# BFCL Baseline Reconstruction Report

Status: baseline reconstruction only. No buyer-profile weights were applied. No paid API calls were made.

## Result

- Published aggregate formula reproduced: YES - reproduced to the official two-decimal precision
- Models included: 109
- Maximum discrepancy: 0.01 percentage points
- Exact two-decimal matches: 98
- Matches within 0.01 percentage points: 109
- Required baseline missing-data issue: none

## Formula Used

Published BFCL V4 overall score was reconstructed from downloaded `data_overall.csv` fields as:

`Agentic = (Web Search Acc + Memory Acc) / 2`

`Overall Acc = 0.40 * Agentic + 0.30 * Multi Turn Acc + 0.10 * Live Acc + 0.10 * Non-Live AST Acc + 0.10 * Irrelevance Detection`

This follows the documented V4 composition: Agentic 40%, Multi-Turn 30%, Live 10%, Non-Live 10%, and Hallucination Measurement 10%. The downloaded public CSV exposes hallucination measurement in the aggregate through `Irrelevance Detection`; substituting `Relevance Detection` does not reproduce the published aggregate.

## Treatment Of N/A And Missing Categories

BFCL documentation states that unevaluated categories appear as `N/A` and summary columns treat unevaluated categories as zero during calculation. The reconstruction used zero for missing required formula fields, but the current downloaded `data_overall.csv` had no missing required formula fields among eligible rows.

Format sensitivity was not included in the reconstructed aggregate because BFCL documents it as non-scoring/no-score-impact. In `data_overall.csv`, `Format Sensitivity Max Delta` is missing for 69 rows; this is expected for many FC/native tool-call variants and does not affect the baseline formula.

Cost and latency were retained in the processed baseline but were not used in the published aggregate reconstruction.

## Discrepancies

All eligible rows reproduced within 0.01 percentage points. The small nonzero differences are consistent with using rounded public component columns to reconstruct a rounded published aggregate.

## Reliability

The baseline is reliable enough for sensitivity analysis if subsequent audit steps retain the frozen raw CSVs and use the same missing-data rules. Buyer-profile weighting should not begin until the frozen manifest and processed baseline are treated as immutable inputs.