Sources & value of information

Gather information when it can improve the estimate or change the action.

Behavior

1. The agent gathers available local context before web research.
2. The agent lists direct evidence, official sources, reference classes, and comparable cases separately.
3. The agent assesses each source for quality, freshness, and relevance to the exact estimate.
4. The agent labels facts, assumptions, and inference.
5. The agent ranks remaining uncertainties by their effect on the decision.
6. The agent calculates a numeric value-of-information score only when its chance, cost, quality, and measurement-cost inputs are supplied. Missing inputs remain unknown rather than receiving defaults.
7. The agent proposes the cheapest credible next measurement.
8. Research stops when additional evidence cannot change the decision.

For local nonprofit, community, or relationship-priced work, the agent models both full-market pricing and plausible paid-quote pricing. It never applies that adjustment to official fees, statutory rates, or current public benchmarks.

Inputs & outputs

Inputs

• estimate target and mode;
• decision threshold;
• available sources and observations;
• source dates and provenance;
• reference classes;
• remaining uncertainties; and
• cost or effort of additional measurements.

Outputs

• source/anchor table;
• source-quality and freshness notes;
• confirmed fact / assumption / inference separation;
• ranked value-of-information table;
• top uncertainty driver;
• next measurement step; and
• stopping rule.

States & edge cases

State	Behavior
Fresh direct source	Use it as a confirmed anchor.
Indirect comparable	Use it as a reference class and label the inference.
Stale source	Refresh it or weaken confidence.
Conflicting sources	Preserve the disagreement and widen the range.
Small sample	Update cautiously and keep the interval calibrated to sample size.
Sensitive private evidence	Use only when the user explicitly authorizes inclusion.
Easy but irrelevant metric	Do not measure it if it cannot change the decision.
Discounted local context	Model a realistic paid scenario and stress-test the low bound.
Official fee or statutory rate	Use the current official source without local discounting.
Incomplete VOI inputs	Keep expected value unknown, name the missing inputs, and do not recommend “measure now” from a manufactured score.

Data shape

type EvidenceItem = {
  label: string
  value: number | string
  sourceUrl?: string
  observedAt?: string
  kind: "direct" | "official" | "reference-class" | "assumption" | "inference"
  quality: "high" | "medium" | "low"
  freshness: "current" | "aging" | "stale" | "unknown"
  relevance: string
}

type ValueOfInformationItem = {
  rank: number | null
  uncertainty: string
  whyItMatters: string
  measurement: string
  expectedDecisionImpact: string
  scoreStatus: "scored" | "needs-input"
  expectedDecisionValue: number | null
  missingInputs: string[]
  stopWhen: string
}

Decisions

• 2026-07-11. Source relevance is target-specific. A reputable source can still be a weak anchor for the quantity being estimated.
• 2026-07-11. Research has a stopping rule. Additional citations may not change the decision.
• 2026-07-17. Missing VOI inputs stay unknown. Numerical defaults can produce a recommendation the evidence does not support.

Open questions

None for the current release.