Skip to content

Historical v0.2.0 evaluation result

Evaluation evidence. This page reports only the independently reviewed real-project-v1 evaluation partition frozen in UsageBench v0.2.0. It does not pool the broader analyzer-informed development corpus.

The generated evidence map is the site-wide index of the frozen v1 and v2 prospective slices, the separately reviewed legacy core, controls, overflow, and remaining development cases. This page remains the detailed historical v1 result; it does not become a proxy for the other strata.

The active development candidate is public Bifrost v0.10.1 at immutable commit 511adaa2733067bb1b7809ab79e06ec0e3d2a146. No result for that release is published on this page: the tables below remain historical v0.2.0 evidence produced with Bifrost v0.8.8, and any upgraded result must be published as a subsequent snapshot or release with its own provenance.

The first immutable UsageBench evaluation compares Bifrost with gopls, Pyright, and TypeScript language server across 12 source-only sampled public repositories and 36 reviewed cases. The claim is deliberately bounded to these repositories, profiles, and the recorded References and Definition operations; it does not estimate language-wide accuracy, latency, memory, or cold-start performance.

Release fact Value
Freeze real-project-v1, evaluation v0.2.0
Source revision 6ea6056fa6b3eb52a656a2b4a62c57956771de78
Evidence Release archive and checksum
Manifest SHA-256 3258e3269c98baa980883e969df97e1c9f4920503072ca8797e2600e2b268614
Publication run Freeze benchmark snapshot
Language Reference profile Repositories Cases Population exclusions Source-review replacements
Go gopls 4 12 5 (small repository × 1; missing root build marker × 2; source size over 150 MiB × 4) 6 (independent source review rejected every earlier assigned candidate × 6)
Python pyright 4 12 35 (small repository × 13; missing SPDX license × 13; missing root build marker × 19; source size over 150 MiB × 18) 4 (independent source review rejected every earlier assigned candidate × 4)
TypeScript typescript-language-server 4 12 27 (small repository × 1; missing SPDX license × 8; missing root build marker × 3; source size over 150 MiB × 23; truncated source tree × 1) 2 (independent source review rejected every earlier assigned candidate × 2)

Exclusion reasons can overlap, so reason counts are not repository totals. The selection, replacements, reviews, and adjudication are hash-bound in the release audit rather than reconstructed from these summary counts.

Artifact Frozen source SHA-256
Protocol protocol.json 9780dc32b34de862f51588bcdf036a004b9a4de76a3c4c52eb667fbbc51ad501
Selection selection.json 964f23de0130dfd48f38b12bfa35806659be12176b828fdc97fa7ef3dfcc46e8
Independent review review.json a801484c37f9071db95359050db098ec282bb75b64e94cf87e8df2cd66a2ad6c
Source lock sources.json 700a9b6927d2fbb57fae005b3f95057132902fba9eeb86ef7deb73db77cde45e
Sol review openai-gpt-5.6-sol.json 4f77c34ba904ea8ad2652d518ff750f66cc32b789d4a4cf6c947c88c8a7486cd
Fable review anthropic-claude-fable-5.json a26abea9f9f265d262d3440bbd82df96a489c960b41e7b07557781d84eae8637
Human adjudication adjudication.json 2e1ce6a986e399330b73e9558d3267a113d737e3d77038d2c6b2927f342f6900
Candidate Requested version Environment Report SHA-256
Bifrost a54be9be9b08b9d9ddbab1c471e26d7f8bd932df Linux x86-64 canonical container 702a73c97d86a4c9c6b2b576c3d411306788581e4659b1ff9e82fae491c7459b
gopls 0.23.0 Linux x86-64 canonical container 679796a7a8f3d5b5eb44cad9a9ef06a2b512f4d53fdd503414842a4024026b94
Pyright 1.1.411 Linux x86-64 native c6899baca9154f464c7baefe1984d0078175813138d6c428234f3573d0665099
TypeScript language server 5.3.0 with TypeScript 5.9.3 Linux x86-64 native 8dda75467e0bd2498905abb8e8e6702c185e9af5954c5513a55248ecda9a2ad4

This view asks whether each tool reached every required destination for a case. It does not forgive extra results; those are exposed by the location-level and strict-contract tables below.

Reference profile Shared Bifrost found Reference found
gopls 0.23.0 12 8/12 (66.7%) 9/12 (75.0%)
Pyright 1.1.411 12 8/12 (66.7%) 11/12 (91.7%)
typescript-language-server 5.3.0 (TypeScript 5.9.3) 12 10/12 (83.3%) 8/12 (66.7%)

TP, FP, and FN are reported without true negatives. Strict precision counts every extra result; policy-adjusted precision excludes authored and policy-allowed extras.

Reference profile Analyzer Cases TP FP FN Destination recall Exact-token recall Strict precision Policy-adjusted precision Exact-set case rate Extras/success
gopls 0.23.0 Bifrost 12 71 0 6 92.2% 48.1% 100.0% 100.0% 75.0% 0.00
gopls 0.23.0 gopls 12 60 0 17 77.9% 77.9% 100.0% 100.0% 75.0% 0.00
Pyright 1.1.411 Bifrost 12 73 2 20 78.5% 48.4% 97.3% 97.3% 58.3% 0.05
Pyright 1.1.411 Pyright 12 90 1 3 96.8% 96.8% 69.2% 98.9% 33.3% 1.70
typescript-language-server 5.3.0 (TypeScript 5.9.3) Bifrost 12 43 0 11 79.6% 66.7% 100.0% 100.0% 83.3% 0.00
typescript-language-server 5.3.0 (TypeScript 5.9.3) typescript-language-server 12 43 2 11 79.6% 79.6% 79.6% 95.6% 41.7% 0.56

Strict conformance requires the complete reviewed location set and exact token ranges. A separating result is a contract disagreement, not automatically an analyzer defect.

Reference profile Shared Both exact Bifrost only Reference only Neither
gopls 0.23.0 12 7 1 2 2
Pyright 1.1.411 12 6 1 4 1
typescript-language-server 5.3.0 (TypeScript 5.9.3) 12 5 5 2 0
Total 36 18 7 8 3

The case comparison lists every case where exactly one side satisfies the strict contract. The broader 24 July development result remains available as historical regression evidence and is not part of these denominators.