ALIGNS.md can help a team turn recurring corrections and agreed decisions into examples for a coding agent. Option 8 supplies local candidate collection, review and local dataset export. The workflow below adds a manual handoff and team review. It is an operating guide, not a hosted submission service or a claim that a trained model will fit the team without evaluation.

For technical review or an adoption decision, share the MLOps and team decision brief, also available in Korean. It includes the implemented checks, reproducible evidence and remaining validation work. The initial readiness audit is retained as the record of the gaps that prompted these changes.

Choose one repository or workstream and a few decisions worth learning, such as when to ask for context, how much complexity to introduce, and what a completion report must verify. Record the team's agreed criteria and their version. Keep individual preferences separately scoped; a maintainer's personal preference or a frequently repeated comment is not automatically a team rule.

Before collecting, agree on:

• Participating contributors and the sources each person permits.
• A private submission destination owned by the team, with an authorized reviewer.
• Who can approve a case as representing the team's criteria and who resolves disagreements. Leave unresolved cases out instead of using majority frequency as a preference label.
• A manual cadence, for example once per sprint, with no required quota.
• Retention, withdrawal, and permission to use the selected code or conversation excerpts for the intended training. Permission to store a record is separate from permission to send it to a model provider.

Do not create a schedule or start collecting teammates' conversations from this guide alone. Each person chooses their own sources through option 8. No private personal ALIGNS.md file needs to be submitted.

At the agreed checkpoint, the contributor asks their agent to review new candidates and prepares a batch for the chosen team destination. They check the exact files before manually submitting them. An empty batch is fine; repeated examples do not make a better dataset.

Use a directory such as submissions/<batch-id>/ in the team's selected private storage. It contains:

In submission.json, each files entry records a relative file path and its SHA-256 hash. Keep hashes current after redaction. In review.json, copy decision_fields into decisions for each record and fill it with the actual review result; the empty template is not approval.

Keep credentials, customer identifiers, unnecessary personal details and full raw logs out. Include code or patches only when permitted and needed to understand the decision. Preserve actual test results and explicit verification gaps. These files can still contain private team information after review.

Submission is manual. A contributor can copy the batch to approved private storage or create a PR in a team-selected private data repository. Creating a batch does not authorize a commit, PR, push, upload, or changing repository visibility. The ALIGNS.md public repository is not the intake destination.

The reviewer checks each record's provenance, source-use permission, applicable criterion version and target answer. Record approved, needs_changes or excluded, with the reason and approval date. Approval of a PR or collection settings does not stand in for approval of the training target.

Keep the original submission intact. The version-2 review binds its exact bytes with submission_sha256 and each original record with record_sha256. Provide an approval reference and an authorized reviewer alias. The team policy lists allowed reviewers, current criteria snapshots, permitted use, withdrawals and shared task-family splits. Use the policy template. Never recompute approval hashes to hide edits made after review.

The validated exporter derives namespaced record IDs and applies the reviewer's shared task family without rewriting originals. Use one family for the same issue, PR or derived scenario across contributors. A reviewer must explicitly establish the team scope; personal records are not relabeled automatically.

For DPO, the reviewer must approve the comparison as well as the preferred answer. Both alternatives need the same available facts and instructions. New requirements added during a correction make it a different context, not an observed preference pair. Keep hypothetical and reconstructed examples labeled and opt them in separately.

Build an export plan over the approved team scope. Assign whole task families to training or evaluation before making variants. Review semantic duplicates across contributors; the exporter detects exact overlap but cannot identify every related task. The option-8 procedure describes the exporter and the supported record shape.

Use the version-2 contract and commands for team submissions. The bundled export_team_dataset.py checks batch and artifact hashes, reviewer declarations against the selected policy, approved record hashes, required context, criterion versions, withdrawals and split assignments. It rejects contradictory preference cycles and stops before writing on invalid input. It does not authenticate reviewers or independently execute the submitted tests.

Prefer sft-completion.jsonl, which separates input from the selected target. The export also includes text DPO, held-out evaluation and a private provenance manifest. Selected tool actions retain schemas/calls/results; code examples need recorded passing checks and hashed evidence. The legacy text exporter remains for personal records and must not substitute for team approval validation.

The formats follow TRL dataset formats. Run the bundled local-tokenizer checker, then check the actual training stack's labels, tool argument schemas and context limits. Tool-action compatibility is model-specific, and full-run trajectory learning is not implemented.

The team chooses the training model, provider and training run. Compare the same base coding model with the same ALIGNS.md and tools before and after fine-tuning. Use held-out tasks to check corrections required, unnecessary clarification, criterion fit and code/test correctness. Keep independent team review where possible. Lower correction counts with worse code are not a successful result. Data volume alone does not show that a model is ready to train or safe to replace the baseline.

Version each exported dataset and keep a private index of included records and training runs. If a record is withdrawn or a criterion changes, exclude affected records from later exports and identify older datasets and runs that used them. Deleting a record does not erase Git history, remote copies or learned weights.

Available now: collection guidance, version-2 templates, validated manual batch exports, selected-target SFT, text DPO, code evidence checks, selected tool-action export, output hashes and a local-tokenizer boundary checker.

Before a real pilot, choose contributors, reviewers, scope, private storage, cadence and a target model. Reviewer authentication, semantic duplicate/conflict review, isolated code execution, trainer integration, model training and measured team outcomes remain the team's work. No shared data store is created by installing the skill.

Suggested request at the manual checkpoint:

Review my newly collected option-8 candidates for this team's approved scope. Prepare a batch with only the examples I have cleared for team training, preserving provenance and verification limits. Show me the selected files and exclusions. Leave submission to the team destination to me.