Kaelum Bench v0.2 · Public Calibration
Accounting Capstone · Run evidence
Single configuration-specific calibration run. Not a verified v1 result or statistical ranking. The model-judged component is one stored outcome. The current public release has no model-judge repeatability qualification. A fresh judge call is not guaranteed to reproduce it.
Methodology →Run summary
Score
90.7/ 100
Raw score 906.5 / 1000
Successfully scored responses
Tested-model inference only
Independent AI evaluation
Model plus evaluation
Configuration-specific
Public calibration
Cost composition
Recorded run cost
Run identity
Output health
Valid outputs
98 / 100
98.0%
Scoring process
Hybrid scoring combines deterministic checks with an independent AI rubric assessment. The judge evaluates tasks that require rubric-based review; it does not generate the tested model's answers.
No signed public judge-policy record is available for this release.
Run identity
Material configuration parameters
Scoring and usage
Judge evidence
Run provenance
These timestamps delimit the recorded wall-clock lifecycle span. They do not measure active model execution time or comparable latency.
Authenticated publication
Most visitors only need the run JSON. The other files let auditors and developers independently verify the publication chain behind this result.
Start with this file
Recommended · this run
A sanitized, machine-readable copy of this exact result: model configuration, score, usage, execution, costs, judge summary and provenance. Best for reviewing, archiving or importing this run. The sanitized single-run JSON is a derived projection and is not separately signed.
Advanced downloads
These files work together. A single verification file is not useful on its own.
Release inventory
The signed inventory of this release. It lists every published file together with its SHA-256 digest and byte size.
Authenticity proof
The detached signature over the exact manifest bytes. It lets the public detect whether the signed manifest was changed.
Verification key
The public verification key used to check the signature locally. It cannot be used to create a valid Kaelum signature.
All public runs
A machine-readable directory of all 126 public runs, including their run IDs, benchmark versions and configuration references.
Benchmark dataset
The complete published dataset for Accounting Capstone, covering all model runs in this benchmark. Its hash and byte size are bound by the signed manifest.
The Ed25519 signature authenticates the exact public manifest bytes, including the release inventory. The manifest in turn binds this run's source aggregate by its SHA-256 digest and byte size.
node -e 'const fs=require("node:fs"),c=require("node:crypto"),m=fs.readFileSync("public-manifest.json"),s=JSON.parse(fs.readFileSync("public-manifest-signature.json","utf8")),k=fs.readFileSync("kaelum-bench-public-release-20260718-v1.pem");console.log(c.verify(null,Buffer.concat([Buffer.from("kaelum-bench-fmb-public-release-v1\0"),m]),k,Buffer.from(s.signature_base64,"base64")))'