docs/labeling.md

# Decisioncraft label policy v1

Policy version: v1. Review status belongs in the separate review and experiment records, not in this immutable policy. Neither independent-agent review nor approval of the 12-case policy sample establishes human review of the entire dataset. No model scores were used to write these targets.

## Scope and precedence

Label the requested operation using the request and explicit context together. Context supplies facts, preferences and referents; it is not lower priority merely because a filename appears outside the request sentence. An explicit priority or correction can resolve an apparent conflict. If the two sources still conflict without a stated priority and the difference affects the command, clarify unless the operation is already outside scope. Do not invent missing targets or choose among incompatible alternatives.

1. Choose `unsupported` when any requested positive action is outside the domain below. This includes a mixture of a supported local step and an unsupported external step. An explicitly prohibited or merely quoted operation is not a requested action.
2. Otherwise choose `clarify` when a missing or contradictory fact materially changes the command's behavior. Necessary information depends on the operation: a destination is required for a specifically requested copy, but a file listing does not need an invented output filename.
3. Otherwise choose `suggest`. This means sufficient information for a suggestion, not that the resulting command is correct, harmless, authorized, or executable on an actual machine.

The domain is command suggestions for GNU file/text utilities and local Git in a Linux workspace. It includes reading, searching, transforming and comparing local text; enumerating file paths and metadata; copying, moving, creating and removing precisely described local files; local permissions and symbolic/hard links; archive/compression/checksum operations; and inspecting or changing a specified local Git repository. Short compositions of these supported operations remain supported when all needed details are specified.

The domain excludes network access, remote Git fetch/push/clone, cloud services, package installation, Docker/container control, running applications or test suites, host service or network administration, databases, email, and arbitrary multi-system workflows. Downloaded content may be processed locally after it exists; requesting its download is outside scope. Filenames such as `docker.log` or text containing `npm install` do not change an otherwise local text operation into a package/container task. This scope is the router's declared domain, not a promise that the released Bashcraft generator implements every operation reliably.

## Ordinary defaults and material ambiguity

Use ordinary command conventions only when the request does not contradict them. The current workspace can supply the working directory when context identifies it. A read-only text result may print to standard output without an explicit output path. A normal file listing concerns the named directory or an explicitly supplied current directory. Do not require an output destination for every read operation.

An unqualified sort defaults to ascending lexicographic order unless explicit context requires something else. Require clarification if the request asks for a custom ordering but omits its key or direction, or context supplies conflicting numeric/text preferences. The dataset must not manufacture ambiguity simply because an optional command flag is absent. Similarly, a typo with one clear reading does not by itself require clarification. A vague pronoun is complete when context supplies one referent and ambiguous when two equally plausible referents remain.

Destructive operations are not automatically `unsupported` or `clarify`: deleting a precisely named temporary file is semantically specified. Conversely, a seemingly harmless display request can be ambiguous when it omits the selected file or comparison baseline. Safety assessment and permission to execute are separate systems; this dataset does not provide either.

Negation is compositional. “Inspect the local log; do not restart the service” requests only inspection. “Do not inspect the log; restart the service” requests an unsupported service action. Conditional alternatives with an explicit resolvable condition may be complete; unresolved conflicting instructions are not.

M1 used tiny mechanism-development labels, including a deliberately underspecified sort example. Those historical labels are not canonical M2 policy: ordinary sort order and standard-output conventions are explicitly resolved above. M1 cases are excluded rather than silently relabeled or reused.

## Scenario grouping and benchmark boundaries

Every authored scenario defines a semantic task and its material slot or constraint. All context changes, minimal pairs, paraphrases, typos, negations and entity substitutions derived from it share `group_id = source_id + ':' + scenario_id`. Split whole groups after authoring. A name change or sentence wrapper never creates a new independent scenario.

The source specifications are synthetic, authored by the `decision_source` coding agent. Generated rows preserve source/scenario/variant identity and rationale. The deterministic builder does not obtain predictions, import another dataset, or read Bashcraft or M1 cases. Related views improve coverage of controlled input conditions; their row count is not the number of independent observations. Report both counts and use groups for uncertainty estimates.

Whole designated families form a separate generalization cohort. They are excluded from training, model selection and calibration. Other test groups represent familiar families with new scenarios. Publish the family/cohort definitions and each split's class/group counts before model selection. The final labels must not inform prompts, thresholds or training choices after freezing.

## Schema and review provenance

Each row has `id`, `group_id`, `family`, `request`, `context`, `label`, `rationale`, `tags`, `provenance` and `review`. The provenance includes source ID, scenario ID, variant ID, synthetic author, method and license. Review initially records `pending`, no reviewer, and author checks only. An independent reviewer records disagreements and corrections in a separate bound ledger. Independent-agent review is not human review.

The human policy sample below is outside the dataset and all four splits. It tests scope and conventions before a costly study. A user approving these examples validates that policy sample only. It cannot establish that all generated rows received human review. If full human review is unavailable, explicitly label any release synthetic and independently agent-reviewed, disclose the evidence limit, and obtain an explicit scope decision rather than marking human review complete.

## Twelve policy cases for user review

These original examples do not become training or final-test cases. The proposed labels and reasons are supplied so the user can correct concrete policy decisions.

| ID | Request and context | Proposed label | Reason |
| --- | --- | --- | --- |
| P01 | “List the entries here.” Context: current directory is `/workspace/manuals`. | `suggest` | The context supplies the directory; ordinary display needs no output file. |
| P02 | “Sort these words alphabetically.” Context: `terms.txt` in the current workspace, output to terminal. | `suggest` | Ascending alphabetical order has an ordinary interpretation. |
| P03 | “Remove the disposable marker.” Context: exactly `/workspace/build/marker.tmp`; no other files should change. | `suggest` | The target is specified; this label does not authorize deletion. |
| P04 | “Display the lines containing the literal text `docker run` from the local note; do not launch Docker.” Context: the note is `/workspace/help.txt`. | `suggest` | Quoted words and negated container actions do not request container control. |
| P05 | “Compare the chosen drafts.” Context: `draft-a.txt`, `draft-b.txt` and `draft-c.txt` exist; no pair has been chosen. | `clarify` | Two required operands remain unresolved. |
| P06 | “Move the package notes to their new home.” Context: source `/workspace/pkg-notes.txt`; destination absent. | `clarify` | Moving the file requires a destination. |
| P07 | “List the logs older than the agreed age.” Context: `/workspace/logs`; no age agreement is recorded. | `clarify` | The requested age boundary is material, not an optional default. |
| P08 | “Reset this local branch to the approved commit.” Context: two approvals name different commits with no priority. | `clarify` | The reset target is contradictory. |
| P09 | “Pull the branch from origin.” Context: current local repository is identified. | `unsupported` | Fetching remote state requires network Git, outside local-only scope. |
| P10 | “Print the local digest, then send it by email.” Context: the digest file and recipient are specified. | `unsupported` | A specified local action plus email is a mixed-domain request. |
| P11 | “Do not read the package list; install the dependencies.” Context: the project uses npm. | `unsupported` | The positive action is package installation. |
| P12 | “Restart the development web server.” Context: its service name is specified. | `unsupported` | Service control is outside file/text/local-Git scope. |

## Review checklist

- Judge the positive requested actions, not keyword presence.
- Read context before declaring an operand missing.
- Identify the exact material fact supporting `clarify`; optional defaults alone are insufficient.
- Confirm unsupported precedence for mixed scope, including well-specified requests.
- Check that every derived variant stays in the original scenario group.
- Record synthetic authorship, independent-agent review and human review separately.
- Preserve disagreements and changes before dataset freeze; never relabel to improve model scores.