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>

configs/neural/N0.yaml

@@ -0,0 +1,42 @@
+experiment: N0_bridge_histogram
+kind: gen_lineage
+seed: 20260704
+n_replicates: 60
+generations: 200
+
+# Bridge / harness-faithfulness check (Layer 1.5 build-order Stage B): run the E2 grounding
+# phase-boundary sweep through the NEURAL runner with the histogram model, which is exactly
+# neutral Wright-Fisher drift with immigration. Stationary H must track the closed form
+# H_eq = H* * m(2n+m-1)/(n+2nm+m^2) and reproduce a critical g* << 1 — i.e. the neural
+# plumbing reproduces the analytic core before any real network is trained. R=1 so every
+# grounding policy reduces to the proportional immigration model H_eq is derived for.
+synthetic:
+  K: 200
+  R: 1
+  tail: zipf
+  zipf_s: 1.1
+  tail_frac: 0.5
+  tail_threshold: 1.0e-3
+  init: truth
+  style_len: 2
+  style_vocab: 4
+  id_base: 2
+
+model:
+  kind: histogram
+
+dynamics:
+  n: 200
+  grounding: {m: 0, policy: proportional}   # m is overwritten per g by the sweep
+  remint: {enabled: false, period: null, H_gate: null}
+
+metrics:
+  kl_floor: 1.0e-9
+  support_eps: 1.0e-9
+
+sweep:
+  - param: g
+    values: [0.0, 0.005, 0.01, 0.02, 0.05, 0.1, 0.2, 0.4]
+
+output:
+  dir: results/N0

configs/neural/N1.yaml

@@ -0,0 +1,50 @@
+experiment: N1_collapse_in_weights
+kind: gen_lineage
+seed: 20260704
+n_replicates: 5
+
+# N1 (Layer 1.5, maps to Layer-1 E1 / blueprint C1): does model collapse appear in REAL
+# trained weights under dry recursive self-training, and does a little grounding arrest it?
+# An autoregressive RNN is retrained each generation on n samples drawn from the previous
+# generation's RNN (drift), optionally mixed with m verifier-grounded samples. Expect: the
+# dry arm (g=0) loses diversity (H falls) and forgets the tail (forward-KL to truth rises,
+# tail_mass shrinks); grounded arms hold. Falsifier: dry inheritance does not degrade in
+# real weights -> the neural collapse claim is unsupported at this scale.
+generations: 25
+
+synthetic:
+  K: 256
+  R: 1
+  tail: zipf
+  zipf_s: 1.3
+  tail_frac: 0.5
+  tail_threshold: 1.0e-3
+  init: truth
+  style_len: 3
+  style_vocab: 5
+  id_base: 2
+
+model:
+  kind: rnn
+  hidden: 128
+  embed: 24
+  epochs: 25
+  lr: 2.0e-3
+  batch_size: 256
+  n_eval: 12000
+
+dynamics:
+  n: 200
+  grounding: {m: 0, policy: proportional}   # m overwritten per g by the sweep
+  remint: {enabled: false, period: null, H_gate: null}
+
+metrics:
+  kl_floor: 1.0e-9
+  support_eps: 1.0e-9
+
+sweep:
+  - param: g
+    values: [0.0, 0.02, 0.05, 0.1]
+
+output:
+  dir: results/N1

configs/neural/N2.yaml

@@ -0,0 +1,50 @@
+experiment: N2_grounding_phase_boundary_neural
+kind: gen_lineage
+seed: 20260704
+n_replicates: 5
+
+# N2 (Layer 1.5 headline, maps to Layer-1 E2): the grounding phase boundary in REAL weights.
+# Sweep the grounding fraction g = m/(n+m) and locate the neural critical g* at which
+# stationary diversity is restored. Layer 1 found g* = 0.048 << 1. The neural regime (finite
+# model capacity, a smaller K so gen-0 fidelity holds) will not reproduce that value exactly
+# -- the claim is directional (blueprint 3.5): a critical g* << 1 exists in trained weights,
+# i.e. a little grounding protects most of the diversity. Falsifier: stationary H flat in g,
+# or only restored as g -> 1.
+generations: 30
+
+synthetic:
+  K: 256
+  R: 1
+  tail: zipf
+  zipf_s: 1.3
+  tail_frac: 0.5
+  tail_threshold: 1.0e-3
+  init: truth
+  style_len: 3
+  style_vocab: 5
+  id_base: 2
+
+model:
+  kind: rnn
+  hidden: 128
+  embed: 24
+  epochs: 25
+  lr: 2.0e-3
+  batch_size: 256
+  n_eval: 12000
+
+dynamics:
+  n: 200
+  grounding: {m: 0, policy: proportional}   # m overwritten per g by the sweep
+  remint: {enabled: false, period: null, H_gate: null}
+
+metrics:
+  kl_floor: 1.0e-9
+  support_eps: 1.0e-9
+
+sweep:
+  - param: g
+    values: [0.0, 0.005, 0.01, 0.02, 0.05, 0.1, 0.2]
+
+output:
+  dir: results/N2

configs/neural/N5.yaml

@@ -0,0 +1,54 @@
+experiment: N5_architecture_generality
+kind: gen_lineage
+seed: 20260704
+n_replicates: 5
+
+# N5 (a new Layer-1.5 axis, no Layer-1 counterpart): is collapse ARCHITECTURE-GENERAL?
+# Run the same dry-vs-grounded protocol across three genuinely different learners that
+# share only the generative-collapse operator: the exact histogram (= Wright-Fisher, the
+# analytic anchor), an autoregressive GRU (recurrent), and an autoregressive MLP (feed-
+# forward). Expect the same SIGN in all: dry (g=0) loses diversity / forgets the tail;
+# grounding arrests it. Falsifier: the signs appear only for the histogram -> real neural
+# inductive biases break the Wright-Fisher mapping. (The sequence VAE is implemented but
+# excluded here: it does not clear the gen-0 fidelity gate on the Zipf-codeword task, so
+# its collapse would be confounded with underfitting; see tasks/todo.md.)
+generations: 22
+
+synthetic:
+  K: 256
+  R: 1
+  tail: zipf
+  zipf_s: 1.3
+  tail_frac: 0.5
+  tail_threshold: 1.0e-3
+  init: truth
+  style_len: 3
+  style_vocab: 5
+  id_base: 2
+
+model:
+  kind: rnn          # overwritten per arm by the model.kind sweep
+  hidden: 192
+  embed: 24
+  epochs: 25
+  lr: 2.0e-3
+  batch_size: 256
+  n_eval: 12000
+
+dynamics:
+  n: 200
+  grounding: {m: 0, policy: proportional}   # m overwritten per g by the sweep
+  remint: {enabled: false, period: null, H_gate: null}
+
+metrics:
+  kl_floor: 1.0e-9
+  support_eps: 1.0e-9
+
+sweep:
+  - param: model.kind
+    values: [histogram, rnn, mlp]
+  - param: g
+    values: [0.0, 0.05]
+
+output:
+  dir: results/N5