ai_analysis.py

import json
import os
import re

from dotenv import load_dotenv

# Load environment variables from .venv file
load_dotenv(".venv")

# ── Gemini via REST (no SDK) ───────────────────────────────
try:
    import requests as _requests
    GEMINI_AVAILABLE = True
except ImportError:
    _requests = None
    GEMINI_AVAILABLE = False

GEMINI_MODEL = "gemini-2.5-flash"
GEMINI_URL = (
    "https://generativelanguage.googleapis.com/v1beta/models/"
    "{model}:generateContent"
)
# Injected into the prompt so the model can give specific, grounded explanations rather than generic RL platitudes.
TICKER_CONTEXT = {
    "NVDA": (
        "NVDA experienced an extreme bull run in 2023–2024 driven by "
        "AI infrastructure demand. Its price appreciation exceeded 200% "
        "during this period — a regime completely unlike the 2015–2022 "
        "training window where NVDA was a volatile but range-bound "
        "semiconductor stock."
    ),
    "GOOGL": (
        "GOOGL underwent a post-pandemic valuation reset followed by a "
        "strong recovery in 2023–2024. The test period coincided with "
        "renewed AI optimism around Google DeepMind and Gemini, which "
        "created momentum patterns different from the training window."
    ),
    "AAPL": (
        "AAPL exhibited a selective, patient strategy with very few trades "
        "and long average hold durations (~145 days). The test period "
        "included Apple's Vision Pro launch and continued services growth, "
        "creating a steady uptrend that rewarded holding."
    ),
    "MSFT": (
        "MSFT was the strongest performer. The test period coincided with "
        "Azure AI growth and the OpenAI partnership driving consistent "
        "institutional buying. The agent learned a moderate-frequency "
        "trading strategy that captured most of this uptrend."
    ),
    "AMZN": (
        "AMZN recovered strongly in 2023–2024 after a severe 2022 drawdown. "
        "The agent learned an aggressive strategy with high trade frequency "
        "and 60% position sizing, effectively catching the recovery wave."
    ),
}

# ── Training period context ────────────────────────────────────
TRAINING_CONTEXT = (
    "The PPO agents were trained on data from 2015 to mid 2026 using "
    "a custom Gymnasium environment with a 30-day lookback window, "
    "MultiDiscrete action space (direction × size), and a Differential "
    "Sharpe Ratio reward function with idle and invalid action penalties. "
    "Hyperparameters were tuned via Optuna TPE with 30 trials per ticker, "
    "optimising validation Sharpe Ratio."
)


def build_analysis_prompt(
    ticker: str,
    kpis_rl: dict,
    kpis_bnh: dict,
    kpis_sma: dict,
    actions: list,
    trades: int,
    period: str,
) -> str:
    # Action breakdown percentages
    total = len(actions) if actions else 1
    hold_pct = actions.count(0) / total * 100 if actions else 0
    buy_pct = actions.count(1) / total * 100 if actions else 0
    sell_pct = actions.count(2) / total * 100 if actions else 0

    ticker_ctx = TICKER_CONTEXT.get(ticker, "No specific context available.")

    prompt = f"""
You are a quantitative finance analyst reviewing the performance of a
Reinforcement Learning trading agent (PPO algorithm) on {ticker} stock.

=== TRAINING CONTEXT ===
{TRAINING_CONTEXT}

=== TICKER-SPECIFIC CONTEXT ===
{ticker_ctx}

=== BACKTEST RESULTS ({period} live data) ===

RL Agent (PPO):
  Final Value      : {kpis_rl.get("Final Value", "N/A")}
  Total Return     : {kpis_rl.get("Total Return", "N/A")}
  Ann. Return      : {kpis_rl.get("Ann. Return", "N/A")}
  Ann. Volatility  : {kpis_rl.get("Ann. Volatility", "N/A")}
  Sharpe Ratio     : {kpis_rl.get("Sharpe Ratio", "N/A")}
  Sortino Ratio    : {kpis_rl.get("Sortino Ratio", "N/A")}
  Max Drawdown     : {kpis_rl.get("Max Drawdown", "N/A")}
  Calmar Ratio     : {kpis_rl.get("Calmar Ratio", "N/A")}
  Total Trades     : {trades}
  Action Breakdown : Hold {hold_pct:.1f}% | Buy {buy_pct:.1f}% | Sell {sell_pct:.1f}%

Buy & Hold:
  Total Return     : {kpis_bnh.get("Total Return", "N/A")}
  Sharpe Ratio     : {kpis_bnh.get("Sharpe Ratio", "N/A")}
  Max Drawdown     : {kpis_bnh.get("Max Drawdown", "N/A")}

SMA Crossover:
  Total Return     : {kpis_sma.get("Total Return", "N/A")}
  Sharpe Ratio     : {kpis_sma.get("Sharpe Ratio", "N/A")}
  Max Drawdown     : {kpis_sma.get("Max Drawdown", "N/A")}

=== YOUR TASK ===

Provide a structured analysis in EXACTLY this JSON format.
Return ONLY the JSON object, no markdown, no preamble:

{{
  "summary": "2-3 sentence executive summary of overall performance",
  "strengths": [
    "Specific strength 1 with numbers from the results",
    "Specific strength 2",
    "Specific strength 3"
  ],
  "weaknesses": [
    "Specific weakness 1 — diagnose the root cause",
    "Specific weakness 2",
    "Specific weakness 3"
  ],
  "failure_diagnosis": "1-2 paragraphs specifically explaining WHY the agent performed the way it did on {ticker}. Reference the training period, regime change, reward function, action space, or hyperparameters as appropriate. Be precise and technical.",
  "improvements": [
    "Concrete improvement 1 with implementation detail",
    "Concrete improvement 2",
    "Concrete improvement 3",
    "Concrete improvement 4"
  ],
  "verdict": "Final verdict in this exact form: start with a letter grade (A/B/C/D/F) based on actual performance, then state the alpha vs Buy & Hold and the Sharpe with real numbers, then a clear deployment call. Grade A = clearly beat B&H with Sharpe >= 1; B = small positive edge; C = roughly matched B&H; D = underperformed; F = never traded. Be specific and non-cliché — no 'shows promise but needs tuning'."
}}
"""
    return prompt.strip()


def get_ai_analysis(
    ticker: str,
    kpis_rl: dict,
    kpis_bnh: dict,
    kpis_sma: dict,
    actions: list,
    trades: int,
    period: str,
) -> dict:
    if _requests is None:
        return _fallback_analysis(ticker, kpis_rl, kpis_bnh, trades)

    try:
        api_key = os.getenv("GEMINI_API_KEY")
        if not api_key:
            return _fallback_analysis(
                ticker,
                kpis_rl,
                kpis_bnh,
                trades,
                note="GEMINI_API_KEY not found in environment.",
            )
        prompt = build_analysis_prompt(
            ticker, kpis_rl, kpis_bnh, kpis_sma, actions, trades, period
        )

        # ── Gemini REST call (no SDK, no websockets dependency) ───────
        url = GEMINI_URL.format(model=GEMINI_MODEL)
        resp = _requests.post(
            url,
            headers={"Content-Type": "application/json"},
            params={"key": api_key},
            json={"contents": [{"parts": [{"text": prompt}]}]},
            timeout=30,
        )
        # Handle HTTP errors WITHOUT raise_for_status: its message embeds the
        # full request URL, which includes ?key=... and would leak the API key.
        if resp.status_code != 200:
            msg = {
                429: "rate limit reached — try again shortly",
                503: "service temporarily unavailable — try again shortly",
                500: "service error — try again shortly",
                403: "request rejected (check API key configuration)",
                400: "bad request",
            }.get(resp.status_code, f"HTTP {resp.status_code}")
            return _fallback_analysis(
                ticker, kpis_rl, kpis_bnh, trades,
                note=f"AI analysis unavailable: {msg}.",
            )
        data = resp.json()

        # Extract text from the first candidate's first part.
        raw_text = ""
        try:
            raw_text = data["candidates"][0]["content"]["parts"][0]["text"]
        except (KeyError, IndexError, TypeError):
            raw_text = ""

        if not raw_text:
            raise ValueError(
                "Gemini returned an empty response. This usually happens due "
                "to safety filters or an invalid API key."
            )
        raw_text = raw_text.strip()
        # Strip markdown code fences if present
        raw_text = re.sub(r"^```(?:json)?", "", raw_text).strip()
        raw_text = re.sub(r"```$", "", raw_text).strip()

        analysis = json.loads(raw_text)
        # Ensure a grade exists for the medallion, derived from real KPIs,
        # so the AI path renders the same performance-graded verdict UI.
        if "grade" not in analysis:
            analysis["grade"] = _compute_grade(kpis_rl, kpis_bnh, trades)
        analysis["score"] = _compute_score(kpis_rl, kpis_bnh, trades)
        return analysis

    except json.JSONDecodeError:
        # Response was not valid JSON — extract what we can
        return _fallback_analysis(
            ticker, kpis_rl, kpis_bnh, trades, note="AI response could not be parsed."
        )
    except Exception as e:
        # Defensive: scrub anything that looks like an API key or the request
        # URL from the exception text before showing it, so a key can never
        # leak into the UI even from an unexpected error path.
        safe = re.sub(r"key=[\w\-]+", "key=***", str(e))
        safe = re.sub(r"https?://\S+", "[endpoint]", safe)
        # Keep it short and generic
        safe = safe[:120]
        return _fallback_analysis(
            ticker, kpis_rl, kpis_bnh, trades,
            note=f"AI analysis unavailable: {safe}",
        )


def _compute_score(kpis_rl: dict, kpis_bnh: dict, trades: int) -> int:
    def parse_pct(v):
        try:
            return float(str(v).replace("%", "").replace("+", "").strip())
        except Exception:
            return 0.0
    rl_ret = parse_pct(kpis_rl.get("Total Return", "0%"))
    bnh_ret = parse_pct(kpis_bnh.get("Total Return", "0%"))
    rl_sharpe = parse_pct(kpis_rl.get("Sharpe Ratio", "0"))
    alpha = rl_ret - bnh_ret
    if trades == 0:
        return 0
    # Alpha component: 0 at -20%, 50 at +20% (capped)
    alpha_score = max(0.0, min(50.0, (alpha + 20.0) / 40.0 * 50.0))
    # Sharpe component: 0 at <=0, 40 at >=2.0 (capped)
    sharpe_score = max(0.0, min(40.0, rl_sharpe / 2.0 * 40.0))
    # Participation: up to 10 for a sane (non-degenerate) trade count
    part_score = 10.0 if 2 <= trades <= 80 else (5.0 if trades > 0 else 0.0)
    return int(round(alpha_score + sharpe_score + part_score))


def _compute_grade(kpis_rl: dict, kpis_bnh: dict, trades: int) -> str:
    """Grade the agent A–F from real KPIs. Single source of truth for both
    the AI and fallback paths."""
    def parse_pct(v):
        try:
            return float(str(v).replace("%", "").replace("+", "").strip())
        except Exception:
            return 0.0
    rl_ret = parse_pct(kpis_rl.get("Total Return", "0%"))
    bnh_ret = parse_pct(kpis_bnh.get("Total Return", "0%"))
    rl_sharpe = parse_pct(kpis_rl.get("Sharpe Ratio", "0"))
    alpha = rl_ret - bnh_ret
    if trades == 0:
        return "F"
    if alpha >= 5.0 and rl_sharpe >= 1.0:
        return "A"
    if alpha >= 0.0 and rl_sharpe >= 0.5:
        return "B"
    if alpha >= -5.0:
        return "C"
    return "D"


def _fallback_analysis(
    ticker: str,
    kpis_rl: dict,
    kpis_bnh: dict,
    trades: int,
    note: str = "",
) -> dict:
    def parse_pct(v):
        try:
            return float(str(v).replace("%", "").replace("+", "").strip())
        except Exception:
            return 0.0

    rl_ret = parse_pct(kpis_rl.get("Total Return", "0%"))
    bnh_ret = parse_pct(kpis_bnh.get("Total Return", "0%"))
    rl_sharpe = parse_pct(kpis_rl.get("Sharpe Ratio", "0"))
    rl_dd = parse_pct(kpis_rl.get("Max Drawdown", "0%"))
    outperform = rl_ret > bnh_ret

    ticker_ctx = TICKER_CONTEXT.get(ticker, "")

    summary = (
        f"The PPO agent on {ticker} achieved a total return of "
        f"{kpis_rl.get('Total Return', 'N/A')} vs Buy & Hold's "
        f"{kpis_bnh.get('Total Return', 'N/A')} over the evaluation period. "
        f"{'The agent outperformed the passive benchmark.' if outperform else 'The agent underperformed the passive benchmark.'}"
    )

    strengths = []
    # A non-participating agent (zero trades) has no genuine strengths — its
    # flat 0% drawdown and untouched capital are artifacts of inaction, not
    # skill. Only credit strengths when the agent actually traded.
    if trades > 0:
        if trades > 20:
            strengths.append(
                f"Active participation: {trades} trades show the agent engaged with the market rather than ignoring it."
            )
        if rl_sharpe > 0.5:
            strengths.append(
                f"Positive Sharpe Ratio of {kpis_rl.get('Sharpe Ratio')} indicates the agent generated risk-adjusted returns above the risk-free rate."
            )
        if outperform:
            strengths.append(
                f"Beat Buy & Hold by {rl_ret - bnh_ret:.1f}% — the policy added value over passively holding the asset."
            )
        if rl_dd > -25:
            strengths.append(
                f"Controlled drawdown of {kpis_rl.get('Max Drawdown')} suggests the reward function's drawdown penalty was effective."
            )

    if not strengths:
        if trades == 0:
            strengths.append(
                "None — the agent never entered the market, so there is no behaviour to credit. The flat capital and zero drawdown reflect inaction, not risk management."
            )
        else:
            strengths.append(
                "The agent traded but produced no clear edge; no individual metric stands out as a genuine strength."
            )

    weaknesses = []
    if rl_ret < bnh_ret:
        weaknesses.append(
            f"Underperformed Buy & Hold by {bnh_ret - rl_ret:.1f}% — the added complexity did not translate to returns."
        )
    if rl_sharpe < 0:
        weaknesses.append(
            "Negative Sharpe Ratio means the strategy's volatility was not compensated by returns."
        )
    if trades == 0:
        weaknesses.append(
            "Zero executed trades indicates a degenerate policy — the agent learned to always attempt invalid actions."
        )
    weaknesses.append(
        "Limited out-of-sample generalisation due to the 2015–2022 training window not capturing recent market regimes."
    )

    failure_diagnosis = (
        f"The agent's performance on {ticker} reflects a distribution shift between "
        f"the training period (2015–2022) and the test period. {ticker_ctx} "
        f"The PPO agent optimises for a Differential Sharpe Ratio reward which "
        f"penalises volatility — in a strongly trending market this causes the agent "
        f"to exit positions too early, missing sustained directional moves that a "
        f"simple Buy & Hold captures fully."
    )

    improvements = [
        "Retrain with data through 2026 to include recent market regimes in the training distribution.",
        "Increase the entropy coefficient (ent_coef) to 0.05+ to prevent premature policy collapse.",
        "Add a regime detection module (HMM or volatility clustering) to switch strategies dynamically.",
        "Implement online learning or periodic retraining to adapt to distribution shift in production.",
    ]

    # ── Performance-graded verdict (driven by actual numbers) ─────────
    # Grade on three axes: alpha vs Buy & Hold, risk-adjusted return (Sharpe),
    # and whether the agent actually traded coherently. No clichés — the
    # verdict states the grade, the numbers behind it, and the deployment call.
    alpha = rl_ret - bnh_ret
    if trades == 0:
        grade, headline, deploy = (
            "F",
            f"The agent never traded on {ticker} — it sat in cash for the entire window and returned {rl_ret:+.1f}% against Buy & Hold's {bnh_ret:+.1f}%.",
            "Not deployable. A non-participating policy has nothing to deploy.",
        )
    elif alpha >= 5.0 and rl_sharpe >= 1.0:
        grade, headline, deploy = (
            "A",
            f"The agent beat Buy & Hold by {alpha:+.1f}% on {ticker} ({rl_ret:+.1f}% vs {bnh_ret:+.1f}%) at a Sharpe of {rl_sharpe:.2f}, across {trades} trades.",
            "Promising enough to paper-trade forward on unseen data before any live capital.",
        )
    elif alpha >= 0.0 and rl_sharpe >= 0.5:
        grade, headline, deploy = (
            "B",
            f"The agent edged Buy & Hold by {alpha:+.1f}% on {ticker} ({rl_ret:+.1f}% vs {bnh_ret:+.1f}%) at a Sharpe of {rl_sharpe:.2f}, but the margin is thin.",
            "Not ready for live capital — the edge is too small to survive real costs and slippage. Validate across more regimes first.",
        )
    elif alpha >= -5.0:
        grade, headline, deploy = (
            "C",
            f"The agent tracked Buy & Hold within {abs(alpha):.1f}% on {ticker} ({rl_ret:+.1f}% vs {bnh_ret:+.1f}%) — it is effectively replicating the benchmark, not beating it.",
            "Not worth deploying over simply holding the asset, which is cheaper and simpler.",
        )
    else:
        grade, headline, deploy = (
            "D",
            f"The agent underperformed Buy & Hold by {abs(alpha):.1f}% on {ticker} ({rl_ret:+.1f}% vs {bnh_ret:+.1f}%) at a Sharpe of {rl_sharpe:.2f}.",
            "Do not deploy. The added complexity actively cost return versus just holding.",
        )

    verdict = f"{headline} {deploy}"

    result = {
        "summary": summary,
        "strengths": strengths,
        "weaknesses": weaknesses,
        "failure_diagnosis": failure_diagnosis,
        "improvements": improvements,
        "verdict": verdict,
        "grade": grade,
        "score": _compute_score(kpis_rl, kpis_bnh, trades),
    }

    if note:
        result["note"] = note

    return result


def format_analysis_as_html(analysis: dict, ticker: str, theme: dict) -> str:
    # ── Claude identity palette (warm, editorial) ─────────────────────
    ink = "#1F1E1C"        # near-black warm ink
    muted = "#6E6B65"      # secondary text
    clay = "#C96442"       # the one bold accent — Claude's signature rust
    ivory = "#FAF9F5"      # page surface
    cream = "#F0EEE6"      # card surface
    line = "#E3DFD3"       # hairline warm border
    on_clay = "#FFFFFF"
    grn = "#3D8C5F"        # desaturated to sit in the warm palette
    amb = "#C7901F"
    red = "#BC4633"

    # Map legacy names used below
    primary, secondary, tertiary = ink, muted, clay
    neutral, surface, on_primary = ivory, cream, on_clay

    SERIF = "'IBM Plex Sans',sans-serif"
    SANS = "'IBM Plex Sans',sans-serif"
    MONO = "'IBM Plex Mono',monospace"

    # Performance score 0–100 → accent color + label
    score = int(analysis.get("score", 0))
    if score >= 75:
        grade_color, grade_label = grn, "Strong"
    elif score >= 55:
        grade_color, grade_label = grn, "Marginal edge"
    elif score >= 40:
        grade_color, grade_label = amb, "Matches benchmark"
    elif score > 0:
        grade_color, grade_label = red, "Underperforms"
    else:
        grade_color, grade_label = red, "No participation"

    def md(t: str) -> str:
        if not isinstance(t, str):
            return t
        t = re.sub(r"\*\*(.*?)\*\*", r"<b>\1</b>", t)
        t = re.sub(r"\*(.*?)\*", r"<i>\1</i>", t)
        t = re.sub(
            r"`(.*?)`",
            f"<code style=\"background:rgba(201,100,66,0.08);padding:2px 6px;border-radius:4px;font-family:{MONO};font-size:0.88em;color:{clay}\">\\1</code>",
            t,
        )
        return t

    def icon(name: str) -> str:
        return {
            "cpu": "◆", "calendar": "", "message-square": "",
            "trending-up": "↑", "alert-triangle": "△", "bug": "◈",
            "flowchart": "", "target": "", "chevron-right": "→",
            "check": "✓", "x": "✕",
        }.get(name, "")

    def pattern(kind: str, color: str) -> str:
        """Light geometric SVG pattern, absolutely positioned, very low opacity.
        Sits behind card content as a subtle texture."""
        pid = f"p{kind}{abs(hash(kind+color))%9999}"
        defs = {
            "grid": f'<pattern id="{pid}" width="14" height="14" patternUnits="userSpaceOnUse"><path d="M14 0H0V14" fill="none" stroke="{color}" stroke-width="0.5"/></pattern>',
            "dots": f'<pattern id="{pid}" width="12" height="12" patternUnits="userSpaceOnUse"><circle cx="2" cy="2" r="1" fill="{color}"/></pattern>',
            "diag": f'<pattern id="{pid}" width="10" height="10" patternUnits="userSpaceOnUse"><path d="M0 10L10 0" stroke="{color}" stroke-width="0.5"/></pattern>',
            "tri": f'<pattern id="{pid}" width="16" height="16" patternUnits="userSpaceOnUse"><path d="M8 2L14 13H2Z" fill="none" stroke="{color}" stroke-width="0.5"/></pattern>',
        }.get(kind, "")
        return (
            f'<svg width="100%" height="100%" style="position:absolute;inset:0;'
            f'opacity:0.06;pointer-events:none" aria-hidden="true">'
            f'<defs>{defs}</defs><rect width="100%" height="100%" fill="url(#{pid})"/></svg>'
        )

    def eyebrow(text: str, color: str) -> str:
        return (
            f"<p style=\"font-family:{MONO};font-size:10px;color:{color};"
            f"text-transform:uppercase;letter-spacing:0.14em;margin:0 0 12px;"
            f"font-weight:500;position:relative\">{text}</p>"
        )

    def card(
        title: str, content: str, icon_name: str, color: str, subtitle: str = ""
    ) -> str:
        return f"""<div style="
            position:relative;overflow:hidden;
            background:{cream};
            border:1px solid {line};
            border-radius:0;
            padding:20px 22px;
            margin-bottom:16px;
        ">
            {pattern("grid", color)}
            <div style="position:relative">
            {eyebrow(title, color)}
            {content}
            {"<p style=\"font-family:" + MONO + ";font-size:10px;color:" + muted + ';margin-top:10px">' + subtitle + "</p>" if subtitle else ""}
            </div>
        </div>"""

    def get_icon_name(color: str) -> str:
        if color == grn:
            return "check"
        elif color == amb:
            return "alert-triangle"
        else:
            return "x"

    def bullet_list(items: list, color: str) -> str:
        icon_char = icon(get_icon_name(color))
        lis = "".join(
            f"<li style=\"margin-bottom:10px;color:{ink};padding-left:22px;position:relative;"
            f"font-family:{SANS};font-size:13.5px;line-height:1.6\">"
            f'<span style="position:absolute;left:0;top:1px;color:{color};font-weight:700">{icon_char}</span>'
            f"{md(item)}</li>"
            for item in items
        )
        return f'<ul style="padding-left:0;margin:0;list-style:none">{lis}</ul>'

    def flow_section(title: str, items: list) -> str:
        nodes = ""
        for i, item in enumerate(items):
            nodes += f"""<div style="display:flex;align-items:flex-start;gap:14px;margin-bottom:14px;position:relative">
                <div style="background:{clay};color:{on_clay};width:26px;height:26px;border-radius:0;display:flex;align-items:center;justify-content:center;font-family:{MONO};font-size:11px;font-weight:600;flex-shrink:0">{i + 1}</div>
                <div style="flex:1">
                    <p style="font-family:{SANS};font-size:13.5px;color:{ink};margin:2px 0 0;line-height:1.55">{md(item)}</p>
                </div>
            </div>"""
        return f"""<div style="position:relative;overflow:hidden;background:{cream};border:1px solid {line};border-radius:0;padding:20px 22px;margin-bottom:16px">
            {pattern("diag", clay)}
            <div style="position:relative">
            {eyebrow("How to improve it", clay)}
            {nodes}
            </div>
        </div>"""

    note_html = ""
    if "note" in analysis:
        note_html = f"<p style=\"font-family:{MONO};font-size:11px;color:{muted};font-style:italic;margin-top:14px\">{analysis['note']}</p>"

    strengths_html = bullet_list(analysis.get("strengths", []), grn)
    weaknesses_html = bullet_list(analysis.get("weaknesses", []), amb)
    improvements_flow = flow_section("How to improve it", analysis.get("improvements", []))

    grade_medallion = f"""
        <div style="display:flex;flex-direction:column;align-items:center;justify-content:center;
            width:88px;height:88px;flex-shrink:0;border-radius:0;
            background:{grade_color}14;border:2px solid {grade_color}">
            <span style="font-family:{MONO};font-size:34px;font-weight:600;color:{grade_color};line-height:1">{score}</span>
            <span style="font-family:{MONO};font-size:9px;color:{grade_color};letter-spacing:0.1em;margin-top:2px">/ 100</span>
        </div>"""

    html = f"""<!DOCTYPE html>
<html>
<head>
    <link rel="stylesheet" href="https://fonts.googleapis.com/css2?family=IBM+Plex+Sans:wght@400;500;600;700&family=IBM+Plex+Mono:wght@400;500&display=swap">
</head>
<body style="font-family:{SANS};background:{ivory};color:{ink};margin:0;padding:0">
<div style="
    background:{ivory};
    border:1px solid {line};
    border-radius:0;
    padding:28px 30px;
    margin-top:12px;
    max-width:920px;
">

    <!-- Header -->
    <div style="display:flex;align-items:baseline;justify-content:space-between;gap:16px;margin-bottom:26px;padding-bottom:18px;border-bottom:1px solid {line}">
        <div>
            <p style="font-family:{MONO};font-size:10px;color:{clay};text-transform:uppercase;letter-spacing:0.16em;margin:0 0 6px;font-weight:500">Agent performance review</p>
            <h1 style="font-family:{SANS};font-size:24px;font-weight:600;color:{ink};margin:0;letter-spacing:-0.01em">{ticker} <span style="color:{muted};font-weight:400">·</span> PPO trading agent</h1>
        </div>
        <span style="font-family:{MONO};font-size:10px;color:{muted};letter-spacing:0.05em;white-space:nowrap">Reviewed by Claude</span>
    </div>

    <!-- Verdict — signature: performance score -->
    <div style="position:relative;overflow:hidden;background:{cream};border:1px solid {line};border-left:3px solid {grade_color};border-radius:0;padding:22px 24px;margin-bottom:20px;display:flex;align-items:center;gap:22px">
        {pattern("dots", grade_color)}
        <div style="position:relative;display:flex;align-items:center;gap:22px;width:100%">
        {grade_medallion}
        <div style="flex:1">
            <p style="font-family:{MONO};font-size:10px;color:{grade_color};text-transform:uppercase;letter-spacing:0.14em;margin:0 0 6px;font-weight:600">Performance score · {grade_label}</p>
            <p style="font-family:{SANS};font-size:15px;font-weight:400;color:{ink};margin:0;line-height:1.55">{md(analysis.get("verdict", ""))}</p>
        </div>
        </div>
    </div>

    <!-- Summary -->
    <div style="position:relative;overflow:hidden;background:{cream};border:1px solid {line};border-radius:0;padding:20px 22px;margin-bottom:16px">
        {pattern("grid", clay)}
        <div style="position:relative">
        {eyebrow("What happened", clay)}
        <p style="font-family:{SANS};font-size:14px;color:{ink};margin:0;line-height:1.62">{md(analysis.get("summary", ""))}</p>
        </div>
    </div>

    <!-- Strengths & Weaknesses -->
    <div style="display:grid;grid-template-columns:1fr 1fr;gap:16px;margin-bottom:16px">
        <div style="position:relative;overflow:hidden;background:{cream};border:1px solid {line};border-radius:0;padding:20px 22px">
            {pattern("dots", grn)}
            <div style="position:relative">
            {eyebrow("What worked", grn)}
            {strengths_html}
            </div>
        </div>
        <div style="position:relative;overflow:hidden;background:{cream};border:1px solid {line};border-radius:0;padding:20px 22px">
            {pattern("tri", amb)}
            <div style="position:relative">
            {eyebrow("What held it back", amb)}
            {weaknesses_html}
            </div>
        </div>
    </div>

    <!-- Diagnosis -->
    <div style="position:relative;overflow:hidden;background:{cream};border:1px solid {line};border-radius:0;padding:20px 22px;margin-bottom:16px">
        {pattern("diag", red)}
        <div style="position:relative">
        {eyebrow("Why it behaved this way", red)}
        <p style="font-family:{SANS};font-size:13.5px;color:{ink};margin:0;line-height:1.65">{md(analysis.get("failure_diagnosis", ""))}</p>
        </div>
    </div>

    <!-- Improvements -->
    {improvements_flow}

    {note_html}

    <p style="font-family:{MONO};font-size:10px;color:{muted};margin:18px 0 0;text-align:center;letter-spacing:0.04em">
        This is research analysis, not financial advice. Validate on unseen data before risking capital.
    </p>
</div>
</body>
</html>
"""
    return html