[Cooki-e]

Summary

Consolidate the current revisions of 13 tasks, including the review fixes from
#16, #18 and #23, the revised public/hidden splits, two reward-baseline changes,
and the existing Luna evaluation evidence. The prior #18 and #23 commits are
included intact via merge commits so their authorship is preserved. This PR
offers one combined review surface; #18 and #23 are now closed as superseded. No shared infrastructure, shared task inventories or unrelated tasks
are changed.

• Reject malformed/nonbinary trajectory-judge results in biomedical, legal,
  negative and long-PDF finalizers while preserving valid fractional rewards.
• Enforce filtered ANN build/service boundaries: terminate detached submission
  processes, clear and verify their System V IPC, restrict service filesystem
  access with Landlock, and reject nonregular index files. Expand the split to 10 public / 25 hidden queries.
• Add lightweight's trajectory transfer, isolated audit execution and final
  reward gate, with CPU-only gate and isolation regression tests.
• Expand biomedical public validation from 3 to 8 independently verified
  examples while preserving its graph and 20 hidden queries.
• Use the approved DeepSeek evidence judge and 30-second query limit for
  long-PDF; keep its agent-facing request concise.
• Apply the current methodology baseline of 0.310109 and quantity baseline of
  0.1838. Preserve their original raw metrics and distinguish offline reward
  recalculation from a new model trial.
• Carry the current structured instructions and add task-local RUN_RESULTS.md
  records with model/effort, version boundaries, judge health, artifact hashes
  and source locations. Existing construction records remain historical.

Task authors in task.toml are preserved. The contribution is submitted by
Cooki-e. CPU tasks are biomedical, code-context, filtered ANN, join-aware,
long-PDF, negative and paper-metadata. Chunqiu, decision, legal, lightweight,
methodology and quantity retain their configured GPU resources. No resource
budget, quality threshold or query deadline is relaxed during this publication.

Published inputs

HF PR #8
is merged at 234406eb7e304478e48905829fc0d3233f5a9f1f.

Only four new runtime data files (17,690 bytes) were needed: the biomedical
v0.2.2 and filtered-ANN v0.11.0 public query/answer pairs. They use new versioned
filenames; all 52,417 previous manifest entries and their assets are preserved.
Uploaded metadata and file bytes were read back, and all four files were
downloaded through a fresh cache and verified against exact sizes/SHA-256.
Unchanged corpus/model files retain their existing published pins. No model
weights, hidden evaluation inputs, author solutions, private trajectories or
credentials were uploaded to HF.

Filtered ANN IPC review fix — 2026-10-23

Version 0.11.1 removes System V shared memory, message queues and semaphore
sets after all submission processes exit. Both owner and creator UIDs are
checked, so changing ownership cannot preserve state between build, recovery
and relocation. Cleanup runs as the dedicated submission user, is bounded,
and verifies that no matching object remains. Inspection/removal failure or
residual state aborts the stage before another service starts.

• The old verifier failed all 12 real-kernel boundary/ownership subcases.
  The patched filtered-ANN suite passes 22 tests, 0 skipped, including
  owner-only/creator-only matches, preservation of unrelated IPC, and injected
  inspection/removal/timeout/residual-object failures.
• The real Faiss fixture still passes all 50 original/relocated requests.
  Actual subprocess recovery with IPC at every load still yields 22/25 after
  three deliberate query failures, with 50 requests and eight service loads.
  Cleanup failures at build, recovery and relocation stop the next process.
• Final task static and selected-package release checks: zero errors/warnings;
  Python syntax and git diff --check: passed. Tests used private IPC namespaces,
  two CPUs, 2 GiB RAM, no network, no GPU and no API credentials.
• Instructions, public/hidden data, asset pins, scoring and the 15 ms deadline
  are unchanged. No HF publication, fresh dependency install, full-corpus build,
  coding-model trial or live judge call was run for this fix. The 0.20 score is
  still the earlier v0.11.0 saved-program replay, not a new v0.11.1 result.

Validation evidence

• At the initial PR head, 199 task-local regression tests passed, counting each test once.
  Filtered ANN passed 19 tests in a disposable CPU container, including real
  detached-process cleanup and Landlock checks. Lightweight passed all 21
  tests in a disposable CPU container, including shell-gate and isolation tests.
  Both containers had no network, GPU or API credentials.
• The initial host filtered-ANN test could not import Faiss; its complete suite
  passed using the already installed verifier dependencies in the container.
  Join-aware has no test_*.py suite; its existing verifier evidence is retained.
• python scripts/check_release.py: 0 errors, 0 warnings.
• For each of the 13 task packages, python scripts/check_tasks.py tasks/<name> --merge-ready --base 125d63be79c04903e06cc6f6896bfe667351d22e:
  0 errors. Checks use process-local core.quotepath=false for the original
  Unicode corpus paths; no Git configuration or corpus filenames are changed.
• git diff --check: passed. Changes are confined to these task packages.

Existing coding model: Codex + OpenRouter openai/gpt-6-luna, effort max.
This publication did not launch another agent, GPU training job or paid judge.
Each task's RUN_RESULTS.md records the evaluation/reward hashes and the
author-local evidence paths. Full sensitive evaluation artifacts remain outside
the public task packages.

Task (tasks/…)	Current version	Available score	Evidence boundary
biomedical-graph-retrieval	0.2.2	0 (historical)	v0.2.1 trial with 3 public examples; new 8-example version not rerun
chunqiu-temporal-retrieval	0.1.0	0.005685	Valid verifier result; original command-timeout exception retained
code-context-retrieval	0.1.0	0.171469	Valid line-F1 result after agent development timeout
decision-model-training	0.1.0	0.561905	2,360/4,200; deterministic offline verifier
filtered-ann-retrieval	0.11.1	0.20 (v0.11.0 replay)	IPC fix has no fresh model trial; saved program on 25 cases; old v0.10.1 full reward was 1
join-aware-table-retrieval	0.6.0	0.10	1/10 complete table-and-join sets
legal-stance-retrieval	0.1.2	0	Submission build PermissionError; retrieval quality was not measured
lightweight-query-retrieval	0.4.0	0	Raw nDCG 0.430310 below 0.45; live audit passed
long-pdf-evidence	0.4.3	0.083333	1/12; evidence evaluation and trajectory audit succeeded
methodology-inspiration-retrieval	0.1.1	0.015669	Saved Recall@3 = 0.320765, recalculated under the current baseline
negative-constraint-retrieval	0.3.0	0.10	Verifier rerun of saved submission and original trajectory after Docker failure
paper-metadata-retrieval	0.1.1	0	Valid 0/8; live audit passed
quantity-aware-retrieval	0.1.1	0.161630	Saved nDCG = 0.315722, recalculated under the current baseline

Existing task packages supply the structural precedent; this revision introduces
no new scaffold. Biomedical supplies the sequential CPU retrieval/audit pattern;
filtered ANN supplies the offline-submission/root-side-audit boundary used by
lightweight. Their task-specific data, metrics, hardware and deadlines remain
independent. Detailed precedent records are retained in the task documents.

Remaining review and validation limits

• Biomedical v0.2.2 and filtered ANN v0.11.1 need fresh complete model trials if
  final-version difficulty measurements are required. ANN's reference maximum
  of 14.985 ms is close to the 15 ms limit; host-timing stability remains open.
• Chunqiu's temporal-training route remains unvalidated. Paper still lacks a
  positive legal retrieval witness under the final constraints. Lightweight's
  final configured reference has not demonstrated positive reward; live
  adversarial audit calibration is separate from the passing gate fixtures.
• Low or failed submissions are preserved as observed; no labels or thresholds
  were changed during this publication to turn them into successful trials.
• Previously disclosed source-rights questions for negative, quantity and
  decision, and independent review of paper's additional authorization, remain
  recorded. This upload adds only authored biomedical public examples and
  attributed CC BY 4.0 filtered-ANN derivatives; it does not reupload those
  unresolved corpora or claim that publication resolves their terms.
• No new dependency-install Docker builds, full corpus/index rebuilds or paid
  model trials were run for this publication. Maintainer review of these
  evidence limits remains required before merging.

Contributor checklist

• Current packages are confined to the 13 named tasks/<task-name> paths.
• Actual task authors and original Add filtered ANN and methodology inspiration retrieval tasks #18/Add lightweight query and paper metadata retrieval tasks #23 commits are preserved.
• Allow edits from maintainers.
• No secrets, downloaded inputs/models, job outputs or author solutions in Git.
• Static checks and applicable runtime evidence recorded.
• Four newly published public inputs have provenance, hashes and merged pins.

Maintainer merge checklist

• Review task behavior, timing stability and remaining validation exceptions.
• PRs Add filtered ANN and methodology inspiration retrieval tasks #18 and Add lightweight query and paper metadata retrieval tasks #23 are closed as superseded; their commits remain included.
• Review the separately disclosed source-rights and label-quality limits.
• Confirm head-specific CI, source pins and asset verification.
• Merge with a merge commit, preserving original contributor history.