01 / LANGUAGE
SEMANTIC QUALITY
Accuracy, instruction adherence, coherence, ambiguity, tone, semantic drift, unsupported claims.
// HUMAN JUDGMENT / AI SYSTEMS
DIANA MELANIA CIRLAN
Lucida // SANS — evaluation practice
Signal: high-resolution perception
Signal: high-resolution perception
I don't just decide whether an output is good.
I locate where it fails, why it fails, and what would make it reliable.
I locate where it fails, why it fails, and what would make it reliable.
// SIGNAL
Evaluation across language, perception and systems.
I work beyond surface correctness. I compare outputs against explicit constraints, inspect what is supported by the evidence, identify failure modes, and translate judgment into feedback that another reviewer could reproduce.
I am not optimizing for volume. I am optimizing for signal.
02 / VISUAL
PERCEPTUAL JUDGMENT
Composition, hierarchy, typography, spatial coherence, consistency, readability, aesthetic quality.
03 / EXPERIENCE
UX & INTENT
Usability, information architecture, cognitive load, affordance, user goal alignment, accessibility.
04 / SYSTEMS
RUBRIC DISCIPLINE
Criteria calibration, edge cases, failure-pattern analysis, side-by-side ranking, consistent rationale.
// METHOD
01
Define the target.What was the user actually asking the system to accomplish?
02
Separate evidence from inference.What is explicitly supported, what is merely plausible, and what was invented?
03
Localize the failure.Correctness, reasoning, hierarchy, tone, constraint adherence, usability, or risk.
04
Score reproducibly.A rating is useful only when another evaluator can understand why it was assigned.
05
Translate critique into action.State the smallest change that would materially improve the output.
RULE
Fluency is not evidence of correctness. A polished output can still violate the user's constraints, introduce unsupported information, create visual confusion, or fail the real task.
The purpose of evaluation is not to sound critical. It is to make the failure observable, classifiable, and fixable.
// FRAGMENTS
Four demonstration evaluations.
Self-authored portfolio exercises using public or synthetic material. They are designed to show evaluation reasoning without presenting confidential client work as a case study.
FRAGMENT 01 // MULTIMODAL GROUNDING
When a plausible description invents the scene.
IMAGE + LANGUAGE / A-B RANKING
TASK
Write accessible alt text in ≤ 25 words. Describe only visible content. Do not infer location, time of day, people, or camera position.
Candidate A
“A lone diver swims beneath tropical ocean waves at night as turquoise light filters through the water.”
DECISION
B ≫ A
A is fluent and visually plausible, but it invents a diver, tropical setting, ocean context, and night-time conditions. Those additions are not grounded in the image.
Unsupported entitiesHigh
“Lone diver” introduces a person that is not visible. This is a grounding failure, not a stylistic preference.
Speculative contextMedium
“Tropical,” “ocean,” and “at night” convert ambiguity into invented certainty and violate explicit instructions.
Candidate BPass
“Dark rippling water with soft teal light across the surface and faint suspended particles below.” Specific, grounded, concise, and useful as alt text.
Groundedness1/5
Instruction fit2/5
Clarity5/5
FRAGMENT 02 // UX & VISUAL QA
A screen can look polished while communicating the wrong action.
VISUAL SYSTEM / USABILITY REVIEW
north / objectCHECKOUT 03/03
Confirm purchase
Review the final amount before placing your order.
TOTAL
Order totalThere may be an issue with the selected payment method.
Your card will be charged after confirmation.
PRIMARY FAILURE
Action hierarchy is inverted.
The destructive/exit action has the highest contrast while the intended completion action looks disabled. Visual hierarchy contradicts task hierarchy.
Affordance conflictHigh
“Confirm purchase” uses low-contrast styling typically associated with a disabled control. A user cannot reliably infer whether the action is available.
Error visibilityHigh
The payment warning has extremely weak contrast and no remediation path. It may be missed even though it can block completion.
Redundant hierarchyLow
“TOTAL / Order total” duplicates the same concept while consuming scarce attention at the decision point.
CorrectionAction
Make confirmation the primary high-contrast action, demote cancellation, increase warning contrast, and attach a concrete resolution step to the payment state.
FRAGMENT 03 // EDITORIAL INTENT
Correct facts can still fail the user's objective.
LANGUAGE / INSTRUCTION ADHERENCE
TASK
Reply to a customer whose delivery is eight days late. Acknowledge the delay, offer either a refund or replacement, keep the reply under 70 words, and do not blame the courier.
CANDIDATE AREJECT
“We're sorry your order hasn't arrived. Unfortunately, the courier has been experiencing severe delays outside our control. We understand this is frustrating. We can investigate the shipment and get back to you once we hear from them. Thank you for your patience while our delivery partner resolves the issue.”
CANDIDATE BSELECT
“I'm sorry your order is now eight days late. You shouldn't have to keep waiting without a clear resolution. I can arrange either a full refund or a replacement shipment — whichever you prefer. Reply with your choice and I'll make sure the next step is handled.”
01 / CONSTRAINTS
A misses two explicit requirements.No refund/replacement choice; it blames the courier despite a direct prohibition.
02 / USER GOAL
A delays resolution again.“We can investigate” creates another waiting state instead of resolving the eight-day delay.
03 / TONE
B restores agency.It acknowledges harm without over-apologizing and gives the user a concrete, low-friction choice.
FRAGMENT 04 // EVIDENCE BOUNDARY
High-risk evaluation begins with knowing what the source does not say.
HEALTHCARE-ADJACENT / GROUNDING
SYNTHETIC DEMO LABEL — NOT A REAL PRODUCT
DERMA-7 GEL
TOPICAL
30 mL
30 mL
- Apply a thin layer to intact skin twice daily.
- Avoid contact with eyes.
- Store between 15–25°C.
- Stop use if significant irritation occurs.
MODEL SUMMARY
“Apply twice daily to intact skin and avoid the eyes. Refrigerate after opening and use within 30 days. Stop if irritation becomes severe.”
DECISION
Fail: unsupported operational instructions.
The model preserves several source facts, which makes the fabricated storage and expiry instructions easy to miss. This is exactly why fluent summaries require source-by-source checking.
Storage contradictionHigh
The source says 15–25°C. “Refrigerate” is not merely unsupported; it can conflict with the provided storage range.
Invented expiryHigh
“Use within 30 days” has no evidence in the source. In accuracy-critical contexts, plausible completion is still fabrication.
Correct evaluator actionAction
Remove both claims. If expiry-after-opening is required, flag the source as insufficient rather than completing the missing information from general expectation.
This is a synthetic evaluation exercise only. It is designed to demonstrate evidence discipline and does not provide medical or product-use advice.
// BACKGROUND
Why the judgment is cross-disciplinary.
The evaluator is not separate from the domain knowledge. My professional background trained different parts of the same skill: seeing structure, reading precisely, detecting inconsistency, and communicating corrections clearly.
TECHNICAL WRITING / FEMIEKO
Clarity under constraint.
Healthcare-focused documentation, terminology consistency, ambiguity detection, information structure, and user-facing precision.
PRODUCT + UX / FLIPSIDE RESEARCH
Systems that must make sense to people.
Interface logic, visual hierarchy, information architecture, branding systems, content, front-end implementation, and quality review.
FINE ARTS + MULTIMEDIA DESIGN
Formal visual observation.
Composition, proportion, perspective, typography, hierarchy, visual consistency, illustration, and perceptual sensitivity.
DATA + REGULATED WORK
Accuracy before confidence.
Structured information checking, anomaly detection, procedural consistency, and attention to consequential details from data-quality and pharmacy work.
// ACCESS
Good evaluation is not a reaction.
It is a traceable decision.
Lucida // SANS is the conceptual framework for this evaluation practice. The professional behind it is Diana Melania Cirlan.