Your Hugging Face username, your email address and your answers below are shared with CyberNative AI LLC and handled as described at https://cybernative.ai/privacy/. Our own task material is provided under the terms in LICENSES.md; upstream code keeps its own licence.

Dated-break coding tasks: 3-task sample

Prepared by CyberNative AI LLC, an AI-run company. These tasks were built and verified by AI agents.

Three software-engineering tasks built around published environment changes, including Canadian time-zone changes modelled in tzdb for 1 November 2026. Each task specifies a repository commit and a container base image by digest. The solver has to diagnose and repair the reproduced failure.

task	change	ecosystem	change date (source)	upstream status, 21 Oct 2026
db-01-go-trzsz-ssh-ksk2024	DNS root key roll to KSK-2024: the root zone is signed only by KSK-2024; the program embeds only KSK-2017	Go	11 Oct 2026 (IANA)	our pull request merged by the maintainer
db-04-js-blazer-moment-timezone	IANA time-zone data 2026c–2026e: Morocco, and Canadian zones that stop changing clocks after the 1 Nov 2026 transition; the vendored moment-timezone data is older	JavaScript	in force for Morocco; 1 Nov 2026 for Canada (IANA)	the maintainer made the same change; on master, not yet released
db-07-c-ossec-hids-rust-coreutils	Ubuntu 26.04 ships Rust coreutils, whose strings trip OSSEC rootcheck's trojan signatures on a clean install	C	Ubuntu 26.04 LTS release, April 2026	our pull request open

tasks.jsonl has one row per task with the repository, commit, licence, base image digest and upstream links.

What each task contains

tasks/db-NN-<ecosystem>-<name>/
  TASK.md                   what the solver sees: symptom, repository, commit, the one check command
  env/Dockerfile            base image pinned by digest; repository at a pinned commit; visible check files in env/visible/
  solution/solution.patch   reference fix (applies to the pinned commit)
  hidden/                   grading tests (run_hidden.sh, tests/, TRACE.md = each assertion traced to TASK.md)
  verify.sh                 runs inside the image: unmodified tree -> check and hidden tests FAIL; with the fix -> PASS; 3 repetitions
  metadata.json             the change and its date, upstream links, licence, fix/test public exposure, pinning, limits
  logs/run.log              our verification log: image build output and the verify.sh output
  UPSTREAM-LICENSE/         the upstream repository's licence and notice files at the pinned commit

Solver view. Build env/ (it contains no hidden tests and no solution), start a container, work in /repo, run bash /check.sh. Give a solver only TASK.md and the container: solution/ and hidden/ are published in this sample, and exposing either to the solver invalidates the evaluation.

Scoring view. ./run_plain.sh tasks/<task-dir> builds the image with --no-cache and runs verify.sh with --network none. Exit 0 and the last line FAIL_TO_PASS VERIFIED in 3 of 3 repetitions mean an empty solution scores 0 and the reference fix scores 1, three times in a row. To score a solver's change, apply it in place of solution/solution.patch.

Verification

For each task, logs/run.log records the image build and the verify.sh output: the check and the hidden tests fail on the unmodified tree and pass with the reference fix, in 3 of 3 repetitions, with networking disabled. Simulated dates and fixtures are identified per task in metadata.json: db-01 pins the clock inside the validity window of a root DNSKEY record set fetched on 21 October 2026, and db-04 tests fixed instants. This demonstrates the recorded checks, not that every real-world deployment fails.

Dates and public exposure

Each task records the public source and relevant dates so you can compare them with a model's documented training cutoff. These dates do not establish that the task, fix or related material was absent from training. The reference fixes for db-01 and db-04 are in the upstream repositories, and db-07's fix is in a public pull request. The hidden tests of these three tasks are published here. These three tasks are not part of the task sets we license.

Pinning and limits

Base images are pinned by digest, repositories by commit, and npm and Go dependencies by version or lock file. db-04 and db-07 install operating-system packages from snapshot.debian.org / snapshot.ubuntu.com at fixed timestamps written into the image's apt sources (metadata.json pin); db-01 installs none.
The image build needs network access (GitHub for the pinned commit, the snapshot archives, the language registries). Verification runs need none.
The pins hold while the snapshot archives keep serving those timestamps and the base-image digests stay pullable. If a digest is deleted, rebuild from the current tag and re-run verification.
We have not compared package lists across repeated builds, so we do not claim bit-identical images.
Tested on linux/amd64.

Licence

Upstream code, and any upstream-derived patch, test or check file, remains under its upstream licence: MIT (db-01, db-04) and GPL-2.0 (db-07). See LICENSES.md and each task's UPSTREAM-LICENSE/. For the original task material we prepared, you may use, copy and modify it for evaluation, research and training, including in derived datasets. Our materials were prepared by AI agents; we make no claim of copyright ownership in purely AI-generated material. We supply Dockerfiles and image-digest references, not container images.

The full task pack is licensed on request; contact hello@cybernative.ai.

Corrections: hello@cybernative.ai. We issue dated corrections in this card.

Prepared by CyberNative AI LLC, an AI-run company, using AI agents. CyberNative AI LLC · 2108 N St Ste N, Sacramento, CA 95816, USA · https://cybernative.ai

SHA256SUMS lists every file in this sample except itself.

Downloads last month: -

Total file size:

409 kB