Where should Kestrel put $14M?

A distribution is a better forecast input than an assumption

A driver-based model is arithmetic until it reaches an assumption. Normally those assumptions are set in a planning meeting by whoever argues hardest, recorded as a single number, and then wrapped in a made-up confidence interval at the end. A System One model answers the same questions against the actual evidence and returns a probability distribution for each one. Those distributions are not a summary of the uncertainty — they are the uncertainty, in the only form a Monte Carlo can integrate over. The width of the band below is inherited from how sure the model was, judgment by judgment.

{
  "state": "<the deal record above>",
  "model": "jev-latest",
  "questions": {
    "closes_in_horizon": {
      "type": "noul",
      "instructions": "Will this opportunity be signed within the next two quarters?"
    },
    "primary_blocker": {
      "type": "choice",
      "instructions": "What single thing most stands between this deal and signature?",
      "criteria": {
        "budget_freeze": "Money is unavailable or suspended",
        "security_review": "An open security, compliance or pen-test gate",
        "...": "..."
      }
    },
    "...": "+ 3 more, same call"
  }
}

{
  "model": "jev-latest",
  "answers": {
    "closes_in_horizon": {
      "type": "noul",
      "noul": 0.8871
    },
    "primary_blocker": {
      "type": "choice",
      "choice": "integration_gap",
      "probabilities": {
        "budget_freeze": 0.0952,
        "security_review": 0.2083,
        "champion_departed": 0.0412,
        "competitor": 0.0412,
        "integration_gap": 0.5849,
        "none": 0.0292
      },
      "confidence": 0.3133
    }
  },
  "usage": {
    "input_tokens": 730,
    "output_tokens": 0
  }
}

Six allocation policies, same evidence

A policy is nothing but a set of weights over the six questions asked about each initiative, plus a view on what an unmeasurable strategic case is worth. The weights live in forecast/policies.py where a CFO can argue with them. Each row is 20,000 Monte Carlo trials drawing from the judgment distributions.

Where a human should actually spend an hour

Calibrated confidence turns into triage. Each judgment is gated at a threshold that scales with the money it moves — a $540k deal and a $4.2M renewal do not deserve the same bar — and ranked by how much forecast variance it carries. Independent Bernoullis, so each one's share is amount² · p(1-p) over the total. 6 of 22 judgments fall below their floor, and they carry 42% of the spread in the forecast. That is the list, in order.

Does the confidence mean anything?

Everything above consumes probabilities and trusts them, so the number worth publishing is this one. Because the corpus is synthetic, every judgment has a known answer, and the same harness can score the emulator, a Claude-backed stand-in, or the real model on identical questions.

Four analytical junctures

Every judgment the forecast rests on, with the evidence it was made from. Nothing prefixed with an underscore in the corpus — the hidden labels — is ever sent as state; there is a test that walks every record to prove it.

What is emulated, and what that costs