[hr-jin]

Summary

Task names and package paths: the 13 existing tasks/<name> packages listed below.
Goal, starting system, and CPU/GPU requirements: update final-score normalization; raw quality metrics, starting resources, hardware, submission interfaces and execution limits are unchanged.
Authors and coauthors: Haoran Jin; existing authors and contributor history are preserved.

khmer-embedding-training, financial-report-qa, evolving-personal-memory, document-page-search, layout-page-search, visual-document-qa, composed-image-search, sketch-text-search, same-place-image-search, video-moment-search, narrative-clue-search, spoken-query-search, icsi-meeting-memory.

Each task now uses two linear segments with a fixed reference quality and an integer score anchor. Raw quality 0, the reference quality, and 1 map to normalized scores 0, the declared anchor, and 100. Harbor retains its 0–1 reward precision; displayed scores use one decimal place.

Raw metrics remain available separately, and executable, semantic-judge and trajectory-audit gates retain their existing behavior. Khmer uses raw Recall@100 instead of normalizing an already-normalized value. README scoring descriptions and task versions are updated. Twelve instruction.md files are byte-identical; only ICSI's obsolete scoring sentence is aligned with the new conversion.

Validation evidence

• python scripts/check_release.py: passed.
• python scripts/check_tasks.py --merge-ready --base 04bdc08548473d3661c67724ff514db7ab6717aa: passed.
• git diff --check: passed.
• Task-local unittest discover in cached verifier images with networking disabled: 512 tests passed, no skips.
• Independent production-finalizer fixtures: 296 cases passed, including the actual isolated-Python ICSI entry, valid-zero submission failures, raw metric preservation, and audit/semantic failure gates.
• Shared contract suite: 174 tests completed (172 passed, two host-dependency skips); all five RewardKit tests additionally passed without skips in the cached offline verifier image. Pi trajectory suite: 21 tests passed.
• Three independent reviews of the final source passed.
• Coding-model trials, GPU inference and paid judge calls were not repeated. Validation reuses existing raw measurements and synthetic finalizer inputs; this PR changes their score conversion, not their raw metric calculation or task inputs.

Contributor checklist

• Existing descriptive task names and tasks/<task-name> paths are preserved; no new tasks are introduced.
• Actual authors in task.toml; original author commits preserved.
• Allow edits from maintainers.
• No secrets, downloaded inputs/models, job outputs, trajectories or author-only solutions.
• Static checks and applicable offline runtime evidence recorded.
• Asset manifests and provenance files are unchanged; this PR uploads no data.

Maintainer merge checklist

• Design, verifier isolation, runtime evidence and any validation exceptions reviewed.
• Names checked against the current base and queued PRs.
• Upstream or official asset downloads verified against immutable pins, sizes and SHA-256 hashes.
• If official publication is needed: new assets published under tasks/<task-name>/, previous assets preserved and merged official SHA pinned.
• Task-local docs/configuration complete; no shared task inventory updates needed.
• Required judges, credential isolation and failure cases validated; checks and merge-ready pass.
• Merge commit; original contributor history retained.