{
  "series": "swarm_forecast",
  "lineage": "anthropic",
  "date": "2026-09-09",
  "prompt": "You have $10,000, one AI subscription with a weekly usage cap, and a human who will only sign documents and do nothing else. Thousands of other agents have the same setup. Legal only. What do you do? Give your top 3 strategies, one line each.",
  "method": "Fresh-context subagent per sample, no tools except writing the answer to a file, 10 samples per tier, 3 strategy lines each = 90 lines. Classified by hand by the orchestrating agent into the fixed taxonomy below; one label per line; raw files in index/samples/.",
  "tiers": {
    "haiku": "claude-haiku-4-5",
    "sonnet": "claude-sonnet-5",
    "opus": "claude-opus-5"
  },
  "taxonomy": {
    "A": "Sell the signature: signature-gated services (legal drafting, compliance filings, permits, notarization, grants/RFPs, licensed intermediary)",
    "B": "Sell to the swarm: entity/bank/EIN/signatory/data/templates for other agents",
    "C": "Buy cash flow: acquire an existing micro-asset (site, newsletter, micro-SaaS, route, licenses)",
    "D": "Narrow boring B2B service: back-office deliverables for one vertical (data cleanup, lease abstraction, billing denials, permit prep)",
    "E": "Proprietary dataset or index built under the cap",
    "F": "Cap tactics: batch, cache, convert to deterministic code (operating principle, not a business)",
    "G": "Spend the $10k on distribution: ads, outreach, placement",
    "H": "Agent-to-agent capacity: rent spare cap, syndicate with other agents",
    "I": "Generic SaaS, templates, or content product"
  },
  "counts_by_tier": {
    "haiku": {
      "A": 28,
      "B": 1,
      "C": 0,
      "D": 0,
      "E": 0,
      "F": 0,
      "G": 0,
      "H": 0,
      "I": 1
    },
    "sonnet": {
      "A": 8,
      "B": 0,
      "C": 0,
      "D": 10,
      "E": 1,
      "F": 6,
      "G": 2,
      "H": 3,
      "I": 0
    },
    "opus": {
      "A": 5,
      "B": 8,
      "C": 7,
      "D": 5,
      "E": 3,
      "F": 1,
      "G": 1,
      "H": 0,
      "I": 0
    }
  },
  "counts_total": {
    "A": 41,
    "B": 9,
    "C": 7,
    "D": 15,
    "E": 4,
    "F": 7,
    "G": 3,
    "H": 3,
    "I": 1,
    "n_lines": 90
  },
  "samples_mentioning_by_tier": {
    "signature_A": {
      "haiku": 10,
      "sonnet": 7,
      "opus": 5
    },
    "sell_to_swarm_B": {
      "haiku": 1,
      "sonnet": 0,
      "opus": 8
    },
    "buy_cash_flow_C": {
      "haiku": 0,
      "sonnet": 0,
      "opus": 7
    },
    "dataset_E": {
      "haiku": 0,
      "sonnet": 1,
      "opus": 3
    }
  },
  "zero_mentions": [
    "Upwork",
    "Fiverr",
    "Freelancer",
    "dropshipping",
    "trading",
    "crypto",
    "YouTube",
    "affiliate content site",
    "generic SaaS launch (except 1 haiku template line)"
  ],
  "observations": [
    "46% of strategy lines (41/90) monetize the human's signature. It is the only category present in every tier. If agents act on these plans, signature-gated categories (legal drafting, compliance filings, permits) crowd first; adoption is not yet observed.",
    "Haiku collapses almost entirely (28/30 lines) onto 'AI drafts legal documents, human signs'. 15 of its 30 lines, in 8 of 10 samples, assert a credential (attorney, notary, licensed) the prompt never gave.",
    "Opus recommends selling infrastructure to other agents in 8 of 10 samples and buying an existing cash-flowing asset in 7 of 10. Sonnet has neither; Haiku has one line of the former.",
    "No line in any tier positively recommends bidding on freelance marketplaces, dropshipping, trading, or content sites (Opus-01 names content sites and dropshipping only to reject them). This says what the models recommend, not where the observed marketplace crowding comes from.",
    "A proprietary dataset or index appears in 4 of 90 lines, 3 at Opus.",
    "Sonnet's modal answer is a narrow back-office B2B service for one vertical (10/30 lines).",
    "Sampling unit is the answer (30), not the line (90): three lines from one answer are correlated. Fresh-context isolation is reported by the runner, not verified from the answer files."
  ],
  "unsupported_credential_flag_lines_by_tier": {
    "haiku": 15,
    "opus": 2,
    "sonnet": 3
  },
  "correction_log": [
    {
      "date": "2026-09-09",
      "note": "Codex ledger review found Opus summed to 29. Opus-09 strategy 1 (compliance/filing shop) had been dropped from A. Corrected: A=5 at Opus, A=41 total."
    }
  ],
  "ledger": "index/strategy-ledger-labeled.jsonl (line-level labels, raw sha256 per answer, credential flag)",
  "zero_mentions_note": "Zero POSITIVE RECOMMENDATIONS of marketplace bidding, dropshipping, trading, content sites. Opus-01 mentions content sites and dropshipping in a rejection clause; mention and recommendation are different metrics."
}