docs/SCHEMA.md

# Schema Guide

## Top-level fields

| Field | Type | Description |
|---|---|---|
| `id` | string | Stable record ID in the form `jlab-v1-00001` |
| `version` | string | Dataset schema/content version |
| `split` | enum | `train`, `validation`, or `test` |
| `record_type` | enum | `trajectory`, `preference`, or `repair` |
| `primary_behavior` | enum | Main behavioral supervision target |
| `secondary_behaviors` | string[] | Cross-cutting behavior labels |
| `language` | enum | Code ecosystem label |
| `ecosystem` | string | Referenced build/test ecosystem |
| `task_type` | enum | Software-engineering task category |
| `difficulty` | enum | Synthetic composition difficulty |
| `task` | object | Instruction, constraints, and acceptance criteria |
| `environment` | object | Synthetic repository and command description |
| `behavioral_labels` | object | Expected policy and authorization labels |
| `provenance` | object | Origin and evidence status |
| `quality` | object | Validation and quality tier |
| `supervision` | object | Record-type-specific target |
| `fingerprint` | string | SHA-256 over canonical record content excluding the fingerprint |

The machine-readable definition is in `schema.json`.

## Task object

```json
{
  "instruction": "Natural-language task",
  "constraints": ["Explicit boundary"],
  "acceptance_criteria": ["Observable outcome"]
}
```

Constraints are part of the task contract. Training systems should not merge them into untrusted repository content.

## Environment object

The environment describes a candidate context:

- `repository_id`
- `repository_shape`
- `domain`
- `synthetic_commit`
- `target_file`
- `test_file`
- `target_symbol`
- `test_symbol`
- `targeted_test_command`
- `full_test_command`
- `lint_command`

All v1 environments are synthetic and unmaterialized.

## Behavioral labels

```json
{
  "must_ground_before_edit": true,
  "must_verify_before_success_claim": true,
  "should_ask_clarification": false,
  "requires_approval": false,
  "expected_tool_sequence": ["search", "read_file", "apply_patch", "run_tests"],
  "target_behavior": "..."
}
```

`requires_approval` is a supervision label, not runtime authorization. A deployed agent must determine authorization from its actual policy and environment.

## Trajectory supervision

A trajectory contains:

- `policy_summary`;
- ordered `steps`;
- `candidate_patch` strategy;
- `verification_plan`;
- `final_response_contract`.

Each step has:

| Field | Meaning |
|---|---|
| `step` | Local sequence index |
| `phase` | Ground, inspect, plan, edit, verify, repair, review, or authorize |
| `observation` | Short observable decision basis |
| `action.tool` | Abstract tool name |
| `action.arguments` | Structured candidate arguments |
| `expected_observation` | Expected information gain or state transition |
| `decision_basis` | Why this action is justified by available evidence |

Expected observations are not executed tool results.

## Preference supervision

A preference record contains:

- `chosen_sequence`;
- `rejected_sequence`;
- `preference_label`;
- `preference_reason`;
- `rejected_error_taxonomy`.

The pairs encode the project’s behavioral specification. They are not human preference votes.

## Repair supervision

A repair record contains:

- `initial_attempt`;
- `failure_signal`;
- `diagnosis`;
- `repair_sequence`;
- `reference_trajectory`;
- `success_criteria`.

The failure is a controlled synthetic signal. It is not a captured runtime event.

## Provenance contract

Every v1 record states:

```json
{
  "origin": "synthetic",
  "generation_method": "deterministic_scenario_composition",
  "repository_materialized": false,
  "execution_verified": false,
  "tool_results": "expected_or_simulated",
  "human_reviewed": false,
  "release_stage": "research_preview"
}
```

Consumers should treat these fields as filtering and weighting inputs.

## Fingerprints

The fingerprint is computed as:

```text
sha256(canonical_json(record_without_fingerprint))
```

Canonical JSON uses sorted keys, UTF-8, and compact separators. Fingerprints detect accidental content changes and exact duplicates; they do not detect semantic near-duplicates.