A more complete dependency graph across 9 languages. Merget leads recall in 11 of 12 tested repositories and verified precision in 10.
Of the calls the reference lists, the share a graph also contains. Low recall means missed calls.
Of a graph’s edges that the reference could check, the share it did not contradict. Low precision means invented or misdirected calls.
A compiler-grade index of each repository (SCIP) that records where every call goes, built by a third-party indexer that uses the language’s own compiler or type checker. Both graphs were scored against the same one.
Merget finds more reference calls in 11 of 12 repositories. Inspect both coverage and verified precision for your stack.
How to read thisEach pair of bars is one tool’s score from 0% to 100%; longer is better. The diamond marks the higher score.
Full results: edge counts per repository
Merget on every repository, with the counts behind each rate.
| Repository | Language | Reference facts | Recall | Precision | Judged | Wrong | Edges |
|---|---|---|---|---|---|---|---|
| TypeScript | TypeScript | 45,011 | 94.1% | 99.8% | 42,622 | 90 | 43,325 |
| prometheus | Go | 26,610 | 95.4% | 99.4% | 25,529 | 143 | 25,529 |
| erpnext | Python | 23,408 | 81.9% | 99.0% | 19,374 | 189 | 22,395 |
| redis | C | 20,497 | 80.3% | 99.2% | 17,923 | 152 | 18,983 |
| django | Python | 9,998 | 87.7% | 98.6% | 8,931 | 126 | 10,436 |
| ripgrep | Rust | 5,449 | 89.1% | 97.8% | 4,965 | 111 | 5,000 |
| commons-lang | Java | 4,699 | 90.1% | 99.4% | 4,258 | 26 | 4,258 |
| eslint | JavaScript | 4,561 | 86.1% | 98.4% | 4,363 | 70 | 6,364 |
| AutoMapper | C# | 1,204 | 83.6% | 98.7% | 1,640 | 21 | 2,683 |
| cats | Scala | 345 | 71.6% | 97.6% | 253 | 6 | 268 |
| flask | Python | 253 | 94.9% | 96.8% | 252 | 8 | 264 |
| express | JavaScript | 56 | 100% | 91.8% | 61 | 5 | 597 |
graphify on the same repositories and reference facts.
| Repository | Recall | Precision | Judged | Wrong | Edges scored | Links not scored |
|---|---|---|---|---|---|---|
| TypeScript | 66.0% | 95.7% | 31,127 | 1,353 | 31,303 | 384 |
| prometheus | 44.0% | 88.0% | 13,318 | 1,604 | 13,318 | 1,103 |
| erpnext | 78.2% | 95.0% | 19,270 | 971 | 19,492 | 1,705 |
| redis | 86.5% | 99.7% | 17,969 | 57 | 18,912 | 436 |
| django | 59.9% | 79.4% | 7,550 | 1,554 | 7,753 | 1,260 |
| ripgrep | 35.4% | 62.1% | 3,111 | 1,179 | 3,142 | 171 |
| commons-lang | 42.8% | 78.7% | 2,557 | 545 | 2,558 | 340 |
| eslint | 52.4% | 95.6% | 2,530 | 112 | 2,664 | 217 |
| AutoMapper | 39.5% | 78.1% | 635 | 139 | 1,480 | 1,677 |
| cats | 4.6% | 11.9% | 135 | 119 | 150 | 1,253 |
| flask | 64.0% | 92.1% | 177 | 14 | 191 | 5 |
| express | 55.4% | 93.9% | 33 | 2 | 36 | 4 |
When either tool says an edge is certain, Merget’s are contradicted less often in 11 of 12 repositories.
Certain means each tool’s most-confident label: precise for Merget; EXTRACTED for graphify, its label for a link found directly in source.
Per repository: certain edges the reference contradicted
Merget’s certain edges are contradicted less often than graphify’s in 11 of 12 repositories.
How to read thisEach bar is the share of a tool’s certain edges that the reference contradicted, out of those it could check; shorter is better. The track ends at 40%; graphify’s cats bar (88.1%) runs past it.
Full results: certain edges
Recall and precision counted over each tool’s certain edges only, with how many of the checked certain edges were contradicted.
| Merget, certain (precise) edges | graphify, certain (EXTRACTED) edges | |||||||
|---|---|---|---|---|---|---|---|---|
| Repository | Recall | Precision | Edges | Contradicted | Recall | Precision | Edges | Contradicted |
| TypeScript | 88.4% | 100% | 40,553 | 11 of 39,920 | 63.9% | 95.5% | 30,365 | 1,345 of 30,205 |
| prometheus | 87.3% | 99.6% | 23,332 | 100 of 23,332 | 29.5% | 83.8% | 9,376 | 1,522 of 9,376 |
| erpnext | 78.5% | 99.6% | 21,224 | 77 of 18,453 | 55.2% | 97.2% | 13,347 | 366 of 13,287 |
| redis | 80.2% | 99.2% | 17,465 | 141 of 17,153 | 30.5% | 99.6% | 6,690 | 27 of 6,360 |
| django | 87.5% | 98.9% | 10,120 | 98 of 8,873 | 45.6% | 75.3% | 6,220 | 1,498 of 6,060 |
| ripgrep | 84.5% | 99.5% | 4,637 | 24 of 4,628 | 33.1% | 60.5% | 3,011 | 1,176 of 2,980 |
| commons-lang | 81.6% | 99.8% | 3,844 | 9 of 3,844 | 39.4% | 78.2% | 2,370 | 516 of 2,369 |
| eslint | 81.5% | 99.3% | 4,384 | 27 of 4,105 | 52.2% | 95.6% | 2,656 | 112 of 2,522 |
| AutoMapper | 68.1% | 99.2% | 2,269 | 11 of 1,447 | 16.6% | 69.6% | 1,056 | 95 of 312 |
| cats | 64.3% | 100% | 230 | 0 of 222 | 4.6% | 11.9% | 138 | 118 of 134 |
| flask | 94.9% | 97.2% | 260 | 7 of 251 | 64.0% | 92.6% | 190 | 13 of 176 |
| express | 100% | 100% | 82 | 0 of 56 | 55.4% | 93.9% | 36 | 2 of 33 |
Merget reaches 76% of later fix targets, versus 49% for graphify. Two calls out from changed code, on 30 scorable Cargo merges.
A graph helps only if walking it leads somewhere. We took 38 merges in cargo, 30 of which a later human fix can score, started from the code both sides changed, and followed each tool’s call edges outward. No model is in the loop.
Full results: reaching the fix
All 30 scorable merges: the share of fix targets each graph reaches, and the merges where it reaches at least one.
| Depth | Share reached, Merget | Share reached, graphify | Reached any, Merget | Reached any, graphify |
|---|---|---|---|---|
| 1 call out | 52% | 15% | 21 of 30 | 17 of 30 |
| 2 calls out | 76% | 49% | 28 of 30 | 25 of 30 |
| 3 calls out | 90% | 61% | 30 of 30 | 28 of 30 |
More of the dependency graph helps a merge queue see where PRs collide. These results support Merget’s analysis foundation across the tested stacks.
Merget’s full graph had higher recall on every tested repository in TypeScript, Go, Python, Rust, Java, JavaScript, C# and Scala.
The merge queue uses this graph to see which open pull requests touch the same definitions and callers, and to give conflict resolution the context it needs.
See it on your code. Start in Advisory mode and read Merget’s findings on your own pull requests.
Sign up freeMeasured October 13, 2026 on 12 open-source repositories at pinned commits. Both graphs were scored against the same SCIP reference index for each repository; graphify ran in its code-only mode.
Merge conflict resolution: 223 files, one call each