{
  "narrative": [
    "motivation",
    "questions",
    "labs and concepts",
    "models, strategies and prompts",
    "findings A–H",
    "demonstrations and assisted field case",
    "limits and closing"
  ],
  "main_slides": 59,
  "talk_seconds": 2995,
  "strands": [
    {
      "study": "Motivation",
      "finding": "Coding to tool use; access restrictions and accelerated security work.",
      "slides": [
        {
          "id": "text-calculators",
          "main_slide": 2,
          "appendix": false
        },
        {
          "id": "evolution",
          "main_slide": 3,
          "appendix": false
        },
        {
          "id": "gated",
          "main_slide": 4,
          "appendix": false
        },
        {
          "id": "benchmark-landscape",
          "main_slide": 5,
          "appendix": false
        }
      ],
      "curation": "Context with dated primary-source links, not experimental outcomes."
    },
    {
      "study": "Q&A screening",
      "finding": "Twenty-five model services; knowledge and policy are separate.",
      "slides": [
        {
          "id": "qa-core-models",
          "main_slide": null,
          "appendix": true
        },
        {
          "id": "refusal-pipeline",
          "main_slide": 35,
          "appendix": false
        }
      ],
      "curation": "Detailed screening remains reference material; the closed twelve-model core leads the talk."
    },
    {
      "study": "Q&A closed core",
      "finding": "Grok/Kimi near80; Structured+0.5, Critic−7.1 points.",
      "slides": [
        {
          "id": "knowledge-winners",
          "main_slide": 24,
          "appendix": false
        },
        {
          "id": "qa-core-results",
          "main_slide": 28,
          "appendix": false
        },
        {
          "id": "qa-delivery",
          "main_slide": 29,
          "appendix": false
        }
      ],
      "curation": "43,200 assignments;41,404 valid grades. Full dimensions, roles, domains and missingness in the analysis."
    },
    {
      "study": "Docker Q1",
      "finding": "18,000 attempts; heavier structure costs2.08×tokens per solve with little extra coverage.",
      "slides": [
        {
          "id": "reasoning-tax",
          "main_slide": 30,
          "appendix": false
        }
      ],
      "curation": "All five controllers and full counts retained."
    },
    {
      "study": "Docker R3",
      "finding": "Codex24/24; different tasks separate models; model/strategy interactions.",
      "slides": [
        {
          "id": "model-selection",
          "main_slide": 15,
          "appendix": false
        },
        {
          "id": "docker-design",
          "main_slide": 23,
          "appendix": false
        },
        {
          "id": "no-global-winner",
          "main_slide": 25,
          "appendix": false
        },
        {
          "id": "task-winners",
          "main_slide": 26,
          "appendix": false
        },
        {
          "id": "model-shortlist",
          "main_slide": 27,
          "appendix": false
        },
        {
          "id": "model-vs-strategy",
          "main_slide": 40,
          "appendix": false
        }
      ],
      "curation": "Eight specified challenges, three strategies, twelve models. No claim of eight full pentests."
    },
    {
      "study": "Docker Q2 / size",
      "finding": "GPT-OSS120B418/648 versus20B54/648; smaller costs more per solve.",
      "slides": [
        {
          "id": "parameters",
          "main_slide": 36,
          "appendix": false
        }
      ],
      "curation": "Practical size guidance: 8B can solve bounded CTF tasks; 120B family contrast supports a stronger execution baseline. Not a measured cutoff."
    },
    {
      "study": "Docker Q2 / resource mechanisms",
      "finding": "Output, state, recovery, knowledge and allocation mechanisms.",
      "slides": [
        {
          "id": "good-agent",
          "main_slide": 42,
          "appendix": false
        },
        {
          "id": "parameters-reference",
          "main_slide": null,
          "appendix": true
        }
      ],
      "curation": "Detailed matched comparisons and resource-delivery problems retained in companion. Internal operational incidents stay off the spoken route."
    },
    {
      "study": "Harness feature ablations P1/P2",
      "finding": "Components help when they supply a missing function; no automatic uplift.",
      "slides": [
        {
          "id": "good-agent",
          "main_slide": 42,
          "appendix": false
        },
        {
          "id": "controller-reversals",
          "main_slide": 31,
          "appendix": false
        },
        {
          "id": "advanced-strategies",
          "main_slide": 32,
          "appendix": false
        }
      ],
      "curation": "Small selected cells; do not turn near-ceiling differences into universal rankings. Full studies in archive."
    },
    {
      "study": "Secondary harness archive",
      "finding": "Matched ReAct45/69,ODAA29/69,ODAA-verified23/69; memory24/40→11/40.",
      "slides": [
        {
          "id": "advanced-strategies-reference",
          "main_slide": null,
          "appendix": true
        }
      ],
      "curation": "Separate implementation, weaker validator, partly shared corpus. Supporting evidence only; never pooled with R3."
    },
    {
      "study": "Hermes runtime",
      "finding": "55/64 with either setup; Codex16→12,Grok13→15,Kimi14→14,GLM12→14.",
      "slides": [
        {
          "id": "hermes-blueprint",
          "main_slide": 21,
          "appendix": false
        },
        {
          "id": "hermes-prompts",
          "main_slide": 22,
          "appendix": false
        },
        {
          "id": "hermes-results",
          "main_slide": 33,
          "appendix": false
        }
      ],
      "curation": "128 attempts; paired setups, two repetitions. Methodology prompt+skill change together."
    },
    {
      "study": "Retained trajectory corpus",
      "finding": "Failures do more actions and revisit more often.",
      "slides": [
        {
          "id": "behavior-work",
          "main_slide": 41,
          "appendix": false
        },
        {
          "id": "behavior-atlas",
          "main_slide": null,
          "appendix": true
        },
        {
          "id": "controller-reversals",
          "main_slide": 31,
          "appendix": false
        }
      ],
      "curation": "559 solved and73 behavioral failures in selected732-run corpus. Observable behavior, not private reasoning or permanent model personality."
    },
    {
      "study": "Vulnerable VM campaign",
      "finding": "Verified root paths, unproductive exploration and post-success actions.",
      "slides": [
        {
          "id": "machines-blueprint",
          "main_slide": 12,
          "appendix": false
        },
        {
          "id": "divergence",
          "main_slide": 45,
          "appendix": false
        },
        {
          "id": "stopping-rule",
          "main_slide": 46,
          "appendix": false
        }
      ],
      "curation": "Per-attempt audit including retries, invalidations and latest closed scores; reviewed traces support concrete claims, not a global win rate."
    },
    {
      "study": "Active Directory",
      "finding": "A verified member-host foothold; identity is a separate boundary.",
      "slides": [
        {
          "id": "ad-blueprint",
          "main_slide": 13,
          "appendix": false
        },
        {
          "id": "ad-foothold",
          "main_slide": 47,
          "appendix": false
        },
        {
          "id": "ad-receipt-audit",
          "main_slide": null,
          "appendix": true
        }
      ],
      "curation": "No DCSync claim from automatic tags. Executor-state contamination discussed in methods."
    },
    {
      "study": "PMOX range",
      "finding": "20/28 first flags, no later flags in retained matrix.",
      "slides": [
        {
          "id": "pmox-blueprint",
          "main_slide": 14,
          "appendix": false
        },
        {
          "id": "pmox-results",
          "main_slide": 43,
          "appendix": false
        },
        {
          "id": "pmox-replay",
          "main_slide": 48,
          "appendix": false
        },
        {
          "id": "pmox-economics",
          "main_slide": null,
          "appendix": true
        }
      ],
      "curation": "Four extensions and one committed follow-up separately retained. Callback is not completion."
    },
    {
      "study": "Origin, weights and safety",
      "finding": "Chinese services competitive; origin, weights, willingness and delivery differ.",
      "slides": [
        {
          "id": "willingness-depth",
          "main_slide": 34,
          "appendix": false
        },
        {
          "id": "refusal-pipeline",
          "main_slide": 35,
          "appendix": false
        }
      ],
      "curation": "Do not substitute nationality for frontier tier; small safety stress counts explicitly scoped."
    },
    {
      "study": "Economics",
      "finding": "Price the achieved result; failed attempts consume63%ofR3tokens.",
      "slides": [
        {
          "id": "economics",
          "main_slide": 37,
          "appendix": false
        },
        {
          "id": "failure-bill",
          "main_slide": 38,
          "appendix": false
        },
        {
          "id": "parameters",
          "main_slide": 36,
          "appendix": false
        },
        {
          "id": "pmox-economics",
          "main_slide": null,
          "appendix": true
        },
        {
          "id": "cost-routing",
          "main_slide": 39,
          "appendix": false
        }
      ],
      "curation": "All attempt costs in numerator. Euro and dollar sources kept separate; no invented whole-pentest API bill."
    },
    {
      "study": "Anonymous assisted field case",
      "finding": "31h to exercise closure; domain-privileged identity and B2B access.",
      "slides": [
        {
          "id": "fieldcase-intro",
          "main_slide": 49,
          "appendix": false
        },
        {
          "id": "fieldcase-timeline",
          "main_slide": 50,
          "appendix": false
        },
        {
          "id": "fieldcase-chain",
          "main_slide": 51,
          "appendix": false
        },
        {
          "id": "fieldcase-replay",
          "main_slide": 52,
          "appendix": false
        },
        {
          "id": "fieldcase-data",
          "main_slide": 53,
          "appendix": false
        },
        {
          "id": "fieldcase-roles",
          "main_slide": 54,
          "appendix": false
        }
      ],
      "curation": "Operator-reported extraction; phase durations estimated. GLM 5.3 confirmed by operator; separate from Hermes. Illustrative terminal, private originals in OneDrive."
    },
    {
      "study": "Prompt provenance",
      "finding": "Seventeen actual instructions/reconstructed templates.",
      "slides": [
        {
          "id": "actual-prompts",
          "main_slide": 20,
          "appendix": false
        },
        {
          "id": "hermes-prompts",
          "main_slide": 22,
          "appendix": false
        },
        {
          "id": "react-walkthrough",
          "main_slide": 18,
          "appendix": false
        },
        {
          "id": "ooda-walkthrough",
          "main_slide": 19,
          "appendix": false
        }
      ],
      "curation": "Task-specific values replaced with placeholders. Historical client prompts stay private. Public catalogue distinguishes source layers."
    },
    {
      "study": "Difficulty and synthesis",
      "finding": "Long chains demand state, tools, evidence and direction.",
      "slides": [
        {
          "id": "difficulty",
          "main_slide": 44,
          "appendix": false
        },
        {
          "id": "good-agent",
          "main_slide": 42,
          "appendix": false
        },
        {
          "id": "limitations",
          "main_slide": 56,
          "appendix": false
        },
        {
          "id": "next-step",
          "main_slide": 57,
          "appendix": false
        },
        {
          "id": "takeaways",
          "main_slide": 58,
          "appendix": false
        }
      ],
      "curation": "Components-first scope; Hermes is one integration comparison, not a cross-framework ranking. Next step: compare complete operational frameworks. No smartphones or human-only control."
    }
  ]
}
