Research material: inputs, execution records, and reusable checks
Find HotAI tests and small workflow experiments with retained inputs. Each result is limited to its samples and supported by a protocol or download record. Experiments without model API execution are not presented as model rankings.
Inspect the sample and scope before reusing a test
Use these records to inspect how a specific task was executed. Read the protocol and failure cases before downloading inputs or scripts for replay. The lexical-retrieval experiment makes no language-model call, and the coding experiment covers one self-authored CSV importer. Interpret results with those limits rather than extrapolating them to whole products.
Reuse the knowledge-base experiment through traceable citations
This experiment uses nine fictional documents and eight queries to compare unfiltered lexical retrieval with supplied metadata filters. The control meets three expected cases, while filtering meets eight. Six citation cases check source fields and exact quotations. The result file retains expectations and actual outputs.
These results include no LLM or embedding model and do not test real authentication. Query metadata is supplied by the fixture, so the result does not establish arbitrary question understanding or semantic entailment. Inspect your own filters and retain failures using the protocol.
Inspect one CSV repair with a frozen contract
The coding experiment freezes eight contract checks before repairing a CSV importer that converts identifiers to integers. Six checks fail initially, and all eight pass after repair. An actual CLI output retains separate 0007 and 7 records. Inspect the patch with both source versions to understand the scope.
This coding test uses a custom CSV importer to generate defects, with Codex assisting in the fix. The test scope is limited to the attached source code and check conditions. When reusing it, keep the contract unchanged before modifying the implementation; if you relax expected results by accident, passing the test loses its original validation meaning.
What the four existing tool tests cover
The OCR test retains clear and degraded anonymous text images showing look-alike-letter and decimal-point errors. The upscaling test decodes downloaded dimensions and compares text and geometric edges. These small samples do not represent all photos, handwriting, or models.
The QR test checks generation, export, and decoding, while the text-cleaning test compares specific inputs across operations. Repeat key-field checks on your material instead of treating one anonymous passing case as universal reliability.
Keep your replay results separate from the published record
Read the README before running a downloaded script and keep a separate results directory. When replacing inputs, define new expectations instead of reusing answers from the old fixture. State untested conditions explicitly so each passing result has a clear scope.
These materials provide inspection points, while product selection still needs comparisons on the same task. Editorial updates and source reviews have separate records; downloading a file again is not a new test. Check the environment’s data handling before replacing anonymous samples with private material.