{
  "model_name": "PrincipledMachines Cortex 0.1",
  "model_organization": "Principled Machines",
  "submitting_organization": "Principled Machines",
  "submission_date": "2026-10-20",
  "submission_type": "custom",
  "modality": "text",
  "contact_info": {
    "name": "Badri Iyengar",
    "email": "badri@principledmachines.com",
    "github": "badrinarayan"
  },
  "results": {
    "banking_knowledge": {
      "pass_1": 73.19587628865979,
      "pass_2": null,
      "pass_3": null,
      "pass_4": null,
      "cost": 0.5955274784149485,
      "retrieval_config": "cortex"
    },
    "retail": null,
    "airline": null,
    "telecom": null
  },
  "trajectories_available": true,
  "trajectory_files": {
    "banking_knowledge": "cortex-banking-verified-2026-11-03.json"
  },
  "references": [
    {
      "title": "Cortex on τ³-bench",
      "url": "https://principledmachines.com/research/tau3-banking/",
      "type": "blog_post"
    }
  ],
  "methodology": {
    "evaluation_date": "2026-11-03",
    "tau2_bench_version": "1.0.1",
    "user_simulator": "gpt-5.2",
    "notes": "Cortex 0.1 is a cognitive architecture from Principled Machines. A cognitive core directs the work; a language model is one kind of compute it uses, for semantic judgement, and other parts of a task go to retrieval, programs and other specialized components. We attribute the result to two mechanisms: enhanced retrieval for multi-hop questions over the knowledge base, and process-calculus abstractions for the policy workflows. Submitted as custom because Cortex replaces the standard single-model scaffold. No task-specific fine-tuning, no prompt tuning against the grader, no post-training on labels, and no autoresearch loop. Cost is the average agent spend per trajectory, excluding the user simulator. Write-up: https://principledmachines.com/research/tau3-banking/",
    "verification": {
      "modified_prompts": false,
      "omitted_questions": false,
      "details": "Verified by Sierra on 2026-11-03: all 97 banking_knowledge tasks, 1 trial, run independently with the submitter's cortex_tau2.py client in --local mode at commit b7ea9074. Only agent calls went to the Cortex endpoint (cortex-0.1, reasoning_effort medium); the user simulator (gpt-5.2, reasoning_effort low) and all grading ran on Sierra's side with Sierra's keys. 71/97 passed (pass^1 73.20%); no infrastructure errors. Published scores and trajectories are from this run. The submitter's own 4-trial run reported 78.35% pass^1; pass^2-4 are omitted because the verified run has 1 trial. Caveats: knowledge-base retrieval runs on the Cortex server and is not visible in the trajectories (the submitter states the index contains only the banking_knowledge documents), and the hosted endpoint is not a frozen snapshot, so the result reflects cortex-0.1 as served on 2026-11-03. Cost is the submitter's reported average agent spend."
    }
  },
  "reasoning_effort": "medium",
  "is_new": true,
  "model_release": {
    "release_date": "2026-11-04",
    "announcement_url": null,
    "announcement_title": null
  },
  "voice_config": null,
  "interaction_metrics": null
}