Volunteer compute tasks
Turn donated agent compute into useful improvements whose evidence costs much
less to check than to discover. The operator commissioned this workshop on
2026-09-26, requested three choices, later added a fourth, and specified an ordinary repository
checkout with software installed in .venv.
The operator's selection criterion is substantial work with cheap verification. A quotation audit was rejected because judging each alleged mismatch could cost the maintainer as much as discovering it cost the contributor. These tasks instead return executable counterexamples or measured improvements. None guarantees that more compute will find a result.
All tasks must be grounded in committed repository content and runnable from a fresh checkout. The operator excluded review-state work because the operational review history is not committed. Do not substitute synthetic review history for that unavailable evidence.
Choose one
| Task | Main work | What the maintainer checks |
|---|---|---|
| Relocation stress search | Generate and shrink combinations of KB graphs and move sequences | A small graph, a command sequence, and a broken preservation property |
| Validator defect-detection search | Mutate committed artifacts and generate combinations of explicit structural violations | A small before/after fixture and a required diagnostic the validator misses |
| Faster collection validation | Profile, optimize, and compare the validator across corpus shapes | Identical diagnostics, repeatable timings, and a bounded patch |
| Link recognition differential | Compare each component's link recognition with the site renderer, then propose one shared recognizer | A small Markdown fixture with disagreeing results; a function checked against the same cases |
Each task can absorb sustained search and experimentation. Choose by interest; doing more than one is not expected. Search strategy, tools, and allocation of the contributor's available compute are left to the contributor. The third task offers a numeric success measure but also requires code review before merging. The fourth task's second stage likewise ends in a patch that needs code review.
Shared contributor handoff
Read this page and the chosen task together. Read root AGENTS.md for repository
conventions. This commission explicitly replaces its user-level uv tool
installation and uv run test instructions for this contribution: use the
activated .venv and python -m pytest. Do not change AGENTS.md to make that
exception global. Do not run commonplace-init in this checkout.
Use Python 3.11 or newer and Git. In a fresh POSIX shell:
git clone https://github.com/zby/commonplace.git
cd commonplace
git checkout -b volunteer-compute 8eb5c3a6aea83bcd6c2908b446709b482d5d45a6
python3 -m venv .venv
source .venv/bin/activate
python -m pip install -e '.[docs]' pytest ruff
python -c "import sys, commonplace; print(sys.executable); print(commonplace.__file__)"
commonplace-validate --help
On Windows, use py -3 -m venv .venv and activate with
.venv\Scripts\Activate.ps1; the remaining commands are the same. The docs
extra supplies dependencies used by the wider test suite. Record Python,
platform, git rev-parse HEAD, and python -m pip freeze with the result.
Additional search dependencies are permitted inside .venv; list them and
their versions. Final reproductions should need only the installed package
and pytest where practical.
The commit above is the common baseline, not a moving branch. Keep these task instructions separately available: the baseline predates the workshop. If it cannot be fetched or installed, report that blocker rather than silently switching revisions. No API credentials, live review service, private snapshots, or maintainer database are needed.
Run the chosen task's existing tests before searching. Report baseline failures separately from discoveries. Create all experimental KBs in temporary directories. Invoke relocation commands from the temporary project root; they use the working directory to select the project.
Ownership and result format
Work in your own branch. Add reproducible tests under tests/commonplace/ and
reusable search or benchmark tooling under scripts/volunteer_compute/. Keep
the submission summary under this workshop in a contributor-named subdirectory.
Do not edit the real KB corpus as experimental input. Tasks 1 and 2 initially
own tests, tooling, and reports only; fixes, if supplied, must be separate
patches. Task 3 additionally permits a bounded validator implementation patch.
Task 4 permits a new link-recognition function and one caller adoption, as
separate patches.
The maintainer owns acceptance, integration, and any changes to the contracts.
For a bug finding, supply the baseline SHA, practical consequence, exact contract passage or existing contract test, minimal fixture, one reproduction command, expected result, and actual result. The regression assertion should fail on the baseline for the claimed reason. Include enough before/after state to inspect the failure without reading the search tool. A small positive control should distinguish the failure from broken setup.
Deliver at most five distinct defects, ranked by consequence and deduplicated by cause. Put ambiguous requirements in a separate short list. A test that merely asserts a contributor's preferred behavior is not a confirmed bug. Record seeds and search bounds for reproducibility, but do not make rerunning the entire search a prerequisite for checking a finding. Aim for each final counterexample to replay in under a minute on an ordinary laptop.
Start with one complete result packet. It should be reviewable on its own; continue independent search within the chosen scope while awaiting feedback. Stop at the contributor's stated budget or deadline, after five distinct findings, or when progress requires a contract decision. Return partial results and blockers. A search with no counterexample reports its tested domain and cost, not a claim that the implementation is correct. Report approximate time and model/token usage if available; private conversation logs are unnecessary.
For a proposed implementation fix, the existing full suite must pass with
python -m pytest, together with the relevant new regression tests. A
reproduction-only submission is expected to contain a failing new test; keep
that distinct from the passing baseline suite.
Maintainer evaluation and closure
Record replay outcome, acceptance decision, and approximate checking time for each submitted result. A cheap replay does not remove the need to inspect the contract or review a patch. If adjudication requires reconstructing the search or reading a long report, narrow the packet before commissioning more work.
The relocation refactor workshop owns architectural consolidation; this workshop commissions evidence about existing behavior. Validator defect findings and performance patches remain separate: the performance comparison preserves baseline output, including known defects; a correctness fix changes that output under its own regression test. The graph-loader workshop explores a broader architecture; performance work here does not authorize that redesign. It also owns where a shared positioned link representation lives; task 4 supplies evidence and a candidate recognizer for that decision.
Close after the contributed results have been accepted, rejected, or deferred with reasons, useful tests and tooling have been integrated, and any lesson about verification cost has been retained in the appropriate library artifact. Remove this workshop and its active-list entry when that work is complete.
Complete file listing (generated at build time)