Calibrated estimates

Calibrated estimates replace a precise-looking guess with a defensible interval and an explicit decision implication.

Behavior

1. The user supplies an uncertain quantity and the decision it informs.
2. The agent identifies the unit, time horizon, threshold, and cost of being wrong.
3. The agent classifies the target mode: paid quote, market value, budget allowance, amount likely paid, official benchmark, or ambiguous.
4. The quantity is decomposed into smaller uncertain components.
5. Each component receives a low, central, and high value with a stated basis.
6. The agent calibrates the rolled-up interval and states its confidence.
7. The final interval is compared with the decision threshold.
8. The memo states the action implication and the largest remaining uncertainty.

The agent does not return only a point estimate.

Inputs & outputs

Inputs

• decision statement;
• quantity of interest;
• unit;
• time horizon;
• decision threshold;
• cost of being wrong;
• evidence, source links, files, observations, and constraints; and
• optional requested confidence level.

Outputs

• calibrated range;
• central estimate;
• confidence statement;
• decomposition table;
• threshold implication;
• top uncertainty driver;
• recommendation; and
• structured HTMA_RESULT appendix.

States & edge cases

State	Behavior
Complete evidence	Return a calibrated memo with estimate_status: "estimated".
Vague decision	Ask for the missing decision or threshold before external research.
Missing identifier	Keep numeric fields null and return needs_identifier.
Missing effective period	Keep numeric fields null and return needs_effective_period.
Current authoritative lookup required	Return lookup_required until the source is refreshed.
Private fact missing	Name it in blocking_missing_inputs; do not infer it.
Inputs too weak for a range	Return not_estimable and the next measurement step.
Range crosses the threshold	Explain the decision sensitivity instead of forcing a yes/no recommendation.
Monte Carlo requested too early	Decompose and calibrate inputs before simulation.

Data shape

type EstimateContext = {
  decision: string
  quantity: string
  unit: string
  timeHorizon: string
  decisionThreshold: number | string | null
  costOfBeingWrong: string
  estimateMode:
    | "paid quote"
    | "market value"
    | "budget allowance"
    | "amount likely paid"
    | "official/public benchmark"
    | "ambiguous"
  evidence: EvidenceItem[]
}

type ComponentRange = {
  component: string
  low: number
  central: number
  high: number
  confidence: string
  basis: string
  whatWouldTighten: string
}

Decisions

• 2026-07-11 — Intervals are mandatory. A central value without bounds hides the uncertainty the skill exists to expose.
• 2026-07-11 — Missing required inputs produce nulls. Fabricated completeness is worse than a blocked result.

Open questions

None for the current release.