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