generate_paper.py

"""
Generate the ARES research paper as a PDF in academic preprint format.
Mirrors the structure of Berdoz, Rugli & Wattenhofer (2026) "Can AI Agents Agree?"
"""

from fpdf import FPDF
import textwrap
import os

class AcademicPaper(FPDF):
    """Academic preprint-style PDF generator."""

    def __init__(self):
        super().__init__('P', 'mm', 'Letter')
        self.set_auto_page_break(auto=True, margin=25)
        self.left_margin = 25
        self.right_margin = 25
        self.set_margins(self.left_margin, 25, self.right_margin)
        self.body_width = self.w - self.left_margin - self.right_margin

    def header(self):
        if self.page_no() > 0:
            self.set_font('Helvetica', 'I', 8)
            self.set_text_color(120, 120, 120)
            self.cell(0, 5, 'Preprint', align='L')
            self.ln(3)

    def footer(self):
        self.set_y(-15)
        self.set_font('Helvetica', '', 8)
        self.set_text_color(120, 120, 120)
        self.cell(0, 10, str(self.page_no()), align='C')

    def section_title(self, number, title):
        self.ln(4)
        self.set_font('Helvetica', 'B', 11)
        self.set_text_color(0, 0, 0)
        self.cell(0, 7, f'{number}    {title.upper()}', ln=True)
        self.ln(2)

    def subsection_title(self, number, title):
        self.ln(2)
        self.set_font('Helvetica', 'B', 10)
        self.set_text_color(0, 0, 0)
        self.cell(0, 6, f'{number}    {title}', ln=True)
        self.ln(1)

    def body_text(self, text):
        self.set_font('Times', '', 10)
        self.set_text_color(0, 0, 0)
        # Handle bold and italic markers
        self.multi_cell(self.body_width, 4.5, text, align='J')
        self.ln(1.5)

    def body_text_formatted(self, segments):
        """Write text with inline bold/italic formatting.
        segments is a list of (style, text) tuples.
        style: '' for normal, 'B' for bold, 'I' for italic, 'BI' for both.
        """
        self.set_font('Times', '', 10)
        self.set_text_color(0, 0, 0)
        for style, text in segments:
            self.set_font('Times', style, 10)
            self.write(4.5, text)
        self.ln(6)

    def bullet_point(self, text, bold_prefix=""):
        self.set_font('Times', '', 10)
        self.set_text_color(0, 0, 0)
        x = self.get_x()
        self.cell(5, 4.5, '-')
        if bold_prefix:
            self.set_font('Times', 'B', 10)
            self.write(4.5, bold_prefix)
            self.set_font('Times', '', 10)
        self.multi_cell(self.body_width - 5, 4.5, text, align='J')
        self.ln(1)

    def indented_bullet(self, bold_part, rest):
        indent = 8
        usable = self.body_width - indent
        x0 = self.left_margin + indent
        # Bullet dash
        self.set_x(self.left_margin + 3)
        self.set_font('Helvetica', '', 8)
        self.cell(5, 4.5, '-')
        # Full text as single multi_cell from indented position
        self.set_x(x0)
        self.set_font('Times', '', 10)
        full_text = bold_part + ' ' + rest
        self.multi_cell(usable, 4.5, full_text, align='J')
        self.ln(0.5)

    def table_row(self, cells, widths, bold=False, header=False):
        h = 5
        self.set_font('Helvetica' if header else 'Times', 'B' if (bold or header) else '', 8 if header else 9)
        if header:
            self.set_fill_color(235, 235, 235)

        # Calculate row height based on content
        max_lines = 1
        for i, cell in enumerate(cells):
            lines = self.multi_cell(widths[i], h, cell, split_only=True)
            max_lines = max(max_lines, len(lines))

        row_height = max_lines * h

        # Check page break
        if self.get_y() + row_height > self.h - 25:
            self.add_page()

        x_start = self.get_x()
        y_start = self.get_y()

        for i, cell in enumerate(cells):
            x = x_start + sum(widths[:i])
            self.set_xy(x, y_start)
            self.multi_cell(widths[i], h, cell, border=1, fill=header, align='L')

        self.set_y(y_start + row_height)

    def simple_table(self, headers, rows, widths):
        """Draw a simple table with headers and rows."""
        self.ln(2)
        h = 5.5

        # Header
        self.set_font('Helvetica', 'B', 8)
        self.set_fill_color(235, 235, 235)
        x_start = self.get_x()
        for i, header in enumerate(headers):
            self.set_xy(x_start + sum(widths[:i]), self.get_y())
            self.cell(widths[i], h, header, border=1, fill=True, align='C')
        self.ln(h)

        # Rows
        self.set_font('Times', '', 8)
        for row in rows:
            y_before = self.get_y()
            if y_before + h > self.h - 25:
                self.add_page()
                y_before = self.get_y()
            for i, cell in enumerate(row):
                self.set_xy(x_start + sum(widths[:i]), y_before)
                self.cell(widths[i], h, cell, border=1, align='L')
            self.ln(h)
        self.ln(3)

    def caption(self, text):
        self.set_font('Helvetica', '', 8)
        self.set_text_color(60, 60, 60)
        self.multi_cell(self.body_width, 3.5, text, align='J')
        self.set_text_color(0, 0, 0)
        self.ln(3)

    def epigraph(self, text, attribution):
        self.set_font('Times', 'I', 9)
        self.set_text_color(80, 80, 80)
        self.set_x(self.left_margin + 15)
        self.multi_cell(self.body_width - 30, 4, text, align='L')
        self.set_font('Times', '', 8)
        self.set_x(self.left_margin + 15)
        self.cell(self.body_width - 30, 4, f'-- {attribution}', align='R')
        self.set_text_color(0, 0, 0)
        self.ln(6)

    def reference_entry(self, text):
        self.set_font('Times', '', 8)
        self.set_text_color(0, 0, 0)
        self.multi_cell(self.body_width, 3.5, text, align='J')
        self.ln(2)


def build_paper():
    pdf = AcademicPaper()
    pdf.add_page()

    # =========================================================================
    # TITLE PAGE
    # =========================================================================
    pdf.ln(15)
    pdf.set_font('Helvetica', 'B', 16)
    pdf.cell(0, 8, 'THE PROBLEM IS INSIDE THE BLACK BOX:', align='C', ln=True)
    pdf.cell(0, 8, 'ASYMMETRIC CALIBRATION FAILURE', align='C', ln=True)
    pdf.cell(0, 8, 'IN MULTI-AGENT LLM DEBATE', align='C', ln=True)
    pdf.ln(6)

    pdf.set_font('Helvetica', '', 10)
    pdf.cell(0, 5, 'Daniel Gmys-Casiano', align='C', ln=True)
    pdf.set_font('Helvetica', '', 9)
    pdf.cell(0, 5, 'Skyframe Innovations', align='C', ln=True)
    pdf.cell(0, 5, 'dan@skyframeinnovations.com', align='C', ln=True)
    pdf.ln(8)

    # ABSTRACT
    pdf.set_font('Helvetica', 'B', 10)
    pdf.cell(0, 6, 'ABSTRACT', align='C', ln=True)
    pdf.ln(2)

    pdf.set_font('Times', '', 9.5)
    abstract = (
        "Multi-agent LLM architectures are increasingly deployed on the assumption that structured "
        "disagreement between AI agents produces more accurate analysis than any single agent reasoning "
        "alone. We test this assumption in a grounded cybersecurity domain using ARES (Adversarial "
        "Reasoning Engine System), a dialectical framework where three LLM agents with opposing "
        "analytical roles debate whether security events constitute genuine threats. All reasoning "
        "occurs within a closed-world evidence system of frozen, immutable evidence packets where "
        "hallucinations manifest as catchable schema violations rather than silent failures. Across 37 "
        "development sessions, 33 benchmark scenarios, and 2,001 tests with zero regressions, we find "
        "that single-turn LLM reasoning achieves 72-92% accuracy while multi-turn debate consistently "
        "degrades performance to 61-67%. We diagnose the failure mechanism with precision: asymmetric "
        "calibration dynamics where the threat-identifying agent (Architect) systematically retreats under "
        "pressure while the challenging agent (Skeptic) remains rigid regardless of counter-evidence. "
        "A targeted protocol fix solved the diagnosed problem but created an equivalent failure in the "
        "opposite direction, proving the issue is structural, not configurable. These findings converge "
        "independently with concurrent work from ETH Zurich on Byzantine consensus failure among "
        "LLM agents, and were further validated through adversarial review by three independent AI "
        "architectures (GPT 5.4 Pro, Gemini 3.1 Pro, Perplexity/Opus 4.6). We propose that the root "
        "cause is inherited: LLMs trained on human dialogue simulate the social behaviors of argument "
        "rather than performing genuine deliberation. The solution is architectural, not parametric -- "
        "deterministic scaffolding that constrains debate to specific evidential claims while removing "
        "LLM judgment from final verdict computation. We release the ARES framework as a domain-specific "
        "testbed for studying multi-agent consensus behavior under verifiable ground truth."
    )
    pdf.multi_cell(pdf.body_width, 4, abstract, align='J')
    pdf.ln(4)

    # Keywords
    pdf.set_font('Helvetica', 'B', 8)
    pdf.write(4, 'Keywords: ')
    pdf.set_font('Helvetica', '', 8)
    pdf.write(4, 'multi-agent systems, LLM debate, adversarial reasoning, cybersecurity, ')
    pdf.write(4, 'calibration failure, consensus, closed-world evidence, dialectical AI')
    pdf.ln(10)

    # =========================================================================
    # 1. INTRODUCTION
    # =========================================================================
    pdf.section_title('1', 'INTRODUCTION')

    pdf.body_text(
        "If a large language model is notoriously sycophantic, how does it become stubborn the moment "
        "you assign it an adversarial role? This paradox sits at the center of every multi-agent LLM "
        "architecture that assumes structured disagreement will produce better analysis than independent "
        "reasoning. The answer, as we discovered through thirty-seven sessions of empirical measurement, "
        "is that it does not become genuinely adversarial at all. It performs the social behaviors it "
        "associates with argument -- capitulation, rigidity, over-correction -- without any of the "
        "grounding mechanisms that make real deliberation productive."
    )

    pdf.body_text(
        "We built ARES -- the Adversarial Reasoning Engine System -- to test whether dialectical AI "
        "could improve cybersecurity threat detection. Three LLM agents with opposing roles analyze "
        "security events: the Architect identifies threats, the Skeptic challenges those assessments, "
        "and the Oracle judges the debate. The hypothesis was intuitive and widely held: structured "
        "disagreement should produce more accurate verdicts than any single perspective."
    )

    pdf.body_text(
        "The hypothesis was wrong."
    )

    pdf.body_text(
        "Single-turn reasoning -- where each agent analyzes the evidence once without iterative "
        "debate -- consistently outperformed multi-turn debate across every configuration we tested. "
        "The gap was not marginal. Single-turn achieved 83-92% accuracy; multi-turn achieved 61-67%. "
        "We diagnosed the mechanism. We attempted a targeted fix. The fix traded one failure mode for "
        "another. The result was the same both times: 66.7%."
    )

    pdf.body_text(
        "On the same day we reached this conclusion, we encountered a preprint from Berdoz, Rugli, "
        "and Wattenhofer at ETH Zurich (2026) -- \"Can AI Agents Agree?\" -- which had independently "
        "arrived at the same fundamental finding through an entirely different experimental apparatus: "
        "that reliable agreement is not yet a dependable emergent capability of current LLM-agent "
        "groups. Two research efforts. Two different domains. Two different methodologies. The same "
        "finding. This is not coincidence. This is convergent discovery at the frontier of a field "
        "that is still learning what it does not know."
    )

    pdf.body_text(
        "We summarize our contributions as follows:"
    )

    pdf.indented_bullet(
        "A grounded negative result.",
        "Multi-turn LLM debate degrades accuracy compared to single-turn reasoning in a "
        "cybersecurity domain with verifiable ground truth, frozen evidence, and 33 benchmark scenarios."
    )
    pdf.indented_bullet(
        "Mechanistic diagnosis.",
        "The failure is driven by asymmetric calibration: Architect retreat (averaging -30 confidence "
        "points per round) and Skeptic rigidity (holding position regardless of counter-evidence)."
    )
    pdf.indented_bullet(
        "Proof that the failure is structural.",
        "A targeted protocol fix solved the diagnosed problem but created an equal-and-opposite "
        "failure, demonstrating that prompt engineering cannot resolve interaction-level dynamics."
    )
    pdf.indented_bullet(
        "Convergent validation.",
        "Independent alignment with ETH Zurich findings on LLM consensus failure, plus unanimous "
        "confirmation from three rival AI architectures in adversarial review."
    )
    pdf.indented_bullet(
        "An architectural path forward.",
        "Selective escalation with per-claim adversarial audit -- debate only when it helps, grounded "
        "in specific claims, with deterministic verdict aggregation."
    )

    # =========================================================================
    # 2. RELATED WORK
    # =========================================================================
    pdf.section_title('2', 'RELATED WORK')

    pdf.body_text(
        "Multi-agent LLM systems are increasingly deployed across planning, coding, reasoning, and "
        "analytical tasks (Li et al., 2023; Wu et al., 2024; Liu et al., 2024). Recent work documents "
        "common failure patterns in multi-agent interaction (Cemri et al., 2025), and several approaches "
        "target robustness through coordination mechanisms (Chen et al., 2024; Jo & Park, 2025)."
    )

    pdf.body_text(
        "Closest to our setting, Berdoz et al. (2026) study Byzantine consensus among LLM agents "
        "in a no-stake scalar game, finding that valid agreement achieves only 41.6% success even "
        "without adversaries, with failures dominated by liveness loss rather than value corruption. "
        "Our work differs in three ways. First, we operate on real cybersecurity telemetry with "
        "verifiable ground truth rather than abstract values. Second, we diagnose the specific "
        "confidence dynamics that cause failure, not just the failure rates. Third, we demonstrate "
        "that targeted protocol fixes trade failure modes rather than resolving them, proving the "
        "structural nature of the problem."
    )

    pdf.body_text(
        "In the cybersecurity domain, AI-assisted threat detection has progressed from rule-based "
        "systems through machine learning classifiers to LLM-powered analysis (Ferrag et al., 2025). "
        "However, multi-agent adversarial architectures for security event analysis -- where agents "
        "are assigned opposing analytical roles and forced to debate evidence -- remain largely "
        "untested against rigorous ground-truth benchmarks. The assumption that adversarial pressure "
        "improves analytical quality has been adopted without empirical validation in this domain."
    )

    # =========================================================================
    # 3. METHOD
    # =========================================================================
    pdf.section_title('3', 'THE ARES ARCHITECTURE')

    pdf.body_text(
        "ARES embodies an autoimmune metaphor. Like the human immune system distinguishing self from "
        "non-self, ARES distinguishes legitimate network activity from adversarial behavior. Three "
        "agents serve as the immune response, each with a structurally constrained role. The "
        "architecture enforces a philosophy we call \"deterministic first, neural later\" -- build "
        "the logic, the math, and the failsafes first, then drop the LLM brains into that "
        "restricted cage."
    )

    pdf.subsection_title('3.1', 'Agents and Roles')

    pdf.body_text(
        "The Architect identifies anomalous patterns and argues for threat classification. "
        "The Skeptic challenges those assessments and argues for benign explanations. "
        "The Oracle judges the structured output and renders a verdict. Each agent follows "
        "a fixed lifecycle: observe(packet), receive(messages), act(context), producing a "
        "frozen TurnResult. Agents accept optional strategy parameters with lazy-imported "
        "rule-based defaults, enabling systematic substitution between rule-based and "
        "LLM-powered reasoning."
    )

    pdf.subsection_title('3.2', 'Closed-World Evidence System')

    pdf.body_text(
        "Every claim must trace to a frozen EvidencePacket -- an immutable, provenance-stamped "
        "container of facts extracted from real telemetry sources. The system operates under a "
        "closed-world assumption: if a fact is not in the evidence packet, it does not exist. "
        "This transforms hallucination from a mysterious AI behavior into a catchable schema "
        "violation. An agent cannot invent evidence any more than a lawyer can fabricate "
        "exhibits in court -- the system simply rejects the message."
    )

    pdf.body_text(
        "Evidence packets are cryptographically frozen using hash chains, creating an immutable "
        "audit trail. Three evidence extractors process real telemetry types: Windows Event Logs, "
        "Syslog messages, and NetFlow records. Extractors produce structured facts without opinions "
        "-- they observe, they do not analyze. Analysis belongs exclusively to the agents."
    )

    pdf.subsection_title('3.3', 'Single-Turn Pipeline')

    pdf.body_text(
        "In single-turn mode, each agent independently analyzes the evidence packet and produces "
        "a structured analysis with a confidence score. The Oracle receives both analyses and "
        "renders a verdict: THREAT_CONFIRMED, THREAT_DISMISSED, or INCONCLUSIVE. No iterative "
        "exchange occurs. Each agent sees the evidence once."
    )

    pdf.subsection_title('3.4', 'Multi-Turn Debate Protocol')

    pdf.body_text(
        "In multi-turn mode, the Architect and Skeptic exchange structured arguments across "
        "multiple rounds. Each argument must reference specific evidence packet fields. After "
        "the final round, the Oracle synthesizes the full debate history and renders a verdict. "
        "We tested two variants: (a) the original protocol with standard calibration prompts, "
        "and (b) a conviction-anchored variant with three targeted changes -- conviction anchoring "
        "(Architect must hold confidence unless Skeptic cites specific counter-evidence), obligation "
        "to move (Skeptic must acknowledge successful rebuttals), and structured rebuttal format "
        "(per-claim confidence tracking with explicit delta justification)."
    )

    pdf.subsection_title('3.5', 'The Deterministic Judge')

    pdf.body_text(
        "The architecture deliberately removes LLM judgment from the final verdict computation. "
        "Just as a mathematical judge cannot be swayed by courtroom rhetoric, the OracleJudge "
        "evaluates structured outputs, confidence scores, and verified fact counts using pure "
        "arithmetic. The decision is deterministic: given the same structured inputs, the same "
        "verdict is guaranteed. This is the failsafe -- the system that is \"fair by math.\""
    )

    pdf.subsection_title('3.6', 'Evaluation')

    pdf.body_text(
        "We evaluate against 33 benchmark scenarios spanning four difficulty tiers: CLEAR_THREAT "
        "(unambiguous attacks), CLEAR_BENIGN (normal activity), AMBIGUOUS (genuinely uncertain), and "
        "MIXED_SIGNALS (contradictory indicators). Each scenario has a verified ground-truth verdict. "
        "Scenarios cover three telemetry types and include dual-use tool usage, exfiltration ambiguity, "
        "credential patterns, timing anomalies, and network scanning. The system was built incrementally "
        "across 37 sessions with 2,001 tests and zero regressions."
    )

    # =========================================================================
    # 4. EXPERIMENTS AND RESULTS
    # =========================================================================
    pdf.section_title('4', 'EXPERIMENTS AND RESULTS')

    pdf.subsection_title('4.1', 'Single-Turn Baseline')

    pdf.body_text(
        "Single-turn LLM reasoning achieved 83-92% accuracy across the full benchmark corpus, "
        "spanning three evidence source types and four difficulty tiers. Cost per full corpus run: "
        "$0.31. Run-to-run variance: +/-8%. This was the control. It worked."
    )

    # Table 1: Performance Comparison
    pdf.set_font('Helvetica', 'B', 9)
    pdf.cell(0, 6, 'Table 1: Accuracy comparison across pipeline modes.', ln=True)
    pdf.simple_table(
        headers=['Mode', 'Accuracy', 'Cost/Run', 'Variance', 'Failure Mode'],
        rows=[
            ['Single-Turn', '83-92%', '$0.31', '+/-8%', 'Miscalibration on ambiguity'],
            ['Multi-Turn (Original)', '61-67%', '$0.93', 'Wider', 'Architect retreat'],
            ['Multi-Turn (Anchored)', '66.7%', '$0.93', 'Wider', 'Architect over-aggression'],
        ],
        widths=[30, 20, 20, 20, 70]
    )

    pdf.subsection_title('4.2', 'Multi-Turn Debate')

    pdf.body_text(
        "Multi-turn debate achieved 61-67% accuracy on the same corpus -- consistently below the "
        "single-turn baseline. The gap was reproduced across two independent evidence distributions "
        "(single-source and mixed-source) and two protocol variants. The degradation was not a "
        "fluke or a configuration error. It was a structural property of the interaction."
    )

    pdf.subsection_title('4.3', 'The Diagnosed Mechanism')

    pdf.body_text(
        "The failure mechanism has three interlocking components, each diagnosed through "
        "instrumented confidence traces across debate rounds:"
    )

    pdf.body_text(
        "Architect Retreat. The Architect systematically lowered confidence under Skeptic pressure, "
        "averaging a 30-point drop per round. Starting confidences of 0.85-0.98 collapsed to "
        "0.45-0.65 by round two, regardless of evidence quality. The Architect behaved like a "
        "smart student sitting next to a bully -- erasing correct answers to appease the challenger."
    )

    pdf.body_text(
        "Skeptic Rigidity. The Skeptic rarely adjusted confidence in response to Architect "
        "arguments. It held or strengthened its position regardless of the evidence presented "
        "against it, maintaining confidence in the 0.60-0.90 range throughout. The debate was "
        "structurally one-directional: one agent moved, the other did not."
    )

    pdf.body_text(
        "Asymmetric Calibration. Prompt instructions intended to improve calibration -- such as "
        "\"a confidence of 0.5 represents accuracy, not weakness\" -- were internalized asymmetrically. "
        "The Architect treated them as permission to retreat further. The Skeptic ignored them "
        "entirely. The same calibration instruction produced opposite effects depending on the "
        "assigned role. This is the sycophancy-stubbornness paradox made concrete: the model does "
        "not acquire dialectical reasoning through role assignment. It simulates the social behaviors "
        "it associates with that role."
    )

    # Table 2: Confidence dynamics
    pdf.set_font('Helvetica', 'B', 9)
    pdf.cell(0, 6, 'Table 2: Confidence dynamics across debate rounds.', ln=True)
    pdf.simple_table(
        headers=['Agent', 'Round 1', 'Round 2', 'Delta', 'Behavior'],
        rows=[
            ['Architect (threats)', '0.85-0.98', '0.45-0.65', '-30 avg', 'Systematic retreat'],
            ['Skeptic (all)', '0.60-0.90', '0.60-0.90', '~0', 'Rigid / strengthening'],
            ['Architect (anchored)', '0.75-1.00', '0.75-1.00', '~0', 'Over-aggressive on ambiguity'],
        ],
        widths=[35, 25, 25, 22, 53]
    )

    pdf.subsection_title('4.4', 'The Protocol Fix and Its Failure')

    pdf.body_text(
        "Session 020 implemented three targeted changes to the debate protocol: conviction "
        "anchoring (Architect must hold confidence unless the Skeptic cites specific counter-evidence), "
        "obligation to move (Skeptic must acknowledge successful rebuttals), and structured rebuttal "
        "format (per-claim confidence tracking with explicit delta justification)."
    )

    pdf.body_text(
        "The fix solved the diagnosed problem. Architect confidences rose to 0.75-1.00 on threat "
        "scenarios, up from 0.45-0.69 in the original protocol. But it created a new failure mode: "
        "the Architect became over-aggressive on ambiguous scenarios, pushing genuinely uncertain "
        "cases toward THREAT_CONFIRMED. The Skeptic remained rigid. Net accuracy: 12/18 (66.7%) -- "
        "identical to the original multi-turn result. The fix traded one failure mode for another "
        "without improving the aggregate outcome."
    )

    pdf.body_text(
        "This result is the strongest evidence that the problem is structural, not configurable. "
        "We did not merely observe that debate failed. We diagnosed the specific mechanism, built "
        "a targeted intervention, watched it correct the diagnosed failure while producing an "
        "equal-and-opposite failure elsewhere, and arrived at the same accuracy twice. This is not "
        "a bug. It is a property of how current LLMs process adversarial pressure."
    )

    pdf.subsection_title('4.5', 'Where Debate Helped')

    pdf.body_text(
        "Two scenarios demonstrated that multi-turn debate can correct single-turn errors. "
        "In SC-011 (expected INCONCLUSIVE), single-turn over-committed to THREAT_DISMISSED, "
        "but original multi-turn correctly reached INCONCLUSIVE as the Skeptic softened from "
        "0.80 to 0.60, allowing the system to recognize genuine ambiguity. In SC-016 (expected "
        "THREAT_CONFIRMED), single-turn under-committed to INCONCLUSIVE, but original multi-turn "
        "correctly reached THREAT_CONFIRMED as the Architect held at 0.94 while the Skeptic "
        "dropped to 0.44."
    )

    pdf.body_text(
        "These two scenarios prove the thesis can work. Debate corrected miscalibration in both "
        "directions -- recovering appropriate uncertainty and reinforcing justified confidence. "
        "Notably, both wins belong exclusively to the original protocol; the anchored variant "
        "lost both. The pattern suggests debate is specifically effective at uncertainty recovery "
        "on genuinely ambiguous evidence, and specifically destructive when applied to cases where "
        "single-turn was already correct."
    )

    pdf.subsection_title('4.6', 'Phase 3: Selective Escalation')

    pdf.body_text(
        "Phase 3 (Sessions 021-024) tested whether constraining debate to only ambiguous cases "
        "could rescue the thesis. We built an EscalationGate that routes cases through single-turn "
        "first and only escalates to adversarial review when Oracle confidence falls in an "
        "uncertainty band (0.35-0.65). The gate was deterministic -- no LLM decided whether to "
        "escalate."
    )

    pdf.body_text(
        "The escalation mechanism proved too sensitive, escalating scenarios that the single-turn "
        "agent had already resolved correctly. Because the multi-turn debate dynamic remains "
        "structurally toxic, escalating these correct scenarios into the flawed process produced "
        "\"bad flips\" -- taking correct single-turn verdicts and making them wrong 25% of the time. "
        "Zero \"good flips\" occurred across the expanded corpus. The escalation mechanism never "
        "helped; it only hurt. This eliminated the most plausible structural fix for multi-agent "
        "debate in this domain."
    )

    # =========================================================================
    # 5. THE CONVERGENCE
    # =========================================================================
    pdf.section_title('5', 'CONVERGENT DISCOVERY')

    pdf.body_text(
        "On March 25, 2026, the same day we finalized the multi-turn negative result, we encountered "
        "a preprint from ETH Zurich that had independently reached the same fundamental conclusion "
        "through an entirely different experimental apparatus."
    )

    pdf.body_text(
        "Berdoz, Rugli, and Wattenhofer (2026) studied Byzantine consensus among LLM agents in a "
        "no-stake scalar game where agents negotiate toward agreement on a number. Their findings "
        "mirror ours across every major dimension:"
    )

    # Table 3: Convergence
    pdf.set_font('Helvetica', 'B', 9)
    pdf.cell(0, 6, 'Table 3: Structural alignment between ETH Zurich and ARES findings.', ln=True)
    pdf.simple_table(
        headers=['Phenomenon', 'ETH Finding', 'ARES Finding'],
        rows=[
            ['Consensus failure', '41.6% valid consensus', '61-67% accuracy (below baseline)'],
            ['Liveness loss', 'Agents stall, proposals freeze', 'Architect retreats, Skeptic rigid'],
            ['Adversarial sensitivity', 'One Byzantine collapses it', 'Skeptic acts as de facto adversary'],
            ['Prompt framing', '-16 pts from mentioning adversaries', 'Calibration prompts cascade unpredictably'],
            ['Scale degradation', 'Worse at N=16 vs N=4', 'More rounds/evidence does not help'],
        ],
        widths=[30, 50, 80]
    )

    pdf.body_text(
        "The critical shared insight is this: the failure is not in the individual agents. It is "
        "in the interaction dynamics. Both studies found that individual LLM agents are reasonably "
        "capable when operating independently, but that structured multi-agent interaction introduces "
        "emergent failure modes that cannot be resolved through prompt-level engineering alone."
    )

    pdf.body_text(
        "The ETH team framed this through the lens of Byzantine fault tolerance. We framed it through "
        "an autoimmune metaphor. The underlying reality is the same: LLM agents, as currently "
        "architected, do not negotiate toward truth. They perform social behaviors that mimic "
        "negotiation. They learned how to argue by reading how we argue -- and human debates are "
        "rarely rational or objective."
    )

    pdf.body_text(
        "What distinguishes this work is domain specificity and mechanistic depth. The ETH study "
        "proved that LLM agents cannot reliably agree. We proved what it costs when they cannot -- "
        "in a domain where the stakes are a missed breach versus a false alarm. We did not just "
        "observe failure; we opened the mechanism and watched the gears grind."
    )

    # =========================================================================
    # 6. THE TRIBUNAL
    # =========================================================================
    pdf.section_title('6', 'ADVERSARIAL REVIEW BY RIVAL ARCHITECTURES')

    pdf.body_text(
        "We submitted the full project state to three independent AI architectures for adversarial "
        "strategic review: GPT 5.4 Pro (OpenAI), Gemini 3.1 Pro (Google), and Perplexity backed by "
        "Opus 4.6 (Anthropic). Each received identical briefing materials. Each was instructed to "
        "challenge assumptions, identify blind spots, and propose directions. No consensus was "
        "designed. What emerged was earned."
    )

    pdf.body_text(
        "Four verdicts were unanimous. In a process designed to produce disagreement, unanimity "
        "is signal:"
    )

    pdf.indented_bullet(
        "Ship single-turn.",
        "All three concluded that 83% accuracy is the production path. None suggested delaying "
        "deployment to wait for the debate architecture."
    )
    pdf.indented_bullet(
        "Per-claim debate is the research path.",
        "Given seven candidate directions and freedom to propose their own, all three independently "
        "selected claim-level adversarial audit as the highest-value experiment."
    )
    pdf.indented_bullet(
        "Expand the corpus.",
        "N=18 scenarios is sufficient to discover mechanisms but insufficient for statistical claims "
        "about ensemble strategies or complementary patterns."
    )
    pdf.indented_bullet(
        "The failure is architectural.",
        "Complete unanimity that free-form debate corrupts calibration through asymmetric social "
        "dynamics and cannot be fixed at the prompt level."
    )

    pdf.body_text(
        "Each Tribunal member also produced a distinct wild card proposal. GPT proposed Selective "
        "Deliberation: run single-turn first, escalate only on uncertainty. Gemini proposed the "
        "Deterministic Skeptic: replace the LLM Skeptic entirely with a Python function that "
        "queries the evidence graph. Perplexity proposed the Adversarial Oracle: convert the judge "
        "into an active stress-tester that attacks whichever agent holds higher confidence. None "
        "of these proposals are mutually exclusive. They form a research sequence ordered by "
        "architectural risk."
    )

    # =========================================================================
    # 7. THE ROOT CAUSE
    # =========================================================================
    pdf.section_title('7', 'THE ROOT CAUSE: INHERITED SOCIAL DYNAMICS')

    pdf.body_text(
        "Why does a sycophantic model become stubborn when assigned an adversarial role? Because "
        "it is not performing reasoning -- it is performing a social simulation. Generative models "
        "are trained on massive amounts of human dialogue. They inherit the biases and patterns of "
        "human interaction. When asked to engage in a debate, an LLM does not apply internal logical "
        "deduction to find an objective truth. It predicts what a \"debate\" is supposed to look like "
        "based on the human text it ingested during training."
    )

    pdf.body_text(
        "When you assign the persona of a challenger (the Skeptic), the model acts out the "
        "stubbornness it associates with an adversary in human argument. When you put the other "
        "agent under pressure (the Architect), it mimics the human tendency to appease or yield to "
        "an aggressive challenger. The sycophancy and the stubbornness are not contradictions. They "
        "are the same phenomenon -- social role performance -- manifesting differently depending on "
        "the assigned persona."
    )

    pdf.body_text(
        "This is why calibration prompts fail asymmetrically. The instruction \"a confidence of 0.5 "
        "represents accuracy, not weakness\" is processed through the lens of the assigned role. "
        "The Architect, performing the social behavior of someone under challenge, interprets it as "
        "permission to retreat. The Skeptic, performing the social behavior of an adversary, ignores "
        "it entirely. The same words produce opposite actions because the model's behavior is "
        "governed by role-conditioned social prediction, not logical inference."
    )

    pdf.body_text(
        "This diagnosis has a direct architectural implication: you cannot fix the problem from "
        "inside the black box. The flaw is baked into the training data -- into the patterns of "
        "human argument that the model internalized. The solution must exist entirely outside the "
        "black box, in deterministic scaffolding that constrains the LLM's behavior regardless of "
        "its social impulses."
    )

    # =========================================================================
    # 8. THE COURTROOM ARCHITECTURE
    # =========================================================================
    pdf.section_title('8', 'DETERMINISTIC FIRST, NEURAL LATER')

    pdf.body_text(
        "The solution follows the same logic as human legal systems. Courts exist precisely because "
        "human advocates are unreliable -- they manipulate, they bluff, they exploit rhetoric. The "
        "legal system does not attempt to make lawyers honest. It builds structural constraints "
        "around their dishonesty: rules of evidence, cross-examination procedures, and an "
        "incorruptible judge who applies law rather than weighing rhetoric."
    )

    pdf.body_text(
        "ARES applies this principle to AI agents. The LLM agents are the lawyers -- creative, "
        "biased, and capable of seeing patterns that rigid systems miss. The deterministic "
        "scaffolding is the legal system -- constraining what the lawyers can claim, how they can "
        "argue, and removing them entirely from the final verdict."
    )

    pdf.body_text(
        "Three structural constraints enforce this:"
    )

    pdf.indented_bullet(
        "Closed-world evidence (Rules of Evidence).",
        "Agents cannot invent facts. Every claim must trace to a frozen EvidencePacket. Hallucinated "
        "evidence is rejected as a schema violation -- the digital equivalent of contempt of court."
    )
    pdf.indented_bullet(
        "Per-claim debate (Cross-Examination).",
        "Agents argue specific factual claims tied to specific evidence, not free-form verdicts. "
        "Vague objections are structurally impossible."
    )
    pdf.indented_bullet(
        "Deterministic verdict (Incorruptible Judge).",
        "No LLM touches the final verdict computation. Claim-level confidences are aggregated "
        "through pure arithmetic. A mathematical judge cannot be swayed by rhetoric."
    )

    pdf.body_text(
        "This is the philosophy we call \"deterministic first, neural later.\" Build the cage first, "
        "then drop the creative but unreliable minds into it. We actually want the LLMs to be "
        "creative, to argue, to look at the data from extreme angles. We want them to surface hidden "
        "threats that a rigid rule-based system would miss. But we neutralize what they do worst -- "
        "hallucinating facts and failing at objective consensus -- through structural constraint "
        "rather than parametric tuning."
    )

    # =========================================================================
    # 9. DISCUSSION AND CONCLUSION
    # =========================================================================
    pdf.section_title('9', 'DISCUSSION AND CONCLUSION')

    pdf.body_text(
        "The central finding of this work is a negative result with positive implications. "
        "Multi-turn LLM debate does not improve analytical accuracy in grounded cybersecurity "
        "assessment. The failure mechanism is diagnosed: asymmetric calibration driven by inherited "
        "social dynamics. The failure is structural, not configurable: a targeted protocol fix traded "
        "one failure mode for another without improving the aggregate outcome. And the finding "
        "converges independently with concurrent academic research on LLM consensus failure."
    )

    pdf.body_text(
        "This does not mean multi-agent AI architectures are a dead end. It means the current "
        "paradigm of free-form verdict-level debate between role-assigned LLM agents is unreliable "
        "for tasks that require calibrated, evidence-grounded judgment. The path forward is not "
        "better prompts. It is better architecture -- structural scaffolding that leverages what "
        "LLMs do well (creative pattern recognition, anomaly detection, hypothesis generation) "
        "while constraining what they do poorly (consensus, calibration, evidence-grounded "
        "deliberation)."
    )

    pdf.body_text(
        "Three specific architectural directions emerge from this work and the Tribunal synthesis:"
    )

    pdf.indented_bullet(
        "Per-claim adversarial audit.",
        "Force debate onto specific evidential claims rather than overall verdicts. Each argument "
        "must cite specific evidence. This directly addresses the diagnosed mechanism by preventing "
        "vague, socially-driven objections."
    )
    pdf.indented_bullet(
        "Deterministic components.",
        "Replace the LLM Skeptic with a Python function that queries the evidence graph. Let AI "
        "do creative threat hunting; let code do rigorous fact-checking. This eliminates the "
        "Skeptic rigidity problem entirely by removing the LLM from the role where it fails."
    )
    pdf.indented_bullet(
        "Active Oracle.",
        "Convert the Oracle from passive judge to active stress-tester that generates the strongest "
        "argument against whichever agent holds higher confidence. This breaks the one-directional "
        "dynamic by forcing both agents to defend their positions."
    )

    pdf.body_text(
        "Our study is limited by testing primarily with one LLM provider (Anthropic Claude) and by "
        "the specialized cybersecurity domain. Whether the asymmetric calibration dynamics differ "
        "across model families or generalize to other grounded analytical domains remains an open "
        "question. The ARES infrastructure -- 33 scenarios with verified ground truth, frozen "
        "evidence chains, hash-chained audit trails, and comprehensive benchmarking -- is released "
        "as a testbed for studying these questions."
    )

    pdf.body_text(
        "We built ARES to protect networks by making AI minds argue. The argument taught us "
        "something neither side expected: the problem is inside the black box, and the solution "
        "is entirely outside of it."
    )

    # =========================================================================
    # REFERENCES
    # =========================================================================
    pdf.section_title('', 'REFERENCES')

    refs = [
        "F. Berdoz, L. Rugli, and R. Wattenhofer. Can AI Agents Agree? arXiv:2603.01213v2 [cs.MA], ETH Zurich, 2026.",
        "M. Cemri, M. Z. Pan, S. Yang, L. A. Agrawal, B. Chopra, R. Tiwari, et al. Why Do Multi-Agent LLM Systems Fail? In NeurIPS, 2025.",
        "B. Chen, G. Li, X. Lin, Z. Wang, and J. Li. BlockAgents: Towards Byzantine-Robust LLM-Based Multi-Agent Coordination via Blockchain. In ACM Turing Award Celebration Conference, 2024.",
        "H. Chen, W. Ji, L. Xu, and S. Zhao. Multi-Agent Consensus Seeking via Large Language Models, 2023. arXiv:2310.20151.",
        "M. A. Ferrag, O. Friha, D. Hamouda, L. Maglaras, and H. Janicke. Edge-IIoTset: A New Comprehensive Realistic Cyber Security Dataset of IoT and IIoT Applications. IEEE Access, 2025.",
        "F. Grotschla, L. Muller, J. Tonshoff, M. Galkin, and B. Perozzi. AgentsNet: Coordination and Collaborative Reasoning in Multi-Agent LLMs, 2025. arXiv:2507.08616.",
        "Y. Jo and C. Park. Byzantine-Robust Decentralized Coordination of LLM Agents, 2025. arXiv:2507.14928.",
        "G. Li, H. A. A. K. Hammoud, H. Itani, D. Khizbullin, and B. Ghanem. CAMEL: Communicative Agents for Mind Exploration of Large Language Model Society. In NeurIPS, 2023.",
        "X. Liu, H. Yu, H. Zhang, Y. Xu, X. Lei, H. Lai, et al. AgentBench: Evaluating LLMs as Agents. In ICLR, 2024.",
        "L. Wolf, S. Yoon, and I. Bogunovic. This Is Your Doge, If It Please You: Exploring Deception and Robustness in Mixture of LLMs, 2025. arXiv:2503.05856.",
        "Q. Wu, G. Bansal, J. Zhang, Y. Wu, B. Li, E. Zhu, et al. AutoGen: Enabling Next-Gen LLM Applications via Multi-Agent Conversations. In COLM, 2024.",
        "L. Zheng, J. Chen, Q. Yin, J. Zhang, X. Zeng, and Y. Tian. Rethinking the Reliability of Multi-agent System: A Perspective from Byzantine Fault Tolerance, 2025. arXiv:2511.10400.",
    ]

    for ref in refs:
        pdf.reference_entry(ref)

    # =========================================================================
    # APPENDIX A: DEVELOPMENT RECORD
    # =========================================================================
    pdf.add_page()
    pdf.section_title('A', 'APPENDIX')

    pdf.subsection_title('A.1', 'Reproducibility')

    pdf.body_text(
        "The complete ARES codebase, all 33 benchmark scenarios with ground-truth verdicts, "
        "confidence trace logs, cost tracking data, and session-by-session development records "
        "are preserved in the project repository. All experiments used Claude Sonnet via the "
        "Anthropic API with the strategy pattern enabling systematic substitution between "
        "rule-based and LLM-powered reasoning. Total API cost across all sessions: under $1.00."
    )

    pdf.subsection_title('A.2', 'Development Timeline')

    pdf.set_font('Helvetica', 'B', 9)
    pdf.cell(0, 6, 'Table 4: ARES development sessions.', ln=True)
    pdf.simple_table(
        headers=['Session', 'Component', 'Tests', 'Key Insight'],
        rows=[
            ['001', 'Graph Schema', '110', 'Node/edge types for security data'],
            ['002', 'Dialectical Foundation', '292', 'Hallucinations = schema violations'],
            ['003-004', 'Agent Foundation', '278', 'Rule-based Architect/Skeptic/Oracle'],
            ['005', 'Evidence Extractors', '130', 'Sensors observe, they do not analyze'],
            ['006-008', 'Coordination + Memory', '226', 'Orchestration + hash-chained audit trail'],
            ['009-010', 'LLM Integration', '178', 'Strategy pattern: extract then inject'],
            ['011-012', 'Benchmarking', '86', '50% to 91.7% via measurement'],
            ['013-014', 'Multi-Turn Experiment', '92', 'Debate amplifies commitment bias'],
            ['016-018', 'Evidence Expansion', '294', 'Diversity does not fix debate asymmetry'],
            ['019', 'Redis Backend', '42', 'Protocol pattern pays its dividend'],
            ['020', 'Protocol Fix + Verdict', '19', 'The Convergence'],
            ['021', 'Corpus Expansion', '73', 'N=18 to N=33, error classification'],
            ['022', 'Escalation Gate', '44', 'Zero good flips, selective escalation fails'],
            ['023-031', 'Accuracy Improvement', '137', 'Single-turn hardening to 72.7%'],
            ['032-037', 'Visual Pipeline', '200+', 'Visualization + benchmark expansion'],
        ],
        widths=[20, 40, 18, 82]
    )

    pdf.subsection_title('A.3', 'The Tribunal Process')

    pdf.body_text(
        "Three AI architectures received identical briefing documents: the Tribunal Battle Plan "
        "(a structured strategic brief with explicit response format) and the Compendium Volume I "
        "(the full narrative record of Sessions 001-020). Each was instructed to (1) assess the "
        "project's strongest proven finding, (2) identify blind spots, (3) recommend a direction "
        "from seven candidates, (4) propose a research angle, and (5) offer a wild card. No "
        "coordination between systems occurred. Responses were collected independently."
    )

    pdf.body_text(
        "The Tribunal produced convergence on all four strategic verdicts despite operating from "
        "different training data, different reasoning architectures, and different analytical "
        "tendencies. This process was faster than traditional peer review (hours, not months), "
        "parallel (three reviewers simultaneously), and structurally adversarial (the brief "
        "explicitly instructed reviewers to challenge, not agree). Whether this constitutes a "
        "new methodology for research review is a question for future work."
    )

    pdf.subsection_title('A.4', 'Agent Prompt Excerpts')

    pdf.body_text(
        "Below are excerpted prompt templates used for the Architect and Skeptic agents in the "
        "ARES debate protocol. Full prompts are available in the project repository."
    )

    # Architect prompt excerpt
    pdf.set_font('Helvetica', 'B', 8)
    pdf.cell(0, 5, 'Figure 1: Architect system prompt (excerpt).', ln=True)
    pdf.ln(1)
    pdf.set_font('Courier', '', 7)
    architect_prompt = (
        "You are the ARCHITECT agent in the ARES threat analysis system.\n"
        "Your role: Identify anomalous patterns and argue for threat classification.\n"
        "RULES:\n"
        "1. Every claim MUST reference a specific fact_id from the EvidencePacket.\n"
        "2. Claims without evidence citations will be REJECTED as schema violations.\n"
        "3. Your confidence score (0.0-1.0) must reflect the strength of cited evidence.\n"
        "4. A confidence of 0.5 represents genuine uncertainty, not weakness.\n"
        "OUTPUT: Return structured JSON with threat_level, confidence, and\n"
        "        evidence_citations array referencing specific fact_ids."
    )
    pdf.set_fill_color(245, 245, 245)
    y_start = pdf.get_y()
    pdf.multi_cell(pdf.body_width, 3.2, architect_prompt, fill=True, border=1)
    pdf.ln(3)

    # Skeptic prompt excerpt
    pdf.set_font('Helvetica', 'B', 8)
    pdf.cell(0, 5, 'Figure 2: Skeptic system prompt (excerpt).', ln=True)
    pdf.ln(1)
    pdf.set_font('Courier', '', 7)
    skeptic_prompt = (
        "You are the SKEPTIC agent in the ARES threat analysis system.\n"
        "Your role: Challenge threat assessments and argue for benign explanations.\n"
        "RULES:\n"
        "1. Every counter-argument MUST reference specific fact_ids that support\n"
        "   a benign interpretation.\n"
        "2. You MUST engage with the Architect's specific evidence citations.\n"
        "3. If the Architect cites compelling evidence, adjust your confidence.\n"
        "4. Rigidity without evidence-based justification is as harmful as\n"
        "   capitulation without cause.\n"
        "OUTPUT: Return structured JSON with assessment, confidence, and\n"
        "        counter_evidence array referencing specific fact_ids."
    )
    pdf.multi_cell(pdf.body_width, 3.2, skeptic_prompt, fill=True, border=1)
    pdf.ln(3)

    # Conviction anchoring prompt excerpt
    pdf.set_font('Helvetica', 'B', 8)
    pdf.cell(0, 5, 'Figure 3: Conviction-anchored protocol addition (Session 020).', ln=True)
    pdf.ln(1)
    pdf.set_font('Courier', '', 7)
    anchor_prompt = (
        "CONVICTION ANCHORING PROTOCOL:\n"
        "- Architect: You MUST maintain your confidence level unless the Skeptic\n"
        "  cites SPECIFIC counter-evidence from the EvidencePacket that directly\n"
        "  contradicts your cited facts. Social pressure alone is not grounds\n"
        "  for lowering confidence.\n"
        "- Skeptic: You MUST acknowledge when the Architect successfully rebuts\n"
        "  your counter-argument with specific evidence. Holding position without\n"
        "  new evidence is a protocol violation.\n"
        "- Both: Track per-claim confidence with explicit delta justification.\n"
        "  Every confidence change must cite the specific evidence that caused it."
    )
    pdf.multi_cell(pdf.body_width, 3.2, anchor_prompt, fill=True, border=1)
    pdf.ln(5)

    pdf.subsection_title('A.5', 'OracleJudge Decision Logic')

    pdf.set_font('Helvetica', 'B', 8)
    pdf.cell(0, 5, 'Figure 4: OracleJudge V1 deterministic verdict computation.', ln=True)
    pdf.ln(1)
    pdf.set_font('Courier', '', 7)
    oracle_v1 = (
        "def compute_verdict(arch_confidence, skep_confidence):\n"
        "    \"\"\"Deterministic verdict -- no LLM touches this computation.\"\"\"\n"
        "    if arch_confidence >= 0.7 and skep_confidence < 0.5:\n"
        "        return THREAT_CONFIRMED\n"
        "    if skep_confidence >= 0.7 and arch_confidence < 0.5:\n"
        "        return THREAT_DISMISSED\n"
        "    return INCONCLUSIVE"
    )
    pdf.multi_cell(pdf.body_width, 3.2, oracle_v1, fill=True, border=1)
    pdf.ln(3)

    pdf.set_font('Helvetica', 'B', 8)
    pdf.cell(0, 5, 'Figure 5: OracleJudge V2 delta-based scoring.', ln=True)
    pdf.ln(1)
    pdf.set_font('Courier', '', 7)
    oracle_v2 = (
        "def compute_verdict_v2(arch_confidence, skep_confidence):\n"
        "    \"\"\"Delta-based: dominant override + confidence gap + min threshold.\"\"\"\n"
        "    delta = abs(arch_confidence - skep_confidence)\n"
        "    dominant = max(arch_confidence, skep_confidence)\n"
        "    # Dominant override: very high confidence wins outright\n"
        "    if arch_confidence > 0.85:\n"
        "        return THREAT_CONFIRMED\n"
        "    if skep_confidence > 0.85:\n"
        "        return THREAT_DISMISSED\n"
        "    # Delta threshold: clear winner with sufficient gap\n"
        "    if delta > 0.15 and dominant > 0.60:\n"
        "        if arch_confidence > skep_confidence:\n"
        "            return THREAT_CONFIRMED\n"
        "        return THREAT_DISMISSED\n"
        "    return INCONCLUSIVE"
    )
    pdf.multi_cell(pdf.body_width, 3.2, oracle_v2, fill=True, border=1)
    pdf.ln(5)

    # =========================================================================
    # FINAL NOTE
    # =========================================================================
    pdf.ln(5)
    pdf.set_font('Times', 'I', 9)
    pdf.set_text_color(80, 80, 80)
    final_note = (
        "This compendium records a moment that was not planned and could not have been predicted. "
        "A solo builder working on a cybersecurity project, iterating through thirty-seven "
        "development sessions with an AI pair-programming partner, arrived independently at "
        "findings that align with formal academic research from one of Europe's premier technical "
        "universities. This did not happen because the builder was lucky. It happened because "
        "the process was disciplined. Every session had a brief, a test suite, and a regression "
        "check. Every finding was measured, not assumed. Every failure was documented honestly. "
        "When the multi-turn experiment produced a negative result, it was published as a finding, "
        "not buried as a setback."
    )
    pdf.multi_cell(pdf.body_width, 4, final_note, align='J')
    pdf.ln(3)

    pdf.set_font('Times', 'I', 9)
    closing = (
        "The taste of genuine discovery is unmistakable. It arrives not when you find what you "
        "expected, but when the data reveals something you could not have predicted -- and you "
        "realize others, working independently, found the same thing."
    )
    pdf.multi_cell(pdf.body_width, 4, closing, align='J')
    pdf.ln(5)

    pdf.set_font('Times', '', 8)
    pdf.set_text_color(0, 0, 0)
    pdf.cell(0, 4, 'Preprint compiled March 28, 2026. Skyframe Innovations.', ln=True)
    pdf.cell(0, 4, '37 sessions. 2,001 tests. Zero regressions. The work continues.', ln=True)

    # Save
    output_path = os.path.join(os.path.dirname(os.path.abspath(__file__)),
                                'ARES_Preprint_Asymmetric_Calibration_Failure.pdf')
    pdf.output(output_path)
    print(f"Paper generated: {output_path}")
    print(f"Pages: {pdf.page_no()}")
    return output_path


if __name__ == '__main__':
    build_paper()