Feature: Consumer overage calibration mode (claude.ai Max/Pro)

[clawSean]

Problem

Tide Pools currently estimates spend using standard API pricing, but users running Anthropic via consumer plans (e.g., claude.ai Max extra usage) see billing that diverges from API-rate estimates.

Result: both numbers can be internally consistent, but operationally confusing.

Goal

Add a calibration layer that can learn effective consumer overage rates from real billing totals and produce a more accurate spend estimate.

Proposed solution

1. Keep existing API-rate estimate as baseline (estimated_charge_api).
2. Add optional calibration input (daily/weekly actual charge totals from claude.ai usage/billing).
3. Fit effective pricing coefficients by model (plus optional global overhead term).
4. Emit a second metric (estimated_charge_calibrated) with confidence/error bounds.

UX / Output

• API estimate: $X
• Calibrated estimate: $Y
• Calibration error: ±Z%
• Last calibration update timestamp

Data model sketch

• usage_by_model_per_day (already derivable from session logs)
• actual_charge_per_day (user-entered/imported)
• Fitted params: {model_rates, intercept, confidence}

Acceptance criteria

• Users can supply at least 7 days of actual charge totals.
• Tide Pools reports both API and calibrated estimates.
• Report includes confidence/error band and warns when calibration data is stale.
• Calibration can be disabled (fallback to API estimate only).

Notes

• This should be provider-agnostic where possible, but Anthropic consumer overage is the immediate driver.
• Keep implementation lightweight and transparent (no black-box surprises).