:
+        p_over_modes (np.ndarray): Distribution over the ``K`` modes to sample from.
+        n (int): Number of observations.
+        cfg (SyntheticCfg): The synthetic configuration.
+        rng (np.random.Generator): Random source.
+        codewords (np.ndarray | None): Optional precomputed identity codewords.
+
+    Returns:
+        tuple[np.ndarray, np.ndarray]: ``(X, modes)`` — token sequences of shape
+        ``(n, seq_len)`` and the length-``n`` true mode indices.
+    """
+    p = np.asarray(p_over_modes, dtype=float)
+    modes = rng.choice(cfg.K, size=n, p=p / p.sum())
+    X = render_modes(modes, cfg, rng, codewords)
+    return X, modes

src/neural/torch_mlp.py

@@ -0,0 +1,109 @@
+"""Autoregressive MLP generative model (Stage C, for the N5 architecture-generality axis).
+
+A causal feed-forward next-token model: token ``i`` is predicted from the concatenated
+(causally-masked) embeddings of all earlier tokens. Deliberately a *different* inductive
+bias from the GRU — if collapse appears here too, it is a property of the transmission
+operator, not of any one architecture.
+"""
+
+from __future__ import annotations
+
+import numpy as np
+
+from .config import ModelCfg, SyntheticCfg
+from .oracle import Oracle
+from .torch_models import _BaseTorchGenerator
+from .train import device_generator, seed_everything
+
+
+def _make_mlp_net(V: int, L: int, embed: int, hidden: int):
+    import torch
+    import torch.nn as nn
+
+    class MLPNet(nn.Module):
+        """Predict every position from a causally-masked flatten of prior embeddings."""
+
+        def __init__(self) -> None:
+            super().__init__()
+            self.V, self.L, self.E = V, L, embed
+            self.bos = V
+            self.embed = nn.Embedding(V + 1, embed)
+            self.net = nn.Sequential(
+                nn.Linear(L * embed, hidden), nn.ReLU(),
+                nn.Linear(hidden, hidden), nn.ReLU(),
+                nn.Linear(hidden, V),
+            )
+            # lower-triangular INCLUSIVE mask over input positions: position i sees inputs
+            # 0..i (the input is already shifted by one, so this is strictly causal on x).
+            mask = torch.tril(torch.ones(L, L))
+            self.register_buffer("mask", mask)
+
+        def _context(self, inp):                       # inp: (B, L) input tokens
+            B = inp.shape[0]
+            emb = self.embed(inp)                       # (B, L, E)
+            m = self.mask.to(emb.dtype)                 # (L, L)
+            # ctx[b, i] = concat_j ( emb[b, j] * mask[i, j] )  -> (B, L, L*E)
+            ctx = emb.unsqueeze(1) * m.unsqueeze(0).unsqueeze(-1)   # (B, L, L, E)
+            return ctx.reshape(B, self.L, self.L * self.E)
+
+        def forward(self, x):                           # x: (B, L) targets
+            B = x.shape[0]
+            bos = torch.full((B, 1), self.bos, dtype=torch.long, device=x.device)
+            inp = torch.cat([bos, x[:, :-1]], dim=1)
+            ctx = self._context(inp)
+            return self.net(ctx)                        # (B, L, V)
+
+        def step_logits(self, prefix):                  # prefix: (B, pos) tokens so far
+            """Logits for the next token given the tokens generated so far."""
+            B, pos = prefix.shape
+            bos = torch.full((B, 1), self.bos, dtype=torch.long, device=prefix.device)
+            inp = torch.cat([bos, prefix], dim=1)[:, : self.L]      # (B, <=L)
+            if inp.shape[1] < self.L:
+                pad = torch.zeros((B, self.L - inp.shape[1]), dtype=torch.long,
+                                  device=prefix.device)
+                inp = torch.cat([inp, pad], dim=1)
+            ctx = self._context(inp)                    # (B, L, L*E)
+            return self.net(ctx[:, pos, :])             # logits at position `pos`
+
+    return MLPNet()
+
+
+class MLPGenerator(_BaseTorchGenerator):
+    """Autoregressive feed-forward generative model over token sequences."""
+
+    def fit(self, X: np.ndarray, rng: np.random.Generator) -> None:
+        import torch
+
+        g = seed_everything(int(rng.integers(2 ** 31)))
+        net = _make_mlp_net(self.V, self.L, self.mcfg.embed, self.mcfg.hidden).to(self.device)
+        net.train()
+        opt = torch.optim.Adam(net.parameters(), lr=self.mcfg.lr)
+        loss_fn = torch.nn.CrossEntropyLoss()
+        data = torch.as_tensor(np.asarray(X), dtype=torch.long, device=self.device)
+        n, bs = data.shape[0], self.mcfg.batch_size
+        for _ in range(self.mcfg.epochs):
+            perm = torch.randperm(n, generator=g).to(self.device)
+            for i in range(0, n, bs):
+                batch = data[perm[i:i + bs]]
+                logits = net(batch)
+                loss = loss_fn(logits.reshape(-1, self.V), batch.reshape(-1))
+                opt.zero_grad()
+                loss.backward()
+                opt.step()
+        net.eval()
+        self.net = net
+
+    def sample(self, n: int, rng: np.random.Generator) -> np.ndarray:
+        import torch
+
+        if self.net is None:
+            raise RuntimeError("MLPGenerator.sample called before fit")
+        g = device_generator(int(rng.integers(2 ** 31)), self.device)
+        prefix = torch.empty((n, 0), dtype=torch.long, device=self.device)
+        with torch.no_grad():
+            for pos in range(self.L):
+                logits = self.net.step_logits(prefix)
+                probs = torch.softmax(logits, dim=-1)
+                tok = torch.multinomial(probs, 1, generator=g)
+                prefix = torch.cat([prefix, tok], dim=1)
+        return prefix.cpu().numpy()

src/neural/torch_models.py

@@ -0,0 +1,146 @@
+"""Torch-backed generative models over the synthetic token grammar (Stage C).
+
+Each model implements the same ``GenerativeModel`` protocol as the histogram bridge
+(``initialise`` / ``fit`` / ``sample`` / ``mode_distribution``), so a lineage is
+architecture-agnostic and N5 is a single sweep over ``model.kind``. Unlike the histogram
+model, a neural model's ``mode_distribution`` is *estimated* by generate-and-classify
+(``n_eval`` samples), which is the honest, slightly-noisy neural readout of ``p_t``.
+
+Implemented so far: ``RNNGenerator`` (autoregressive GRU). ``VAEGenerator`` and
+``MLPGenerator`` follow and reuse the shared measure/initialise helpers.
+"""
+
+from __future__ import annotations
+
+import numpy as np
+
+from .config import ModelCfg, SyntheticCfg
+from .oracle import Oracle, measure_distribution
+from .synthetic import sample_synthetic
+from .train import device_generator, resolve_device, seed_everything, set_determinism
+
+
+def _measure(model, rng: np.random.Generator, oracle: Oracle, K: int, n_eval: int) -> np.ndarray:
+    """Estimate a model's mode distribution by generate-and-classify."""
+    X = model.sample(n_eval, rng)
+    return measure_distribution(X, oracle, K)
+
+
+class _BaseTorchGenerator:
+    """Shared plumbing: device, gen-0 initialisation, and mode measurement."""
+
+    def __init__(self, cfg: SyntheticCfg, model_cfg: ModelCfg, oracle: Oracle) -> None:
+        self.cfg = cfg
+        self.mcfg = model_cfg
+        self.oracle = oracle
+        self.device = resolve_device(model_cfg.device)
+        self.V = cfg.vocab
+        self.L = cfg.seq_len
+        self.net = None
+        set_determinism()
+
+    def initialise(self, p0: np.ndarray, rng: np.random.Generator) -> None:
+        """Train generation 0 on a sample drawn from ``p0`` (fidelity-gated in Stage C)."""
+        n_init = max(self.mcfg.n_eval, 4000)
+        X, _ = sample_synthetic(p0, n_init, self.cfg, rng)
+        self.fit(X, rng)
+
+    def mode_distribution(self, rng: np.random.Generator) -> np.ndarray:
+        return _measure(self, rng, self.oracle, self.cfg.K, self.mcfg.n_eval)
+
+
+# --- autoregressive GRU -----------------------------------------------------------------
+
+def _make_ar_net(V: int, embed: int, hidden: int):
+    """Build an autoregressive GRU next-token network (built lazily to avoid a torch import
+    at module load)."""
+    import torch.nn as nn
+
+    class ARNet(nn.Module):
+        """Predict token ``i`` from tokens ``0..i-1`` via a GRU (BOS-prefixed)."""
+
+        def __init__(self) -> None:
+            super().__init__()
+            self.bos = V                      # extra input id for the start token
+            self.embed = nn.Embedding(V + 1, embed)
+            self.gru = nn.GRU(embed, hidden, batch_first=True)
+            self.out = nn.Linear(hidden, V)
+
+        def forward(self, x):                 # x: (B, L) target tokens
+            import torch
+
+            B = x.shape[0]
+            bos = torch.full((B, 1), self.bos, dtype=torch.long, device=x.device)
+            inp = torch.cat([bos, x[:, :-1]], dim=1)   # teacher forcing
+            h, _ = self.gru(self.embed(inp))
+            return self.out(h)                # (B, L, V)
+
+    return ARNet()
+
+
+class RNNGenerator(_BaseTorchGenerator):
+    """Autoregressive GRU generative model over token sequences."""
+
+    def _train_net(self, X, rng: np.random.Generator):
+        import torch
+
+        g = seed_everything(int(rng.integers(2 ** 31)))
+        net = _make_ar_net(self.V, self.mcfg.embed, self.mcfg.hidden).to(self.device)
+        net.train()
+        opt = torch.optim.Adam(net.parameters(), lr=self.mcfg.lr)
+        loss_fn = torch.nn.CrossEntropyLoss()
+        data = torch.as_tensor(np.asarray(X), dtype=torch.long, device=self.device)
+        n = data.shape[0]
+        bs = self.mcfg.batch_size
+        for _ in range(self.mcfg.epochs):
+            perm = torch.randperm(n, generator=g).to(self.device)
+            for i in range(0, n, bs):
+                idx = perm[i:i + bs]
+                batch = data[idx]
+                logits = net(batch)                       # (b, L, V)
+                loss = loss_fn(logits.reshape(-1, self.V), batch.reshape(-1))
+                opt.zero_grad()
+                loss.backward()
+                opt.step()
+        net.eval()
+        return net
+
+    def fit(self, X: np.ndarray, rng: np.random.Generator) -> None:
+        self.net = self._train_net(X, rng)
+
+    def sample(self, n: int, rng: np.random.Generator) -> np.ndarray:
+        import torch
+
+        if self.net is None:
+            raise RuntimeError("RNNGenerator.sample called before fit")
+        g = device_generator(int(rng.integers(2 ** 31)), self.device)
+        out = torch.empty((n, self.L), dtype=torch.long, device=self.device)
+        tok = torch.full((n, 1), self.V, dtype=torch.long, device=self.device)  # BOS
+        h = None
+        with torch.no_grad():
+            for pos in range(self.L):
+                emb = self.net.embed(tok)
+                hid, h = self.net.gru(emb, h)
+                logits = self.net.out(hid[:, -1, :])          # (n, V)
+                probs = torch.softmax(logits, dim=-1)
+                tok = torch.multinomial(probs, 1, generator=g)
+                out[:, pos] = tok[:, 0]
+        return out.cpu().numpy()
+
+
+# --- dispatch ---------------------------------------------------------------------------
+
+def make_torch_model(model_cfg: ModelCfg, cfg: SyntheticCfg, oracle: Oracle):
+    """Construct a torch generative model of the requested ``kind``."""
+    kind = model_cfg.kind
+    if kind == "rnn":
+        return RNNGenerator(cfg, model_cfg, oracle)
+    if kind == "vae":
+        from .torch_vae import VAEGenerator  # noqa: PLC0415
+
+        return VAEGenerator(cfg, model_cfg, oracle)
+    if kind == "mlp":
+        from .torch_mlp import MLPGenerator  # noqa: PLC0415
+
+        return MLPGenerator(cfg, model_cfg, oracle)
+    raise ValueError(f"unknown torch model kind {kind!r}")

src/neural/torch_vae.py

@@ -0,0 +1,104 @@
+"""Sequence VAE generative model (Stage C, for the N5 architecture-generality axis).
+
+A GRU encoder maps a token sequence to a Gaussian latent ``z``; a GRU decoder (its initial
+hidden state projected from ``z``) reconstructs the sequence. Trained by the ELBO
+(reconstruction CE + ``beta`` * KL). A latent-variable generator is a third, distinct
+inductive bias — and the canonical model in which generative collapse was first studied — so
+its collapse under dry self-training is strong evidence the effect is operator-driven.
+"""
+
+from __future__ import annotations
+
+import numpy as np
+
+from .config import ModelCfg, SyntheticCfg
+from .oracle import Oracle
+from .torch_models import _BaseTorchGenerator
+from .train import device_generator, seed_everything
+
+
+def _make_vae(V: int, L: int, embed: int, hidden: int, latent: int):
+    import torch
+    import torch.nn as nn
+
+    class SeqVAE(nn.Module):
+        def __init__(self) -> None:
+            super().__init__()
+            self.V, self.L, self.bos = V, L, V
+            self.embed = nn.Embedding(V + 1, embed)
+            self.enc = nn.GRU(embed, hidden, batch_first=True)
+            self.to_mu = nn.Linear(hidden, latent)
+            self.to_lv = nn.Linear(hidden, latent)
+            self.z_to_h = nn.Linear(latent, hidden)
+            self.dec = nn.GRU(embed, hidden, batch_first=True)
+            self.out = nn.Linear(hidden, V)
+
+        def encode(self, x):
+            _, h = self.enc(self.embed(x))              # h: (1, B, H)
+            h = h[-1]
+            return self.to_mu(h), self.to_lv(h)
+
+        def decode_logits(self, z, x):                  # teacher forcing
+            B = x.shape[0]
+            bos = torch.full((B, 1), self.bos, dtype=torch.long, device=x.device)
+            inp = torch.cat([bos, x[:, :-1]], dim=1)
+            h0 = torch.tanh(self.z_to_h(z)).unsqueeze(0)   # (1, B, H)
+            out, _ = self.dec(self.embed(inp), h0)
+            return self.out(out)
+
+        def forward(self, x):
+            mu, lv = self.encode(x)
+            std = torch.exp(0.5 * lv)
+            z = mu + std * torch.randn_like(std)
+            logits = self.decode_logits(z, x)
+            kl = -0.5 * torch.sum(1 + lv - mu.pow(2) - lv.exp(), dim=1).mean()
+            return logits, kl
+
+    return SeqVAE()
+
+
+class VAEGenerator(_BaseTorchGenerator):
+    """Sequence VAE generative model over token sequences."""
+
+    def fit(self, X: np.ndarray, rng: np.random.Generator) -> None:
+        import torch
+
+        g = seed_everything(int(rng.integers(2 ** 31)))
+        net = _make_vae(self.V, self.L, self.mcfg.embed, self.mcfg.hidden,
+                        self.mcfg.latent).to(self.device)
+        net.train()
+        opt = torch.optim.Adam(net.parameters(), lr=self.mcfg.lr)
+        ce = torch.nn.CrossEntropyLoss()
+        data = torch.as_tensor(np.asarray(X), dtype=torch.long, device=self.device)
+        n, bs, beta = data.shape[0], self.mcfg.batch_size, self.mcfg.beta
+        for _ in range(self.mcfg.epochs):
+            perm = torch.randperm(n, generator=g).to(self.device)
+            for i in range(0, n, bs):
+                batch = data[perm[i:i + bs]]
+                logits, kl = net(batch)
+                recon = ce(logits.reshape(-1, self.V), batch.reshape(-1))
+                loss = recon + beta * kl / batch.shape[0]
+                opt.zero_grad()
+                loss.backward()
+                opt.step()
+        net.eval()
+        self.net = net
+
+    def sample(self, n: int, rng: np.random.Generator) -> np.ndarray:
+        import torch
+
+        if self.net is None:
+            raise RuntimeError("VAEGenerator.sample called before fit")
+        g = device_generator(int(rng.integers(2 ** 31)), self.device)
+        z = torch.randn(n, self.mcfg.latent, generator=g, device=self.device)
+        h = torch.tanh(self.net.z_to_h(z)).unsqueeze(0)     # (1, n, H)
+        tok = torch.full((n, 1), self.net.bos, dtype=torch.long, device=self.device)
+        out = torch.empty((n, self.L), dtype=torch.long, device=self.device)
+        with torch.no_grad():
+            for pos in range(self.L):
+                dec_out, h = self.net.dec(self.net.embed(tok), h)
+                logits = self.net.out(dec_out[:, -1, :])
+                probs = torch.softmax(logits, dim=-1)
+                tok = torch.multinomial(probs, 1, generator=g)
+                out[:, pos] = tok[:, 0]
+        return out.cpu().numpy()

src/neural/train.py

@@ -0,0 +1,79 @@
+"""Torch determinism, seeding, and device helpers for the neural (Stage C) models.
+
+Layer 1.5's neural tiers are *statistically* reproducible, not bitwise (blueprint 4 rider):
+we derive every torch seed from the same ``SeedSequence`` stream the rest of the study uses,
+set all available determinism flags, and report per-seed points. ``CUBLAS_WORKSPACE_CONFIG``
+must be set before the first CUDA op, so it is set at import time.
+"""
+
+from __future__ import annotations
+
+import os
+
+os.environ.setdefault("CUBLAS_WORKSPACE_CONFIG", ":4096:8")
+
+import numpy as np
+
+_DETERMINISM_SET = False
+
+
+def torch_seed_from(seed) -> int:
+    """Reduce an int or ``np.random.SeedSequence`` to a 32-bit torch seed."""
+    if isinstance(seed, np.random.SeedSequence):
+        return int(seed.generate_state(1)[0])
+    return int(seed) & 0xFFFFFFFF
+
+
+def resolve_device(device: str):
+    """Resolve ``{auto, cpu, cuda}`` to a concrete ``torch.device``."""
+    import torch
+
+    if device == "auto":
+        device = "cuda" if torch.cuda.is_available() else "cpu"
+    return torch.device(device)
+
+
+def set_determinism() -> None:
+    """Set torch/cuDNN determinism flags once per process (best-effort)."""
+    global _DETERMINISM_SET
+    if _DETERMINISM_SET:
+        return
+    import torch
+
+    torch.use_deterministic_algorithms(True, warn_only=True)
+    torch.backends.cudnn.deterministic = True
+    torch.backends.cudnn.benchmark = False
+    _DETERMINISM_SET = True
+
+
+def seed_everything(seed) -> "object":
+    """Seed torch (CPU+CUDA) from ``seed`` and return a seeded CPU ``torch.Generator``.
+
+    Args:
+        seed: An int or ``np.random.SeedSequence`` from the study's seed stream.
+
+    Returns:
+        torch.Generator: A CPU generator seeded for sampling ops (e.g. ``randperm``).
+    """
+    import torch
+
+    s = torch_seed_from(seed)
+    torch.manual_seed(s)
+    if torch.cuda.is_available():
+        torch.cuda.manual_seed_all(s)
+    g = torch.Generator()
+    g.manual_seed(s)
+    return g
+
+
+def device_generator(seed, device) -> "object":
+    """Return a ``torch.Generator`` on ``device`` seeded from ``seed``.
+
+    ``torch.multinomial`` requires the generator to live on the same device as the
+    probabilities, so sampling ops use this rather than the CPU generator.
+    """
+    import torch
+
+    g = torch.Generator(device=device)
+    g.manual_seed(torch_seed_from(seed))
+    return g