RUBRIC.md

# SkillJev Competency Rubric v0.1

## Core principle: UNKNOWN ≠ LEVEL 0
- **Unknown** = the provided evidence does not let us estimate a level (skill absent, only listed, or ambiguous). This is the DEFAULT when evidence is missing. Radar shows a gap/hatched axis, never 0.
- **Level 0** = there is *explicit* evidence of no experience (e.g. "no production experience with Kubernetes", "currently learning Go, never used it at work", interview confirmed none).
- A skill appearing only in a "Skills:" keyword list ⇒ evidence `insufficient` (or at most `partial`), proficiency **Unknown** unless context gives level signals.
- Self-reported (CV) and verified (interview/assessment) are separate layers: `source_type` in schema. CV-only evidence caps confidence; it never caps level, but levels ≥4 from CV need concrete outcome/scope evidence.

## Generic levels (apply per skill, then skill-specific anchors refine)
| Level | Name | Evidence signals |
|---|---|---|
| 0 | None (explicit) | Candidate/interviewer explicitly states no experience |
| 1 | Awareness | Coursework, tutorial, toy/personal project, "familiar with" |
| 2 | Assisted | Used at work on tasks under guidance; small features; < ~1 yr hands-on |
| 3 | Independent | Owned features/services in production end-to-end; debugged real issues; ~2+ yrs |
| 4 | Advanced | Designed systems/subsystems with this skill; performance/scale outcomes; set team practices; mentored |
| 5 | Expert / Authority | Org-wide standards, architecture across teams, OSS maintainer/contributor of core project, talks/papers, solved novel hard problems at large scale |
| unknown | Unknown | Not enough evidence to pick a level |

Years are weak signals only; scope, ownership, and outcomes dominate.

## Evidence classes (evidence head)
| Label | Definition |
|---|---|
| `supported` | Evidence directly shows the requirement is met (skill used at ≥ required level / required context) |
| `partial` | Evidence shows related or lower-level use, adjacent tech, or only part of the requirement |
| `contradicted` | Evidence conflicts with the requirement (explicit no-experience, timeline conflict, clearly below in a way stated) |
| `insufficient` | No evidence, or keyword-only mention, or too vague to judge |

Mapping guide: `insufficient` → proficiency normally `unknown`. `contradicted` → often level 0 or a stated low level. `partial`/`supported` → a numeric level.

## Skill-specific anchors (examples)
**kubernetes (cloud_devops)**: L1 ran minikube tutorial · L2 deployed apps with existing Helm charts · L3 wrote/maintained manifests/Helm, debugged pods in prod · L4 designed cluster setup, autoscaling, multi-env rollout, operators · L5 platform owner for many clusters/teams, upstream contributions.
**system_design (system_design)**: L1 describes patterns · L2 contributed components to an existing design · L3 designed a single service incl. data model & failure modes · L4 designed multi-service systems with explicit trade-offs at scale (throughput/latency numbers) · L5 org-level architecture, led migrations across teams.
**postgresql (data_storage)**: L1 basic CRUD · L2 schema changes, joins, ORM usage · L3 indexing, query plans, migrations in prod · L4 partitioning, replication, tuning at high load · L5 deep internals, extension dev, large-scale DBA leadership.

## Example axes — Senior Backend Engineer (JD target)
| Axis (group) | Target level |
|---|---|
| lang_core (Go/Java) | 4 |
| api_design | 4 |
| data_storage | 4 |
| system_design | 4 |
| cloud_devops | 3 |
| reliability_obs | 3 |
| security | 3 |
| testing_quality | 3 |
| leadership | 3 |

## Confidence (calibrated)
- Model outputs softmax per head; calibrated with temperature scaling on the dev split (ECE target ≤ 0.05).
- `review_flag = true` if max prob < 0.7, or evidence=`contradicted`, or level ≥4 from CV-only.

## Real-CV labeling convention v0.2 (2026-10-16, adopted as standard; derived from the locked held-out reviewer's conventions — no held-out text copied into train)
Applies to single CV excerpts (Djinni). Requirement target is 3–4.
| Pattern in excerpt | evidence | proficiency |
|---|---|---|
| Keyword/stack list only ("Languages: …", "Technologies: …") | insufficient | unknown |
| Skill absent / wrong sense (e.g. "go-to-market", "3 node cluster") | insufficient | unknown |
| Only *wants/plans/interested in* learning the skill | insufficient | unknown |
| Courses, diploma, certificate, tutorial, "basics of", homework | partial | 1 |
| *Practising/learning while doing*: mentorship/teaching program, "learning X while working on project", "know basics + have experience working with X" among many tools | partial | 2 |
| Single project mention with the skill but no ownership/duration ("Worked on a project (C# Unity)") | partial | 2 |
| Skill named among several in a generic usage claim ("projects using C++, C, Python, C#") | partial | unknown |
| < 2 years, junior, intern, small mentoring (1–3 people) | partial | 2 |
| ≥ 2 years commercial with the skill, or owned/delivered features in production; a single metric improvement (e.g. "reduced deploy time 50%") without design/lead scope | supported | 3 |
| Designed/architected/led systems or teams with this skill at scale (multi-team, large migration, platform) | supported | 4 |
| Org-wide authority / core OSS maintainer | supported | 5 |