glayguo/evalarc-casebook

EvalArc Casebook

The same 93.75% score can pass one acceptance gate and fail another. Inspect the rules, actual failed checks and original Docker records in a filterable table. This is the data companion to the interactive evidence lab.

In the default suite_jobs view, compare support-partial and support-protected. Both use the same frozen defective policy, score 93.75% and fully resolve 0/2 attempts. The deliberately permissive rule accepts partial progress; the rule requiring every notes check rejects it. gate_accepted and fully_resolved are separate columns.

What is included

Configuration	Rows	Unit and purpose
suite_jobs	3	One configured gate per job, backed by five Docker attempts; compare acceptance and full resolution.
audit_cases	167	One case execution per scripted control: 135 coding and 32 support cases across two references and 15 declared faults.
repetition_attempts	6	One recorded attempt of a frozen support policy; three reference and three faulty attempts.

Each configuration has a single development split. These are different units, so their row counts must not be summed into a number of independent trials or benchmark examples. evaluation_score in the case table is the parent evaluation's score, repeated for navigation; averaging that column across case rows would reweight evaluations incorrectly.

These are saved scripted controls on public development tasks, not runs of a trained language model. Support tickets, identifiers and messages are synthetic fixtures. No customer data or private logs were collected.

Use without the web viewer

from datasets import load_dataset

jobs = load_dataset("glayguo/evalarc-casebook", "suite_jobs", split="development")
for job in jobs:
    if not job["fully_resolved"]:
        print(job["id"], job["mean_score"], job["gate_accepted"])

No dataset loading script, candidate execution or model credential is required. You can also download the three JSONL files and read them with Python's standard json module. The datasets package is only needed for the example above.

Inspect a row's evidence

source_file points to an unchanged JSON file in this dataset repository. source_pointer is a JSON Pointer locating the exact original object; an empty pointer means the whole evaluation. source_sha256 authenticates that file. source_url links the corresponding file at a fixed GitHub commit.

All configurations preserve candidate, grader and case fingerprints and the recorded runtime. The case table additionally includes original check outcomes and case_json; this retains available support states and tool traces. The suite configuration, plan, all five attempts and JUnit are included under evidence/examples/suite/. Repetition summaries and all six attempts are included under their original example directories.

The original records come from EvalArc 0.2, 0.4 and 0.5. Their recorded_evalarc_version and grader fingerprints are preserved. Do not treat the three configurations as matched version comparisons; use the lab's separate matched comparison for that question.

Build source: f1b5272c86c518b1c56c8a566d2d3d4a73ea6635. manifest.json lists every exported file hash.

git clone https://github.com/noteflowai/evalarc.git
cd evalarc
git checkout f1b5272c86c518b1c56c8a566d2d3d4a73ea6635
python3 scripts/build_dataset.py --output dist/casebook

The build validates evaluations, recomputes repetitions and suite decisions, and checks JUnit before exporting. It does not rerun candidates. See the methodology for the task contracts and limitations.

Intended uses and limits

Use this small casebook to learn grader auditing, inspect retries and idempotency, test report readers, and discuss acceptance criteria. Reference implementations and deliberate faults were authored for these tasks; this is not an exhaustive collection of possible defects.

The repeated controls show no observed check variation. They do not establish independence, a population reliability estimate or performance on unseen tasks. The dataset does not support model rankings, arbitrary reward-hack resistance or RL improvement claims. Public development cases should not be presented as a held-out benchmark. No hosted CI importer was exercised.

Authorship and license

Published by the EvalArc maintainer, with AI-assisted development and writing. Code, task fixtures and these derived records are MIT-licensed; see LICENSE. No third-party model weights are included.

Related maintainer projects: Robot Reel inspects recorded Physical AI experiments; Skills Anywhere provides reusable skill discovery and file checks. The project collection groups these independent tools; it does not imply a shared model evaluation or upstream endorsement.