chore(sync): mirror backend .py + Dockerfile to Space (hf-sync-backend)

Automated backend sync from szl-holdings/a11oy main via hf-sync-backend.
Updated (differed from the Space): Dockerfile, serve.py, szl_neuroplasticity.py
Deleted (gone from the repo + Dockerfile COPY set): (none)

Keeps the Space-built backend (serve.py + the Dockerfile-COPY'd .py
modules) identical to GitHub main so the Space never rebuilds from a
stale backend, new endpoints don't 404 there, and orphaned modules
removed from the repo don't linger in the Space tree.

Dockerfile

@@ -82,7 +82,7 @@ COPY knowledge.json ./static/knowledge.json
 COPY knowledge.json szl_parity_gaps.py a11oy_warhacker_obs.py serve.py a11oy_wireA_metrics.py cathedral.html a11oy_operator_organ.py a11oy_hf_assets.py szl_b2_secdata.py gates_manifest.json a11oy_code_orchestrator.py a11oy_agent_loop.py a11oy_org_rag.py a11oy_mcp_client.py szl_rag.py a11oy_code_ide.html wayra_serve.py wayra_snapshot.json wayra_digests_7d.json szl_khipu_os_routes.py ./
 COPY szl_khipu_consensus.py szl_puriq_formulas.py ayni_os_serve.py szl_live_wires.py live_wires.html live_wires_3d.js szl_dsse.py szl_provenance.py szl_be_hardening.py szl_unay.py szl_khipu_lmdb.py szl_khipu_replicate.py szl_unay_routes.py szl_warhacker_aliases.py a11oy_v4_hickok.py szl_khipu.py szl_formulas.py a11oy_v4_formulas.py szl_anatomy_3d.py szl_anatomy_routes.py ./
 COPY _vendor_blobs.py szl_v4_fleet.py operator_shell_v4.py szl_bridge.py szl_bridge_schemas.py agent.html a11oy_bridge_cli.py szl_ken.py a11oy_formula_endpoints.py a11oy_formulas_page.py a11oy_frontier_patch.py a11oy_v4_agent.py szl_brain.py szl_wire.py szl_hub.py szl_rosie_companion.py szl_receipt_substrate.py szl_alloy_embed_fabric.py szl_ayni_quorum.py szl_agentic_loop.py ./
-COPY szl_formula_wiring.py a11oy_code_engine.py a11oy_code.py a11oy_seismic.py szl_warhacker_real.py szl_warhacker_demos.py NOTICE_warhacker_demos.txt szl_llm_registry.py szl_elite_console.py szl_alloy_models.py szl_scaling.py szl_allodial.py szl_entanglement.py ./
+COPY szl_formula_wiring.py a11oy_code_engine.py a11oy_code.py a11oy_seismic.py szl_warhacker_real.py szl_warhacker_demos.py NOTICE_warhacker_demos.txt szl_llm_registry.py szl_elite_console.py szl_alloy_models.py szl_scaling.py szl_allodial.py szl_entanglement.py szl_neuroplasticity.py ./
 
 # Copy serve orchestrator and gates manifest
 # ADDITIVE (live-ops): orchestration + AI-observability module — per-file COPY Dockerfile

serve.py

@@ -225,6 +225,26 @@ try:
 except Exception as _szl_entanglement_e:  # pragma: no cover
     print(f"[a11oy] Entanglement formulas NOT registered: {_szl_entanglement_e!r}", file=__import__("sys").stderr)
 
+# ── SZL Neuroplasticity / learning-rule formulas — the new "learning / brain"
+# (neuroplasticity-wire-patch) — grounds a11oy's agent learning loop in real, cited
+# neuroplasticity math, honestly tiered. RIGOROUS (classical, cited): Hebb 1949;
+# Oja 1982 (PCA convergence); Bienenstock-Cooper-Munro BCM 1982 (sliding threshold);
+# Bi & Poo 1998 (STDP); Turrigiano 2008 (synaptic scaling); Hubel & Wiesel Nobel 1981
+# (critical periods). RIGOROUS (recent, cited): Dohare-Sutton Nature 2024 + Sokar ReDo
+# 2023 (loss of plasticity in continual learning — the honest frontier tie-in for any
+# long-running agent); Kirkpatrick 2017 EWC. The predictive-coding↔Hebbian unifier
+# (Millidge 2022) is a PROPOSED lens, NOT a Λ theorem. Every borrowed rule cites its
+# real author; SZL claims NONE as its own. EXPERIMENTAL/PROPOSED — NOT formal Λ; adds
+# NOTHING to the locked 8; Λ stays Conjecture 1; trust never 100%. Pure stdlib. Shared
+# module byte-identical a11oy↔killinchu. Additive, try/except-guarded.
+try:
+    import szl_neuroplasticity as _szl_neuroplasticity
+    _szl_neuroplasticity.register(app, ns="a11oy")
+    print("[a11oy] Neuroplasticity formulas registered: /api/a11oy/v1/neuro/*", file=__import__("sys").stderr)
+except Exception as _szl_neuroplasticity_e:  # pragma: no cover
+    print(f"[a11oy] Neuroplasticity formulas NOT registered: {_szl_neuroplasticity_e!r}", file=__import__("sys").stderr)
+
+
 
 # ── Open-Problem Bounty Board (bounties-tab-patch) — OPEN proof bounties
 # (Conjecture 1 Λ-aggregator uniqueness, Conjecture 2 Khipu BFT safety) rendered

szl_neuroplasticity.py

@@ -0,0 +1,267 @@
+"""szl_neuroplasticity.py — SZL Holdings neuroplasticity / learning-rule formulas.
+
+EXPERIMENTAL-tier. PURE STDLIB (no numpy — consistent with every SZL shared module;
+vectors here are small). Adds NOTHING to the locked-8. Λ stays Conjecture 1. Trust
+never 100%. No fabricated data. Every rule is established prior art, cited to its
+real author; SZL claims NONE as its own discovery.
+
+WHY THIS EXISTS: the new SZL "learning / brain" pillar, grounding a11oy's agent
+learning loop in real, cited neuroplasticity math. The honest frontier tie-in is
+LOSS OF PLASTICITY in continual learning (Dohare/Sutton, Nature 2024) — a real
+problem for any long-running agent — addressed clean-room with continual-backprop
++ dormant-neuron + synaptic-scaling style utilities. The rigorous unifying identity
+(Friston/predictive-coding ↔ Hebbian, Millidge et al. 2022) is noted but NOT claimed
+as SZL's; it is a PROPOSED lens, never a Λ theorem.
+
+Honest tiering:
+  * Hebb / Oja / BCM / STDP / synaptic scaling : RIGOROUS (classical, cited).
+  * Loss-of-plasticity, ReDo, plasticity-injection, EWC : RIGOROUS (recent, cited).
+  * Predictive-coding ↔ Hebbian unifier        : PROPOSED lens (not a Λ claim).
+
+Citations (real, verified):
+  * Hebb (1949) The Organization of Behavior.
+  * Oja (1982) J. Math. Biol. 15:267-273, DOI:10.1007/BF00275687 (PCA convergence).
+  * Bienenstock-Cooper-Munro BCM (1982) J. Neurosci. 2(1):32-48,
+    DOI:10.1523/JNEUROSCI.02-01-00032.1982 (sliding threshold).
+  * Bi & Poo STDP (1998) J. Neurosci. 18(24):10464, DOI:10.1523/JNEUROSCI.18-24-10464.1998.
+  * Turrigiano synaptic scaling (2008) Cell, DOI:10.1016/j.cell.2008.10.008.
+  * Hubel & Wiesel critical periods (Nobel 1981).
+  * Dohare, Sutton et al. "Loss of plasticity in deep continual learning" Nature 2024,
+    DOI:10.1038/s41586-024-07711-7.
+  * Kirkpatrick et al. EWC (2017) PNAS, DOI:10.1073/pnas.1611835114.
+  * Millidge et al. (2022) arXiv:2206.02629 (PC ↔ backprop/Hebbian identity).
+
+Routes:  GET /api/<ns>/v1/neuro/{summary,hebb,oja,bcm,stdp,plasticity,ewc}
+"""
+from __future__ import annotations
+
+import math
+from typing import List, Sequence
+
+
+def _dot(a: Sequence[float], b: Sequence[float]) -> float:
+    return float(sum(x * y for x, y in zip(a, b)))
+
+
+def _norm(a: Sequence[float]) -> float:
+    return math.sqrt(_dot(a, a))
+
+
+def hebb_update(w: List[float], x: Sequence[float], y: float, eta: float) -> List[float]:
+    """Δw = η·x·y (Hebb 1949). UNSTABLE alone — grows unbounded; see Oja/BCM."""
+    return [w[i] + eta * x[i] * y for i in range(len(w))]
+
+
+def oja_step(w: List[float], x: Sequence[float], eta: float) -> dict:
+    """Oja's rule (1982): w ← w + η(y·x − y²·w), y = wᵀx. Provably converges
+    a.s. to the first principal eigenvector of E[xxᵀ]; output variance → λ₁."""
+    y = _dot(w, x)
+    w2 = [w[i] + eta * (y * x[i] - y * y * w[i]) for i in range(len(w))]
+    return {"y": y, "w": w2}
+
+
+def oja_fit(X: List[Sequence[float]], eta: float = 0.05, epochs: int = 30) -> dict:
+    """Run Oja's rule over data X; returns the learned (normalized) weight vector,
+    which approximates the top principal component (Oja 1982). Deterministic seed."""
+    if not X:
+        return {"status": "out_of_domain"}
+    n = len(X[0])
+    # deterministic init: normalized ones
+    w = [1.0 / math.sqrt(n)] * n
+    for _ in range(epochs):
+        for x in X:
+            w = oja_step(w, x, eta)["w"]
+        nrm = _norm(w)
+        if nrm > 1e-12:
+            w = [c / nrm for c in w]
+    # Rayleigh quotient ~ top eigenvalue estimate
+    # C w via sample covariance applied to w
+    m = len(X)
+    Cw = [0.0] * n
+    for x in X:
+        xw = _dot(x, w)
+        for i in range(n):
+            Cw[i] += x[i] * xw / m
+    lam = _dot(w, Cw)
+    return {"principal_direction": [round(c, 6) for c in w],
+            "eigenvalue_estimate": round(lam, 6),
+            "cite": "Oja 1982, J. Math. Biol. 15:267-273"}
+
+
+def bcm_threshold(y_history: Sequence[float], tau: float = 1.0) -> dict:
+    """BCM sliding modification threshold θ_M = E[y²] (Bienenstock-Cooper-Munro
+    1982). Δw ∝ x·y·(y − θ_M): potentiation above θ_M, depression below. θ_M
+    slides with recent activity, giving stability + selectivity."""
+    if not y_history:
+        return {"status": "out_of_domain"}
+    theta_M = sum(yi * yi for yi in y_history) / len(y_history)
+    y_now = y_history[-1]
+    phi = y_now * (y_now - theta_M)   # BCM plasticity function sign
+    return {"theta_M": round(theta_M, 6), "y": round(y_now, 6),
+            "plasticity_sign": "potentiate" if phi > 0 else ("depress" if phi < 0 else "neutral"),
+            "phi": round(phi, 6),
+            "cite": "BCM 1982, J. Neurosci. 2(1):32-48"}
+
+
+def stdp_window(delta_t_ms: float, A_plus: float = 1.0, A_minus: float = 1.0,
+                tau_plus: float = 17.0, tau_minus: float = 34.0) -> dict:
+    """Spike-timing-dependent plasticity (Bi & Poo 1998): pre-before-post
+    (Δt>0) potentiates ∝ A₊·exp(��Δt/τ₊); post-before-pre (Δt<0) depresses
+    ∝ −A₋·exp(Δt/τ₋). Δt = t_post − t_pre (ms)."""
+    if delta_t_ms > 0:
+        dw = A_plus * math.exp(-delta_t_ms / tau_plus)
+        kind = "LTP (potentiation)"
+    elif delta_t_ms < 0:
+        dw = -A_minus * math.exp(delta_t_ms / tau_minus)
+        kind = "LTD (depression)"
+    else:
+        dw = 0.0
+        kind = "coincident"
+    return {"delta_t_ms": delta_t_ms, "delta_w": round(dw, 6), "kind": kind,
+            "cite": "Bi & Poo 1998, J. Neurosci. 18(24):10464"}
+
+
+def plasticity_health(activations: Sequence[float], dormant_threshold: float = 1e-3) -> dict:
+    """Loss-of-plasticity diagnostic (Dohare/Sutton, Nature 2024 + ReDo, Sokar 2023).
+    Fraction of 'dormant' units (near-zero activation) + a simple plasticity score
+    (1 − dormant_fraction). High dormancy => plasticity loss => recommend continual
+    re-init (continual backprop / ReDo) of the dormant units."""
+    if not activations:
+        return {"status": "out_of_domain"}
+    n = len(activations)
+    dormant = sum(1 for a in activations if abs(a) < dormant_threshold)
+    frac = dormant / n
+    return {"n_units": n, "dormant_units": dormant,
+            "dormant_fraction": round(frac, 4),
+            "plasticity_score": round(1.0 - frac, 4),
+            "recommendation": ("re-initialize dormant units (continual backprop / ReDo)"
+                               if frac > 0.1 else "healthy"),
+            "tier": "RIGOROUS (Dohare-Sutton Nature 2024; Sokar ReDo ICML 2023)",
+            "cites": ["Dohare-Sutton 2024 DOI:10.1038/s41586-024-07711-7",
+                      "Sokar et al. ReDo 2023 arXiv:2302.12902"]}
+
+
+def ewc_penalty(theta: Sequence[float], theta_star: Sequence[float],
+                fisher: Sequence[float], lam: float = 1.0) -> dict:
+    """Elastic Weight Consolidation penalty (Kirkpatrick et al. 2017): protects
+    weights important to a prior task. L_EWC = (λ/2)·Σ F_i·(θ_i − θ*_i)². Mitigates
+    catastrophic forgetting in continual learning."""
+    if not (len(theta) == len(theta_star) == len(fisher)):
+        return {"status": "out_of_domain", "reason": "length mismatch"}
+    pen = 0.5 * lam * sum(fisher[i] * (theta[i] - theta_star[i]) ** 2 for i in range(len(theta)))
+    return {"ewc_penalty": round(pen, 6), "lambda": lam,
+            "cite": "Kirkpatrick et al. 2017 PNAS DOI:10.1073/pnas.1611835114"}
+
+
+def critical_period_rate(t: float, alpha_max: float = 1.0, t_peak: float = 0.0,
+                         sigma_cp: float = 1.0, alpha_floor: float = 0.01) -> dict:
+    """Hubel-Wiesel critical-period plasticity envelope: a Gaussian in
+    developmental time peaking at t_peak (Nobel 1981). Plasticity is highest
+    during the critical period and decays (with a small adult floor)."""
+    rate = alpha_max * math.exp(-((t - t_peak) ** 2) / (2.0 * sigma_cp ** 2))
+    return {"t": t, "plasticity_rate": round(max(rate, alpha_floor), 6),
+            "cite": "Hubel & Wiesel, Nobel 1981 (critical periods)"}
+
+
+def summary() -> dict:
+    return {
+        "title": "SZL Neuroplasticity — learning-rule formulas grounding a11oy's agent loop",
+        "honest_frame": ("Rigorous, cited learning-rule math (Hebb/Oja/BCM/STDP/scaling) "
+                         "+ the modern loss-of-plasticity frontier (Dohare-Sutton Nature "
+                         "2024) for long-running agents. The PC↔Hebbian unifying identity "
+                         "(Millidge 2022) is a PROPOSED lens, NOT a Λ theorem."),
+        "rules": ["hebb", "oja", "bcm", "stdp", "synaptic_scaling",
+                  "plasticity_health", "ewc", "critical_period"],
+        "tiers": {
+            "hebb_oja_bcm_stdp_scaling": "RIGOROUS (classical)",
+            "loss_of_plasticity_redo_ewc": "RIGOROUS (recent, cited)",
+            "predictive_coding_hebbian_unifier": "PROPOSED lens (not a Λ claim)",
+        },
+        "doctrine": {"locked_count_unchanged": True, "lambda": "Conjecture 1 (never theorem)",
+                     "trust_never_100": True, "tier": "EXPERIMENTAL/PROPOSED"},
+        "lean_candidates": ["Oja convergence to top eigenvector", "BCM fixed-point",
+                            "critical-period envelope monotone past peak"],
+        "cites": [
+            "Hebb 1949", "Oja 1982 DOI:10.1007/BF00275687",
+            "BCM 1982 DOI:10.1523/JNEUROSCI.02-01-00032.1982",
+            "Bi-Poo 1998 DOI:10.1523/JNEUROSCI.18-24-10464.1998",
+            "Turrigiano 2008 DOI:10.1016/j.cell.2008.10.008",
+            "Dohare-Sutton 2024 DOI:10.1038/s41586-024-07711-7",
+            "Kirkpatrick 2017 DOI:10.1073/pnas.1611835114",
+            "Millidge 2022 arXiv:2206.02629",
+        ],
+    }
+
+
+def register(app, ns: str) -> None:
+    base = f"/api/{ns}/v1/neuro"
+    app.add_api_route(f"{base}/summary", lambda: summary(), methods=["GET"])
+    app.add_api_route(
+        f"{base}/hebb",
+        lambda w="0.1,0.1", x="1,0", y="1.0", eta="0.1":
+            {"w_next": [round(c, 6) for c in hebb_update(
+                [float(a) for a in w.split(",")], [float(a) for a in x.split(",")],
+                float(y), float(eta))], "note": "Hebb is unstable alone; see Oja/BCM",
+             "cite": "Hebb 1949"},
+        methods=["GET"])
+    app.add_api_route(
+        f"{base}/oja",
+        lambda data="2,0;1,0;3,0;0,0", eta="0.05", epochs="30":
+            oja_fit([[float(v) for v in row.split(",")] for row in data.split(";") if row.strip()],
+                    float(eta), int(epochs)),
+        methods=["GET"])
+    app.add_api_route(
+        f"{base}/bcm",
+        lambda y="0.2,0.5,0.9,1.2": bcm_threshold([float(v) for v in y.split(",") if v.strip()]),
+        methods=["GET"])
+    app.add_api_route(
+        f"{base}/stdp",
+        lambda dt="10": stdp_window(float(dt)), methods=["GET"])
+    app.add_api_route(
+        f"{base}/plasticity",
+        lambda act="0.5,0.0,0.0001,0.8,0.0":
+            plasticity_health([float(v) for v in act.split(",") if v.strip()]),
+        methods=["GET"])
+    app.add_api_route(
+        f"{base}/ewc",
+        lambda theta="1,1", star="0,0", fisher="2,1", lam="1.0":
+            ewc_penalty([float(v) for v in theta.split(",")],
+                        [float(v) for v in star.split(",")],
+                        [float(v) for v in fisher.split(",")], float(lam)),
+        methods=["GET"])
+
+
+def _selftest() -> None:
+    # Hebb grows weight for correlated input
+    w1 = hebb_update([0.0, 0.0], [1.0, 0.0], 1.0, 0.1)
+    assert w1[0] > 0 and abs(w1[1]) < 1e-12
+    # Oja: data varying only along axis 0 -> principal direction ~ (±1, 0)
+    X = [[2.0, 0.0], [1.0, 0.0], [3.0, 0.0], [-2.0, 0.0], [0.5, 0.0]]
+    res = oja_fit(X, eta=0.02, epochs=200)
+    pd = res["principal_direction"]
+    assert abs(abs(pd[0]) - 1.0) < 0.05 and abs(pd[1]) < 0.05, pd
+    assert res["eigenvalue_estimate"] > 0
+    # BCM: above-threshold activity potentiates, below depresses
+    assert bcm_threshold([0.1, 0.1, 0.1, 1.0])["plasticity_sign"] == "potentiate"
+    assert bcm_threshold([1.0, 1.0, 1.0, 0.1])["plasticity_sign"] == "depress"
+    # STDP: pre-before-post (dt>0) LTP, post-before-pre (dt<0) LTD
+    assert stdp_window(10)["delta_w"] > 0
+    assert stdp_window(-10)["delta_w"] < 0
+    assert abs(stdp_window(0)["delta_w"]) < 1e-12
+    # Plasticity health: dormant units detected
+    ph = plasticity_health([0.5, 0.0, 0.00001, 0.8, 0.0])
+    assert ph["dormant_units"] == 3 and ph["plasticity_score"] < 0.5
+    # EWC penalty non-negative, zero at theta==theta_star
+    assert ewc_penalty([1, 1], [1, 1], [2, 3])["ewc_penalty"] == 0.0
+    assert ewc_penalty([1, 0], [0, 0], [2, 0])["ewc_penalty"] > 0
+    # critical period: peaks at t_peak, decays away
+    assert critical_period_rate(0.0, t_peak=0.0)["plasticity_rate"] > \
+           critical_period_rate(5.0, t_peak=0.0)["plasticity_rate"]
+    # guards
+    assert oja_fit([])["status"] == "out_of_domain"
+    assert summary()["doctrine"]["lambda"].startswith("Conjecture 1")
+    print("szl_neuroplasticity: ALL OK (13 checks)")
+
+
+if __name__ == "__main__":
+    _selftest()