تصنيف tasnif
Tasnif is Arabic for classification. This site is the working home of two design-science artifacts that share one conviction: a classifier operating under regulation should cite the provision behind every decision, decline to answer rather than guess, and route the uncertain and the high-stakes to a human — and its evaluation should count silent errors, not just accuracy.
NDMO-inspired, SAMA-aware, human-gated classifier for Saudi banking metadata
A deterministic engine that derives a candidate security level from a four-dimension impact profile, cites policy on every decision, keeps PDPL obligations on a separate axis, abstains on what it does not recognise, and gates the rest. Evaluated with classification quality and workflow safety kept apart.
- 100% agreement where a rule fires on external metadata — at 23.3% coverage
- 0 silent under-classifications at every gate threshold tested
- Credit Data flag: 0 / 18 — reported, not repaired
TierGuard: a deterministic triage-and-abstention baseline for Article 5, Annex III and Article 50 of the EU AI Act
From a free-text system description: a prohibited-practice or high-risk candidate bound to its Article or Annex, a minimal outcome with concurrent Article 50 flags, or a first-class INDETERMINATE record carrying candidate provisions and missing facts. Evaluated with error-direction metrics, compliance-washing and a multilingual probe.
- 58% correct-or-referred on the post-freeze set, with 40% silent under-classification, stated plainly
- 91% on the Commission's draft-guideline examples
- Pinned to Regulation (EU) 2024/1689 as amended by the Digital Omnibus
What the two artifacts share
Cite every decision
Each outcome carries the clause it rests on — an NDMO impact rule, an Article 5 point, an Annex III use-case — so "why?" has an answer a regulator can check.
Abstain, don't guess
An unrecognised field gets an interim default; an ambiguous system gets an INDETERMINATE record. Uncertainty is a first-class outcome, never a label.
Gate the consequential
High-stakes candidates and every abstention go to a named human. The gate's workload is reported as a cost, not hidden as coverage.
Count the silent errors
The metric that matters is the wrong label nobody was asked to review. Both evaluations report it first, with confidence intervals, and separate it from the safe direction.
Sovereign by design. Every tool on this site is a browser port of the engine its paper evaluated, loaded from the same frozen rule files and replayed against that engine on every benchmark row before it ships. Nothing you type is sent anywhere.