SCHEMA.md

# SkillJev Skill Assessment Schema v0.1 (JSON)

Three objects: `JDRequirement`, `CandidateEvidence`, `Assessment`. All carry `schema_version: "0.1"`, `taxonomy_version: "0.1"`.

## JDRequirement
```json
{
  "requirement_id": "jd-0001-r03",
  "jd_id": "jd-0001",
  "role_title": "Senior Backend Engineer",
  "text": "5+ years building high-throughput services in Go or Java",
  "skill_id": "go",
  "group_id": "lang_core",
  "target_level": 4,
  "importance": "must|nice",
  "source_span": {"start": 412, "end": 470}
}
```
JSON Schema (abridged):
```json
{"type":"object","required":["requirement_id","text","group_id","target_level","importance"],
 "properties":{"group_id":{"enum":["lang_core","api_design","data_storage","system_design","cloud_devops","reliability_obs","security","testing_quality","frontend","leadership","other"]},
  "target_level":{"type":"integer","minimum":0,"maximum":5},
  "importance":{"enum":["must","nice"]}}}
```

## CandidateEvidence
```json
{
  "evidence_id": "cand-0042-e07",
  "candidate_id": "cand-0042",
  "source_type": "cv|portfolio|interview|assessment",
  "section": "experience|projects|skills_list|summary|education|interview_note",
  "text": "Led migration of 30 services to Kubernetes (EKS), wrote Helm charts and HPA policies",
  "source_span": {"doc_id": "cv-0042", "start": 1203, "end": 1290},
  "date_range": {"start": "2022-03", "end": "2024-06"}
}
```
`source_type` separates **self-reported** (cv, portfolio) from **verified** (interview, assessment).

## Assessment (model + human output, one per requirement×candidate)
```json
{
  "requirement_id": "jd-0001-r05",
  "candidate_id": "cand-0042",
  "group_id": "cloud_devops",
  "evidence_label": "supported|partial|contradicted|insufficient",
  "proficiency": 0,
  "proficiency_label": "0|1|2|3|4|5|unknown",
  "confidence": {"skill": 0.97, "evidence": 0.91, "proficiency": 0.74},
  "review_flag": false,
  "evidence_ids": ["cand-0042-e07"],
  "citations": [{"evidence_id": "cand-0042-e07", "quote": "wrote Helm charts and HPA policies"}],
  "verification": "self_reported|verified",
  "rationale": "short, references quotes only",
  "model": {"name": "niko0xdev/skilljev", "revision": "..."},
  "human_override": null
}
```
Constraints: `proficiency` is `null` iff `proficiency_label == "unknown"`. If `evidence_ids` is empty ⇒ `evidence_label = insufficient` and `proficiency_label = unknown` (never 0). Citations must be exact substrings of evidence text.

## Radar aggregation
Per axis: JD target = max(target_level of must requirements in group). Candidate = confidence-weighted median of numeric levels; if all requirements in group are `unknown` → axis state `unknown` (not 0). Three layers: `jd_target`, `cv_evidence`, `verified`.

## Training sample (intermediate format) — `data/samples_*.jsonl`
One line = one requirement–evidence pair with all head labels:
```json
{"id":"sj-v001-0001","requirement":{...JDRequirement subset...},"evidence":[{...}],
 "labels":{"group_id":"cloud_devops","evidence_label":"supported","proficiency_label":"4"},
 "meta":{"difficulty":"easy|hard","case_type":"...","source":"handwritten_v001"}}
```
Exported to julia/OpenJev trainer rows `{state, question, options, type, target, meta}` — 3 rows per sample (one per head) — by `scripts/export_trainer.py` → `data/trainer_v001.jsonl`. Confidence is not a trained label; it comes from calibrated softmax.