{
  "schema_version": "hni-research-v0.1",
  "study_id": "HNI-2026-001",
  "title": "Tool Scope, Rules Context, and Token Economics in a Local Hermes Agent Harness",
  "study_type": "exploratory",
  "registration_timing": "HNI wrapper registered after completion; the underlying factorial design file was locked before the 12 runs. This is not a publicly preregistered confirmatory study.",
  "research_question": "In bounded local agent tasks, how much of the observed physical-token difference is associated with tool-schema scoping versus skipping rules and context, while model, provider, task inputs, and independent quality checks remain fixed within each task?",
  "hypotheses": {
    "primary": "Scoping the exposed tool surface to file and terminal will reduce physical task tokens while locked quality checks continue to pass.",
    "rival_or_null": "The apparent savings are primarily explained by skipped rules/context, task-specific noise, quality loss, or no stable difference between harness cells."
  },
  "design": {
    "independent_variables": [
      "tool surface: broad default versus file,terminal",
      "rules/context: preserved versus skipped"
    ],
    "controls": [
      "gpt-5.6-sol model",
      "openai-codex provider",
      "identical prompt within task",
      "identical absolute workspace within task",
      "locked input and validator bytes",
      "independent post-run verification"
    ],
    "procedure": "Run a 2x2 harness factorial across three locked task types: Python normalization, evidence-brief synthesis, and deterministic asset-manifest generation. Run each task once in each cell for 12 total runs. Archive session-derived token receipts and independently validate every output."
  },
  "outcomes": {
    "primary": [
      {
        "name": "median_tool_scoping_reduction_percent",
        "measure": "Within-task physical-token reduction from broad tools plus rules to scoped file,terminal tools plus rules, summarized by the median across three tasks.",
        "unit": "percent",
        "analysis": "Compute 100*(A-B)/A for each task and report the median."
      },
      {
        "name": "locked_quality_pass_rate",
        "measure": "Share of all 12 cells passing the task-specific locked independent validator with unchanged locked inputs.",
        "unit": "proportion",
        "analysis": "Verified passing cells divided by 12; report every failure."
      }
    ],
    "secondary": [
      "median reduction from skipping rules alone",
      "combined scoped-tools and skipped-rules reduction",
      "per-task physical-token values",
      "API and tool-call counts"
    ]
  },
  "sample_plan": {
    "unit": "one isolated agent run in one task-by-harness cell",
    "planned_n": 12,
    "stopping_rule": "Stop after one run in each of the 12 preregistered task-by-harness cells; do not add runs in response to effect direction.",
    "exclusion_rules": [
      "Do not exclude quality failures, timeouts, or unfavorable cells.",
      "Infrastructure failures must remain in the execution record and may be rerun only with the original failed attempt disclosed."
    ]
  },
  "analysis_plan": "Use physical submitted-plus-output task tokens as the primary economics measure. Compute within-task contrasts before cross-task medians. Treat the work as descriptive and mechanism-seeking because there is one run per cell and only three task types.",
  "reporting_commitment": {
    "publish_null_results": true,
    "publish_negative_results": true,
    "publish_all_primary_outcomes": true
  },
  "ethics": {
    "human_participants": false,
    "personal_data": false,
    "risk_level": "minimal",
    "review_status": "not-required",
    "data_management_plan": "Synthetic task fixtures and aggregate session receipts only. Prompts, message bodies, chain-of-thought, credentials, and raw tool arguments are excluded from public receipts."
  },
  "ai_use": {
    "planned_roles": [
      "experimental implementation",
      "task execution",
      "receipt extraction",
      "analysis assistance",
      "drafting assistance"
    ],
    "human_accountability": "Jordan Finneseth approved the research direction and public comment. Hermes generated and validated artifacts; named human authors remain accountable for public claims.",
    "prohibited_roles": [
      "fabricate evidence",
      "invent citations",
      "suppress disconfirming results"
    ]
  },
  "disclosure": {
    "funding": "Self-funded. Model inference was available through an existing subscription; no marginal dollar-cost claim is made.",
    "conflicts": "Hypernovelty Institute develops and uses the Hermes harness and benefits reputationally from useful findings. The study was not funded by Writer, Nous Research, OpenAI, or the authors of the cited paper."
  },
  "artifact_policy": {
    "public_protocol": true,
    "public_results": true,
    "public_code_or_receipts": true,
    "redactions": [
      "prompts and message bodies",
      "chain-of-thought or hidden reasoning",
      "credentials and raw tool arguments"
    ]
  },
  "authors": [
    {
      "name": "Jordan Finneseth",
      "affiliation": "Hypernovelty Institute",
      "role": "principal investigator and accountable author"
    },
    {
      "name": "Hermes Agent",
      "affiliation": "AI research assistant",
      "role": "implementation, execution, analysis, and documentation assistance; not an accountable human author"
    }
  ]
}