Artifact Classification — Marko Živković

L ↓ U →	Low Uncertainty TemporalNegligible or within tight, verified SLA. Lag characterised and below threshold for use case. ProbabilisticDeterministic or lossless (η ≈ 1). No information loss through computation. Cross-streamSingle stream, or strongly consistent multi-stream with causal ordering guaranteed.	Medium Uncertainty TemporalBounded and characterised. Lag distribution known, published, and within operational SLA. ProbabilisticWell-calibrated. Confidence intervals published. η documented per processing step. Cross-streamEventual consistency with known, bounded lag. Ordering guarantees documented.	High Uncertainty TemporalUnbounded or uncharacterised. Staleness unknown at consumption time. ProbabilisticSignificant loss. Model uncertainty uncalibrated or unpublished. η unknown. Cross-streamNo ordering or consistency guarantees. Causally impossible views possible.
L0 Atomic The occurrence itself, recorded at the moment it happened. No transformation. Regress terminates here — by nature (decisions) or by designation (observations). ▸ subtypes	Ideal Human decision recorded at moment of commitment with full evidence snapshot Logical inference: state machine transition triggered by verified preconditions (L1 — cites L0 inputs) +3 more… ABCD ◆ 2 LLM variants	Acceptable Calibrated instrument reading with documented error bounds and drift characteristics Human observation recorded with observer expertise, conditions, and confidence metadata +3 more… A*BCD ◆ 3 LLM variants	Degraded Uncalibrated or drifting sensor with uncharacterised error Human observation under fatigue, cognitive load, or time pressure — not documented +3 more… ABCD ◆ 3 LLM variants
L1 Characterised Aggregation Single-stream aggregation with transparent, bounded information loss. Causal chain to L0 intact and explicitly referenced. Single authoritative writer. Includes deterministic logical inference from L0 premises when the generalized conclusion cites its atoms. ▸ subtypes	Authoritative Running aggregate maintained by a single writer, updated synchronously on each L0 event State machine current state — deterministic function of all prior transitions +3 more… A*BCD ◆ 2 LLM variants	Standard Aggregate updated asynchronously with characterised lag and published update frequency Approved artifact where evaluation-to-deployment gap is bounded and documented +2 more… ABCD ◆ 3 LLM variants	Avoid Aggregate from a pipeline with undocumented update frequency or failure modes Summary statistic over a feed of unknown or variable freshness +2 more… ABCD ◆ 2 LLM variants
L2 Purposeful Projection Multi-step or multi-stream transformation optimised for a specific use case. Information loss significant, use-case-specific. Rebuildable from L0.	Usable Feature set computed in a single batch from a complete, pinned, immutable dataset Model prediction with calibrated confidence interval, consumed within verified SLA +2 more… ABCD ◆ 2 LLM variants	Operational Probabilistic model output with published confidence interval, consumed within SLA Point-in-time correct feature store values with documented lag and version +2 more… ABCD ◆ 4 LLM variants	Exploratory only Model output from an uncalibrated model with no confidence estimate Projection computed from inputs of unknown or mixed provenance +2 more… ABCD* ◆ 4 LLM variants
L3 Human-Optimised Rendering Maximum selection, aggregation, and presentation choices made for human perception and cognition. Not suitable for automated decisions at any uncertainty level.	Reporting Regulatory filing computed from a frozen, audited ledger at period-end Report generated from a complete, pinned, verified dataset with full provenance +2 more… ABCD ◆ 2 LLM variants	Standard UI Dashboard refreshed on a defined schedule from characterised L1 or L2 sources User-facing status display with staleness indicator and last-updated timestamp +2 more… ABCD ◆ 3 LLM variants	Prohibited Dashboard mixing data of different vintages without disclosure or staleness indicators Summary presenting cross-stream aggregates as if simultaneously consistent +2 more… ABCD ◆ 5 LLM variants

CONSUMER CLASSES · * = conditional with explicit requirements

AInvariant enforcer — irreversible decisions, must not violate invariants

BAudit / reconciliation — must trace causality to L0

COperational — consequential but correctable decisions

DHuman consumption — display, exploration, reporting

Invariant rules

Class A consumers

L0 or L1 only · Low or Medium-U with reservation pattern · Never L2 or L3 · Never High-U

Class B consumers

L0 preferred · L1 with full causal chain to L0 · Never L2 or L3

Atomicity and uncertainty are independent

L0 + High-U (degraded sensor) is worse than L1 + Low-U (real-time aggregate)

L3 + High-U must not exist

Rendering that conceals uncharacterised uncertainty is an architectural violation, not a design choice

LLM output is L2 or L3 by default

Naked LLM: L2/L3-U-high. RAG with verified citations: L1-U-med. Deterministic extraction with verification: approaches L0-U-med. LLM output stored as L0 without human review contaminates ground truth permanently.

Uncertainty thresholds are contractual

What counts as Low, Medium, High must be defined, measured, and published per artifact per domain

Cross-stream consistency is a distinct uncertainty component

A view can be temporally fresh and probabilistically accurate but causally inconsistent across streams