#!/usr/bin/env python3
"""Build a plain, evidence-first English introduction to the SDIL project."""

from pathlib import Path

from PIL import Image
from pptx import Presentation
from pptx.dml.color import RGBColor
from pptx.enum.text import MSO_ANCHOR, PP_ALIGN
from pptx.util import Inches, Pt


ROOT = Path(__file__).resolve().parents[1]
SLIDES = ROOT / "slides"
ASSETS = SLIDES / "assets"
OUTPUT = SLIDES / "SDIL_project_intro.pptx"

FIG3 = ROOT / "results/figs/figure3_innovation.png"
FIG4 = ROOT / "results/figs/figure4_resnet_confirmation.png"
FIG5 = ROOT / "results/figs/figure5_bci_v2.png"
FIG6 = ROOT / "results/figs/figure6_standard_depth_scaling.png"

FONT = "DejaVu Sans"
BLACK = "111111"
GRAY = "111111"
LIGHT_GRAY = "B5B5B5"
WHITE = "FFFFFF"


def rgb(value: str) -> RGBColor:
    return RGBColor.from_string(value)


def add_text(
    slide,
    text: str,
    x: float,
    y: float,
    w: float,
    h: float,
    *,
    size: float = 18,
    bold: bool = False,
    color: str = BLACK,
    align=PP_ALIGN.LEFT,
    valign=MSO_ANCHOR.TOP,
    margin: float = 0.02,
    font: str = FONT,
):
    box = slide.shapes.add_textbox(Inches(x), Inches(y), Inches(w), Inches(h))
    frame = box.text_frame
    frame.clear()
    frame.word_wrap = True
    frame.margin_left = Inches(margin)
    frame.margin_right = Inches(margin)
    frame.margin_top = Inches(margin)
    frame.margin_bottom = Inches(margin)
    frame.vertical_anchor = valign
    paragraph = frame.paragraphs[0]
    paragraph.alignment = align
    run = paragraph.add_run()
    run.text = text
    run.font.name = font
    run.font.size = Pt(size)
    run.font.bold = bold
    run.font.color.rgb = rgb(color)
    return box


def add_bullets(slide, items, x, y, w, h, *, size=18, gap=7, color=BLACK):
    box = slide.shapes.add_textbox(Inches(x), Inches(y), Inches(w), Inches(h))
    frame = box.text_frame
    frame.clear()
    frame.word_wrap = True
    frame.margin_left = Inches(0.02)
    frame.margin_right = Inches(0.02)
    frame.margin_top = Inches(0.01)
    frame.margin_bottom = Inches(0.01)
    for index, item in enumerate(items):
        paragraph = frame.paragraphs[0] if index == 0 else frame.add_paragraph()
        paragraph.space_after = Pt(gap)
        paragraph.line_spacing = 1.05
        paragraph.text = ""
        bullet = paragraph.add_run()
        bullet.text = "•  "
        bullet.font.name = FONT
        bullet.font.size = Pt(size)
        bullet.font.color.rgb = rgb(color)
        body = paragraph.add_run()
        body.text = item
        body.font.name = FONT
        body.font.size = Pt(size)
        body.font.color.rgb = rgb(color)
    return box


def add_title(slide, title: str, number: int, subtitle: str | None = None):
    add_text(slide, title, 0.52, 0.28, 11.95, 0.55, size=27, bold=True)
    add_text(slide, str(number), 12.35, 0.35, 0.42, 0.25, size=10, color=GRAY, align=PP_ALIGN.RIGHT)
    if subtitle:
        add_text(slide, subtitle, 0.55, 0.92, 12.10, 0.36, size=13.5, color=GRAY)


def add_footer(slide, text: str):
    add_text(slide, text, 0.55, 7.17, 12.20, 0.18, size=8.5, color=GRAY)


def add_takeaway(slide, text: str, y: float = 6.55):
    add_text(slide, text, 0.70, y, 11.93, 0.40, size=16, bold=True, align=PP_ALIGN.CENTER, valign=MSO_ANCHOR.MIDDLE)


def prepare_crops():
    ASSETS.mkdir(parents=True, exist_ok=True)

    image5 = Image.open(FIG5)
    w5, h5 = image5.size
    image5.crop((0, 0, w5 // 2, h5 // 2)).save(ASSETS / "figure5a_actor_critic.png")
    image5.crop((w5 // 2, h5 // 2, w5, h5)).save(ASSETS / "figure5d_outcome.png")

    image6 = Image.open(FIG6)
    w6, h6 = image6.size
    image6.crop((0, 0, w6, h6 // 2)).save(ASSETS / "figure6_top_row.png")


def build_deck():
    prepare_crops()

    presentation = Presentation()
    presentation.slide_width = Inches(13.333)
    presentation.slide_height = Inches(7.5)
    blank = presentation.slide_layouts[6]

    def new_slide():
        slide = presentation.slides.add_slide(blank)
        slide.background.fill.solid()
        slide.background.fill.fore_color.rgb = rgb(WHITE)
        return slide

    # 1. Title.
    slide = new_slide()
    add_text(slide, "Somato-Dendritic Innovation Learning", 0.70, 1.02, 11.95, 0.75, size=34, bold=True)
    add_text(slide, "Local credit assignment from soma-unpredicted dendritic signals", 0.72, 1.96, 11.75, 0.52, size=22)
    add_text(
        slide,
        "Inspired by the neuron-specific somato-dendritic residuals reported by Francioni et al., Nature 2026",
        0.72,
        2.70,
        11.50,
        0.55,
        size=16,
        color=GRAY,
    )
    add_text(
        slide,
        "Question: should a local learner use apical activity itself, or the component that is unexpected from the same neuron's somatic state?",
        0.72,
        4.35,
        11.60,
        1.00,
        size=23,
        bold=True,
    )
    add_text(slide, "Project overview · August 2026", 0.72, 6.80, 4.0, 0.30, size=11, color=GRAY)

    # 2. Biological observation and learning rule.
    slide = new_slide()
    add_title(slide, "From the biological observation to a local learning rule", 2)
    add_text(
        slide,
        "The Harnett-lab study fits each neuron's normal soma–dendrite relationship. The residual is decorrelated from soma and carries outcome and signed task information.",
        0.72,
        1.18,
        11.75,
        0.76,
        size=18,
    )
    add_text(slide, "aₗ = sₗ + nₗ", 1.08, 2.25, 3.10, 0.52, size=25, bold=True, align=PP_ALIGN.CENTER)
    add_text(slide, "mixed apical activity", 1.08, 2.78, 3.10, 0.32, size=13, color=GRAY, align=PP_ALIGN.CENTER)
    add_text(slide, "âₗ = Pₗ(hₗ)", 5.10, 2.25, 3.10, 0.52, size=25, bold=True, align=PP_ALIGN.CENTER)
    add_text(slide, "neutral-period prediction", 5.10, 2.78, 3.10, 0.32, size=13, color=GRAY, align=PP_ALIGN.CENTER)
    add_text(slide, "rₗ = aₗ − âₗ", 9.13, 2.25, 3.10, 0.52, size=25, bold=True, align=PP_ALIGN.CENTER)
    add_text(slide, "somato-dendritic innovation", 9.13, 2.78, 3.10, 0.32, size=13, color=GRAY, align=PP_ALIGN.CENTER)
    add_text(
        slide,
        "ΔWₗ = η (rₗ ⊙ φ′(uₗ)) hₗ₋₁ᵀ",
        2.63,
        3.66,
        8.10,
        0.68,
        size=29,
        bold=True,
        align=PP_ALIGN.CENTER,
    )
    add_text(slide, "presynaptic activity × local postsynaptic gain × dendritic innovation", 2.62, 4.38, 8.12, 0.36, size=15, color=GRAY, align=PP_ALIGN.CENTER)
    add_bullets(
        slide,
        [
            "Training uses locally available quantities, forward observations, and manual synaptic updates.",
            "Node-perturbation feedback supports the small-network experiments; reciprocal Kolen–Pollack plasticity supports the standard ResNets.",
            "The inherited credit path transports layer-specific directions. SDIL removes soma-predictable traffic from those directions.",
        ],
        1.02,
        5.05,
        11.15,
        1.45,
        size=15.5,
        gap=4,
    )
    add_footer(slide, "Sources: Francioni et al. (Nature 2026); Lansdell et al. (ICLR 2020); Akrout et al. (NeurIPS 2019).")

    # 3. Controlled ablation.
    slide = new_slide()
    add_title(
        slide,
        "Residualization is load-bearing under predictable traffic",
        3,
        "Controlled MNIST intervention · depth 3 · width 256 · five paired seeds",
    )
    slide.shapes.add_picture(str(FIG3), Inches(0.42), Inches(1.37), width=Inches(12.48), height=Inches(4.87))
    add_takeaway(slide, "At traffic strength ρ = 0.5: raw 10.38%, norm-matched raw 10.31%, innovation 97.35%.")
    add_footer(slide, "Norm matching equalizes per-example update magnitude. The remaining difference comes from teaching direction.")

    # 4. Standard ResNet-20 confirmation.
    slide = new_slide()
    add_title(
        slide,
        "Dynamic innovation survives a standard ResNet-20 test",
        4,
        "CIFAR-10 · 200 epochs · five untouched paired seeds · four-times-RMS soma-predictable traffic",
    )
    slide.shapes.add_picture(str(FIG4), Inches(0.43), Inches(1.47), width=Inches(12.46), height=Inches(3.82))
    add_bullets(
        slide,
        [
            "Dynamic SDIL under traffic: 91.584% mean test accuracy; clean reciprocal KP: 91.388%.",
            "Mean early-layer teaching-signal cosine: 0.999687.",
            "Cost: 1.326× the matched BP MAC estimate, zero task-loss queries, and one neutral observation per example.",
        ],
        1.20,
        5.56,
        10.95,
        1.08,
        size=15.5,
        gap=3,
    )
    add_footer(slide, "Exact gradients serve diagnostic measurement; local signals drive every learning update.")

    # 5. Standard-depth scaling.
    slide = new_slide()
    add_title(
        slide,
        "Scaling from ResNet-20 to ResNet-56",
        5,
        "Complete 60-endpoint CIFAR-10 panel · five seeds per method and depth · shared hyperparameters",
    )
    figure6_top = ASSETS / "figure6_top_row.png"
    slide.shapes.add_picture(str(figure6_top), Inches(0.56), Inches(1.42), width=Inches(12.22), height=Inches(4.38))
    add_text(
        slide,
        "SDIL under 4× traffic: 91.584% → 92.254% → 92.760%; all five R20/R56 pairs improve.",
        0.73,
        6.01,
        11.90,
        0.40,
        size=14.8,
        bold=True,
        align=PP_ALIGN.CENTER,
    )
    add_text(
        slide,
        "Reciprocal KP provides the scalable credit path. SDIL preserves this path under mixed apical traffic.",
        0.73,
        6.48,
        11.90,
        0.34,
        size=14.5,
        align=PP_ALIGN.CENTER,
    )
    add_footer(slide, "BP and clean reciprocal KP follow the same depth trend; fixed DFA remains near 31%.")

    # 6. Dynamical task.
    slide = new_slide()
    add_title(
        slide,
        "A local actor–critic produces temporal outcome signals",
        6,
        "Synthetic BCI task · six task clusters × five model seeds · manual local updates",
    )
    panel_a = ASSETS / "figure5a_actor_critic.png"
    panel_d = ASSETS / "figure5d_outcome.png"
    slide.shapes.add_picture(str(panel_a), Inches(0.58), Inches(1.42), width=Inches(5.96), height=Inches(4.43))
    slide.shapes.add_picture(str(panel_d), Inches(6.80), Inches(1.42), width=Inches(5.96), height=Inches(4.43))
    add_text(
        slide,
        "Final success: 100% · terminal outcome decoding: 99.83% · outcome-lesion effect: 0.400.",
        0.68,
        6.06,
        11.98,
        0.36,
        size=14.8,
        bold=True,
        align=PP_ALIGN.CENTER,
    )
    add_text(
        slide,
        "The synthetic task supplies terminal reward and tests neuron-specific outcome surprise with causal lesions.",
        0.68,
        6.52,
        11.98,
        0.32,
        size=14.0,
        align=PP_ALIGN.CENTER,
    )
    add_footer(slide, "Outcome labels and exact causal roles serve evaluation.")

    # 7. Current scientific position.
    slide = new_slide()
    add_title(slide, "Current scientific position", 7)
    add_text(slide, "Established", 0.78, 1.30, 3.30, 0.38, size=20, bold=True)
    add_bullets(
        slide,
        [
            "Soma-predictable apical traffic can rotate a raw local teaching signal.",
            "Per-neuron innovation preserves learning direction under controlled traffic.",
            "Dynamic innovation works with a scalable BP-free reciprocal credit path on ResNet-20/32/56.",
            "A separate local actor–critic reproduces neuron-specific outcome signaling in a controlled dynamical task.",
        ],
        0.82,
        1.82,
        11.50,
        2.15,
        size=17,
        gap=7,
    )
    add_text(slide, "Attribution", 0.78, 4.15, 3.30, 0.38, size=20, bold=True)
    add_bullets(
        slide,
        [
            "Reciprocal KP provides clean-setting scale and feedback tracking.",
            "SDIL provides soma-conditioned traffic removal and the associated mechanism tests.",
        ],
        0.82,
        4.66,
        11.50,
        0.92,
        size=17,
        gap=7,
    )
    add_text(slide, "Next decisive experiment", 0.78, 5.82, 4.20, 0.38, size=20, bold=True)
    add_text(
        slide,
        "Evaluate raw reciprocal credit and SDIL under naturally generated, state-dependent mixed traffic with component-level bias hidden from the learner.",
        0.82,
        6.30,
        11.45,
        0.58,
        size=18,
    )
    add_footer(slide, "Target claim: scalable local credit assignment with adaptive removal of predictable feedback contamination.")

    SLIDES.mkdir(parents=True, exist_ok=True)
    presentation.save(OUTPUT)
    print(OUTPUT)


if __name__ == "__main__":
    build_deck()