Research Context and Historical Scope
This study documents the engineering decisions and lessons from an early validation prototype of Perfume Engine — an experimental system created to evaluate claim-level provenance within recommendation architectures.
Status & Scope Clarification: This case study examines a specific, historical engineering prototype (12 perfume profiles, exact deterministic search, immutable static projection). Later iterations of Perfume Engine expanded far beyond these constraints into a broader persistent product and data platform. Subsequent design and product reviews determined that the compact prototype was not an acceptable final consumer product contract, triggering an extensive recovery process. This document evaluates the architectural mechanics of isolating evidence from generative heuristics, not the commercial launch of a finished service.
The Core Problem: The Collapse of Provenance in AI Outputs
Conventional recommendation engines and generative AI interfaces typically act as opaque statistical aggregators. When presented with a recommended alternative or summary, the end user cannot distinguish:
- What represents a verified primary fact (e.g., an official olfactory pyramid published by the brand)?
- What represents subjective editorial assessment from an evaluator or perfume critic?
- What is an algorithmic inference derived from note overlap?
- What is the actual degree of uncertainty behind the pairing?
When teams introduce generative large language models (LLMs) into the runtime request loop, this collapse intensifies. The model generates fluent, authoritative explanations (“this perfume shares an amber depth and smoky finish”) that may have no basis in verified formulation data or documented olfactory pyramids.
The engineering challenge was to construct the inverse: a recommendation engine where every assertion retains explicit provenance, where computation is fully reproducible, and where publishing fails closed if evidence standards are not satisfied.
Five-Layer Architecture: Isolating Evidence from Inference
To eliminate stochastic behavior, the prototype decoupled data handling into five strictly bounded layers:
[Layer 1: Raw Evidence]
↓ (Manual acquisition, primary source URL, timestamp)
[Layer 2: Normalized Claims]
↓ (Canonical entities, alias and flanker resolution)
[Layer 3: Editorial Assessment]
↓ (Dual-assessor consensus, structured agreement)
[Layer 4: Deterministic Inference]
↓ (Versioned mathematical formulas, zero LLM)
[Layer 5: Immutable Public Projection]
↓ (Content-hashed artifacts, fail-closed build)
[Client Interface: Exact Search & Alternative Graph]
1. Raw Evidence
Every assertion regarding an olfactory profile, release year, perfumer, or formulation was recorded as an atomic data point tethered to a primary URL (such as manufacturer technical sheets or certified documentation). Public client interfaces were strictly isolated from raw records.
2. Normalized Claims
Fragrance taxonomy contains extensive terminology fragmentation (“Calabrian bergamot”, “bergamot peel”, “citrus accord”). This layer enforced a controlled vocabulary. No unmapped or ambiguous term was permitted to pass into the similarity calculation pipeline.
3. Editorial Assessment
Perceived similarity between fragrances is not an absolute physical constant; it contains subjective sensory evaluation. The prototype formalized this through a dual-assessor consensus mechanism. An alternative relationship became valid only when two independent evaluators agreed across defined sensory dimensions (scent profile, dominant accords, seasonal drydown).
4. Deterministic Inference
Rather than querying an LLM or computing cosine similarity across high-dimensional vector embeddings, the engine applied transparent, versioned formulas:
- Accord Overlap Index: Jaccard similarity across canonicalized accord sets;
- Consensus Coefficient: Weighted multipliers reflecting assessor agreement;
- Confidence Metric: A transparent score quantifying the completeness of collected evidence rather than a simulated probability.
5. Immutable Release Projection and Fail-Closed Gates
Valid recommendations were compiled into immutable JSON artifacts tagged with SHA-256 integrity digests. If an entry lacked source verification, contained an unresolved assessor conflict, or failed schema validation, the release generation aborted immediately. Broken or unverified data could never be deployed to the client.
Why Omitting a Runtime LLM Was an Architectural Advantage
Excluding generative LLMs from the request path yielded decisive operational benefits:
- Zero Marginal Inference Cost: Serving deterministic static data requires minimal compute and avoids fluctuating API token expenses.
- Sub-Millisecond Response Times: Exact lookups execute instantaneously without queuing for token generation.
- Deterministic Reproducibility: Identical queries guarantee identical outputs and identical explanations, eliminating non-deterministic regressions.
- Absolute Hallucination Immunity: The system cannot invent olfactory notes or spurious brand relationships because claims are constrained to the validated graph.
Failure Modes and Prototype Limits
The primary value of the prototype was revealing where deterministic isolation reached its practical limits:
- The Exact Search Trap: Requiring exact title and flanker matches without client-side fuzzy heuristics frustrated users who misremembered specific product concentrations (e.g., Eau de Toilette vs. Eau de Parfum).
- Editorial Bottlenecks: Dual-assessor manual verification works reliably for 12 records, but becomes an unsustainable operational bottleneck across a multi-thousand perfume catalog.
- The Gap Between Technical PASS and Product Value: The prototype passed all unit, schema, and projection tests with 100% compliance. However, internal product reviews highlighted that technical correctness alone does not produce an engaging, intuitive consumer product without flexible search, rich UX, and comprehensive catalog depth.
Architectural Implications for AI Engineering
The Perfume Engine validation experiment underscores a fundamental principle of intelligent systems architecture: true system intelligence is reflected in how cleanly an architecture demarcates verified facts from subjective interpretation and algorithmic inference.
Deterministic logic and claim-level provenance remain indispensable engineering tools whenever accountability, auditability, and verifiable correctness take precedence over open-ended generative text.