How does Kamin move from evidence to judgment?
This page describes the behavior of the current public release. Validation targets are not presented as measured results, and Kamin does not display a numeric Fit percentage until documented research calibration exists.
Digital-interest analysis
A local experimental keyword method covers six technical and career topics in Arabic and English. It is not a personality model or a Cambridge Analytica algorithm. Mention may reflect criticism or coursework, so suggestions enter the graph only after student confirmation and pathway explanations require separate permission. Rule R4 joins a confirmed interest to a pathway topic without changing fit, eligibility or capability gaps. No uncalibrated accuracy or confidence percentages are shown. Topic, matched keywords, method version, time and consent are retained; original text is not.
Semantic inference in this release
Person360 contains sourced relationships. Bounded JavaScript rules join capability evidence with a matching requirement, a declared goal with a supporting pathway, a preference with compatible pathway context, or a student-confirmed digital interest with a topic-related pathway after separate permission. The UI exposes premises, the rule and the inferred connection. RDF is exported through JSON-LD with separate named graphs for profile inputs, reference knowledge and conclusions.
This is a bounded inference core inspired by the research pattern. SELECT and ASK SPARQL 1.1 queries run locally in the browser with Oxigraph. Reference knowledge and synthetic examples are separate from personal data; querying a profile requires independent permission for the current session. Updates and remote queries are blocked. General OWL reasoning and a SHACL engine are not implemented. Person-to-person matching is not implemented. Preferences never become capability evidence, and missing evidence is not an absent skill. Ontology and shape files are development assets, not proof that their processors are running.
Try the inference example1. Pipeline
Extract, then review
The record is processed locally. The user corrects extracted courses and gives explicit approval before they can enter inference.
Governed mapping
A course title alone is insufficient. Only an explicit course → learning outcome → capability mapping can create capability evidence.
Explain from structured data
Judgments expose reasons, limits, and gaps from evidence plus target requirements; they are not based on free-form LLM text alone.
2. Evidence levels
| Level | Meaning | Current status |
|---|---|---|
| Self-declared | A preference or goal entered by the user | Personalizes a decision but is not academic evidence |
| Document-backed | Information reviewed by the user from a document | Available in the public release |
| Governed mapping | A capability created through an approved educational mapping | Available for selected courses only |
| Institution-verified | Direct verification from a university or issuer | Not operational yet |
| Observed/tested | Documented assessment or observed performance | Roadmap |
3. What enters Fit?
Included
Structured target requirements, evidence-backed capabilities, formal constraints, relevant declared preferences, and explicitly visible evidence gaps.
Not treated as fact
Guesses about personality, sensitive traits, uncalibrated percentages, or capabilities inferred without a traceable source.
4. Pilot validation targets
These are Targets, not Results. Numbers should be published only after documented testing against human review.
| Field extraction accuracy | Target ≥ 95% |
|---|---|
| Skill precision | Target ≥ 90% |
| Evidence linkage rate | Target 100% |
| Unsupported public skills | Target 0% |
| Time to first value | Target ≤ 10 minutes |
5. Psychometrics: research track, not production scoring
O*NET RIASEC and IPIP Big Five remain research candidates only. Self-declared preferences are separate from psychometric scores. No IPIP/RIASEC score enters Fit or eligibility before documented Saudi validation covering reliability, factor structure, cultural appropriateness, and measurement invariance across relevant groups. Saudi norms or percentiles must come from documented local validation rather than direct conversion of external norms.
6. National positioning
Kamin is not a parallel national platform. Its role is a learner-owned university evidence and trust layer, designed standards-first and ready for future interoperability if an official API path becomes available. Kamin currently makes no claim of government integration or endorsement.
7. Current limits
- No published accuracy results yet.
- No automated hiring decision.
- Job and training matching is a pilot aid, not a replacement for a recruiter.
- Central institutional storage and employer sharing are not enabled in the public release.
- Psychometric scores do not enter Fit before Saudi validation; only low-risk self-declared preference signals are currently used.
Local semantic recall and its limits
We use a pinned q8 ONNX paraphrase-multilingual-MiniLM-L12-v2 model, normalised mean pooling and a bilingual index built from the same reference catalog. Up to three candidates pass an operational similarity threshold and stay within a margin of the best result; it is not a readiness percentage or calibrated accuracy. The phrase stays in screen memory and is not added to Person 360.
The catalog is limited and experimental. Arabic/English retrieval needs expert and user evaluation before career-effectiveness claims. Recall creates no graph fact and does not change rule-based judgments.
AI-native recommendation governance
We apply the seven questions to Kamin’s current experience: local rules, inspectable semantic recall and human review. These are product governance mechanisms; we do not claim autonomous agents or institutional verification.
Decompose the work
Extraction, evidence mapping, pathway suggestions and human review are separate stages. Review one recommendation without approving the others.
Define acceptance criteria
The review shows the recommendation’s requirements, present and missing evidence, mapping source and rule version. Acceptance or rejection requires checking evidence, the goal and the judgment’s limits; unresolved remains an available outcome.
Isolate the stages
Reviews are stored separately from the knowledge graph. Accepting or contesting a recommendation adds no capability and changes neither ranking nor evidence level. Contests are not automatically sent.
Reversal — local support with explicit limits
Withdraw a decision while retaining its history, or delete reviews in Privacy. Changed inputs or rules, and restored encrypted backups, require fresh review. Links and files already delivered cannot be recalled from recipients.
Inspectable evidence
Premises, sources, gaps and the rule version remain alongside the judgment. A review is tied to the input and recommendation version; it is not an institutional signature or skill certificate.
Red team — local feedback with human review
“Contest this judgment” records a reason locally. Export a test candidate without names, free-text notes or profile data. The team turns a confirmed defect into a reproducible synthetic case and a regression check. A contest alone is not a confirmed defect; no central triage or automatic closure service is active.
Human accountability
The reviewer chooses exploratory acceptance, rejection, contest or unresolved, and records a name or alias and role. The role is locally declared, not authenticated. Final learning or employment decisions remain with the authorized people and institutions.
How will we measure decision time and errors escaping verification?
The measured local assistant
One narrowly defined function: from a project description, suggest capabilities with the supporting sentence for each suggestion, or abstain. The method is bilingual local lexicon matching; no external model and no data leaves the browser. It is measured on a frozen labelled set of 24 cases (12 Arabic, 12 English) with pre-registered thresholds: precision ≥ 0.8, recall ≥ 0.75 and a non-zero abstention rate per language. Current engineering result: precision 1.0 and recall 1.0 in both languages. These are engineering results on a small set, not psychological or professional validity, and they change no evidence level or eligibility.