Takeoff Agreement Benchmark

A gear list you can level quotes against.

A switchgear takeoff to level gear quotes against: how often Makeoff's gear list matches a working estimator's, on 32 real projects.

On this page

ResultsProject by projectDisagreementsMethodReproduce

Gear slice · type level · mean per project

Recall: the estimator’s gear Makeoff found

Precision: Makeoff’s gear the estimator priced

Projects at 100% recall

Priced gear units found, pooled

Method

The test

Why a takeoff to level against

Leveling needs a reference list that came from the drawings and not from any vendor: the estimator's own gear takeoff. Producing it is the first pass on every bid and it is skilled work, reading the one-line, the riser and every panel schedule, and it is the pass that gets squeezed when six bids are due in a week.

What was tested

Thirty-four projects were attempted and 32 were scored. They came from 13 electrical contractors who uploaded bid documents to Makeoff between March and August 2026 and attached their own takeoff to the project record. Every such pair on the platform that passed a document gate went into the set; none was chosen for how it would score.

How agreement is measured

Recall is the share of the estimator's gear that Makeoff also found: for leveling, the units the reference list would have to carry. Precision is the share of Makeoff's gear that the estimator also had: for leveling, whether the reference list is padded with units a vendor would rightly refuse to quote.

Results

Mean agreement per project, by set

The calibration set of seven projects, scored on the same rules, sits at 90.8% recall and 87.2% precision. It is lower than the reported set because it carries the hardest layouts and because it was never allowed to benefit from the review; it is shown so a reader can see that the reported number is not an artifact of tuning.

On found units the stated amperage or kVA agreed 93.5% of the time. Quantity agreed on 84.4% of found units, with the differences concentrated on the two vendor-quote projects.

Set

Recall

Precision

Calibration (7 projects)

90.8%

87.2%

Batch 1, harness alone (32)

97.2%

88.2%

Batch 1, after review (32)

99.2%

89.6%

Same harness (makeoff-bench 0.3.1), same rules hash. The calibration set was used to settle the rules and is kept out of the headline.

Project by project

Agreement on every scored project

The mean per project is driven by many small jobs that agree exactly, and the pooled figure is driven by four large ones (Publix, The Bridge, Neurona, Prologis) that carry more than half the units. Both are shown here, after review.

Sorted by recall, then precision. Thirty of 32 projects reach 100% recall; precision is where the lists differ.

RecallPrecision
0%25%50%75%100%

Hover a row for the project and reference document.

Project size against precision

The median project priced two gear units. Four large ones carry more than half the units. Bubble area is the number of projects at that point.

Hover a bubble for the projects at that point.

100%80%60%40%
0102030

Gear units priced by the estimator

100% recallRecall below 100%Bubble area is the number of projects at that point

Leveling in practice

Two references were the switchgear vendor's quote itself

Scoring Makeoff's list against a quote is the leveling job done backwards. What comes out is what a contractor leveling that quote would want to know.

Hover a unit for what it is.

Quoted, foundQuoted, credited on reviewOn the drawings, not on the quote

Disagreements

Where the two readers disagreed

Units the estimator priced that were not found: eleven of 210. They fall on six projects, and seven of the eleven sit on the two vendor-quote references, where the quote and the drawings disagree about how many units there are.

Units Makeoff listed that the estimator did not price: thirty-nine of 238, on 13 projects, classified by what the drawings say about each one.

Each block is one unit. Every disagreement is printed in the paper with its cause.

Hover a row for the cause.

Gear slice, type level · recall / precision

Appendix A: the 32 scored pairs

Pair
Project
Reference
Units
Harness alone
After review

makeoff-bench 0.3.1 · 32 scoreable of 34 attempted

Reproduce

Everything behind the figures is a file

The harness, its rule tables and their hash, the filed pairs and the reviewer decisions are published so the number can be regenerated by anyone with the files.

Harness

makeoff-bench 0.3.1

Rules hash

3e98800f311c

Regression tests

41, pinning the calibration projects