Confidence Scoring Methodology¶

How data quality is assessed and communicated.

Overview¶

Every emission estimate includes a confidence score (0.0–1.0) indicating data quality. This enables:

• Transparent uncertainty communication
• Prioritization of data improvement efforts
• Compliance with reporting standards

Confidence Scale¶

Score	Label	Meaning
0.8–1.0	Very High	Primary data from supplier
0.6–0.8	High	Published secondary data
0.4–0.6	Medium	Research-based estimates
0.2–0.4	Low	Model extrapolation
0.0–0.2	Very Low	Fallback estimates

Factor Sources¶

Confidence depends on how factors were derived:

Source	Typical Confidence	Description
measured	0.8+	Direct measurement by provider
research	0.5–0.7	Academic research on similar models
estimated	0.3–0.5	Extrapolation from model characteristics
fallback	0.1–0.2	Generic estimate for unknown models

Calculation¶

Per-Trace Confidence¶

Each trace inherits the factor's confidence:

trace.confidence = factor.confidence

Aggregated Confidence¶

Summary confidence is token-weighted average:

avg_confidence = Σ(trace.confidence × trace.totalTokens) / Σ(trace.totalTokens)

Weighting by tokens ensures high-volume models dominate the average.

GHG Protocol Mapping¶

Confidence maps to GHG Protocol Data Quality Score (DQS):

Confidence	DQS	GHG Protocol Description
≥0.8	1	Primary data from suppliers
≥0.6	2	Published secondary data
≥0.4	3	Average secondary data
≥0.2	4	Estimated data
<0.2	5	Highly uncertain

function confidenceToDataQuality(confidence: number): 1 | 2 | 3 | 4 | 5 {
  if (confidence >= 0.8) return 1;
  if (confidence >= 0.6) return 2;
  if (confidence >= 0.4) return 3;
  if (confidence >= 0.2) return 4;
  return 5;
}

Uncertainty Conversion¶

For ISO 14064 reporting, confidence converts to uncertainty bounds:

Confidence	Uncertainty	Range
≥0.7	±15%	85%–115%
≥0.5	±30%	70%–130%
≥0.3	±50%	50%–150%
<0.3	±100%	0%–200%

function confidenceToUncertainty(confidence: number): { lower: number; upper: number } {
  if (confidence >= 0.7) return { lower: 0.85, upper: 1.15 };
  if (confidence >= 0.5) return { lower: 0.70, upper: 1.30 };
  if (confidence >= 0.3) return { lower: 0.50, upper: 1.50 };
  return { lower: 0.00, upper: 2.00 };
}

Display¶

CLI Output¶

Environmental Impact [PASSIVE]
  Grid carbon: 400 gCO₂/kWh (default)  |  Confidence: low (32%)

Dashboard¶

Confidence shown as colored badge: - 🟢 High (≥60%) - 🟡 Medium (≥40%) - 🔴 Low (<40%)

Export¶

All exports include confidence/DQS fields:

{
  "confidence": 0.32,
  "dataQualityScore": 4,
  "uncertainty_percent": 50
}

Improving Confidence¶

1. Provider Data¶

If a provider publishes emission data: - Update factor with new values - Set source: "measured" - Increase confidence to 0.8+

2. Academic Research¶

When new research available: - Validate against existing factors - Update if significant difference - Document source

3. Direct Measurement¶

For dedicated deployments: - Measure actual power consumption - Apply real grid carbon - Set confidence to 0.9+

Current Status¶

Most AI providers don't publish per-request emissions:

Provider	Data Available	Typical Confidence
Anthropic	No	0.25–0.35
OpenAI	No	0.25–0.35
Google	Partial	0.35–0.45
Others	No	0.15–0.25

Confidence Philosophy¶

Conservative by Default¶

When uncertain, we overestimate emissions: - Larger model size assumptions - Higher energy per token - Average (not clean) grid carbon

This ensures reported emissions are upper bounds.

Transparent Uncertainty¶

Users always know data quality: - Confidence displayed prominently - Uncertainty ranges in exports - Methodology documentation

Continuous Improvement¶

Track confidence over time: - Annual factor review - Provider engagement - Research monitoring