Neurologyca platform: architecture, validation and declared limits

01 · System architecture

Four components, one direction of data flow

The platform is a pipeline with hard boundaries between measurement, inference and presentation. Each boundary is a versioned contract, and each component can be deployed on its own.

1. Capture — on-device feature extraction · no image leaves the device
2. Engine — protocol orchestration · proprietary models · calibration
3. Konstellation — deterministic case resolution · evidence grounding · narrative
4. Client application — host-supplied interface

Two properties matter more than the boxes. Data flows one way only: inference never asks measurement for a different reading, so every result can be reproduced from the stored measurement alone. And every arrow is a versioned contract, so components can be scaled or replaced independently.

• Capture — Owns: Feature extraction on the user's device; Does not own: Any interpretation
• Engine — Owns: Protocols, model runtime, calibration, reliability, delivery; Does not own: What a number means; thresholds; cases
• Konstellation — Owns: Case resolution, evidence retrieval, narrative, audience profiles; Does not own: Signal processing; recalculating a measurement
• Host application — Owns: Interface, identity, storage, business workflow; Does not own: Nothing in the pipeline

Thresholds and interpretive cases live in Konstellation's configuration, never in Engine. An indicator has no opinion about its own meaning, which lets the same signal be read differently in different contexts without forking a model.

Multi-tenancy by request

Tenant isolation is enforced on every call, not per deployment. Retrieval over the shared scientific corpus is filtered by the caller's visibility scope, so one client's private domain can never surface in another's results.

02 · Capture layer

Geometry on the device, no image transmitted

The capture layer emits facial geometry, not pixels. This is a privacy decision with architectural consequences for every model downstream.

• Signal: dense 3D facial geometry (plus expression coefficients and head pose)
• Where it runs: in the browser (standard consumer camera, no install)
• What crosses the network: numbers only (no image, video or audio of the subject)

Because only numeric frames leave the device, the standard deployment never transmits a picture of the person being measured. A server-side mode exists as an extension point for settings that require it, without changing any contract.

Limits: What geometry-only capture costs

Appearance-based signals are unavailable to the camera models: skin texture and anything that needs photometric information cannot be inferred from geometry. Landmark quality degrades before it fails: occlusion, extreme pose and low light reduce stability, and that shows up as a lower reliability value rather than as an error.

Every channel therefore carries a reliability value per window. Composite channels inherit the reliability of their weakest input, so one degraded input is never hidden behind several healthy ones. Some expression coefficients from the extraction layer carry little or no usable information; this is tracked per input, and channels that depend on them are flagged as open research lines.

Additional modalities

Individual contexts can add their own capture: EEG, electrodermal activity and eye tracking in laboratory settings, and passive behavioural telemetry where the context is defined by interaction rather than by a session. These additions do not change the core contract.

03 · Engine · measurement

Measurement designs as data, not code

Engine executes protocols. A protocol is configuration, so a new measurement design is a configuration change rather than a software release.

1. Protocol — what is measured, and how often
2. Mechanism — what one sitting looks like
3. Step — one stimulus or task: baseline · video · open answer
4. Session — one execution, bound to a subject

Published protocols are immutable. Editing one creates a new version, so every stored result points to the exact definition it ran under. That makes comparison across months auditable rather than approximate.

Calibration

Every channel is calibrated twice: against the population, so scores are comparable, and against the individual's own baseline once one exists, so changes are personal. The payload always states which calibration was applied; consumers never have to guess.

Delivery

Three delivery modes carry the same payload: live streaming for interactive sessions, polling where long connections are impractical, and webhooks with retry for back-office processing.

04 · Taxonomy and configuration

Four levels and one precedence rule

The taxonomy is why a new application domain is a set of configuration records rather than a software project.

1. Context — e.g. Sport, Wellness, Hiring
2. Subcontext — e.g. a competitive moment, a wellbeing state
3. Analytical domain — indicators + thresholds + interpretive cases
4. Situation — one resolved interpretive outcome

The most specific definition that matches wins.

A level that defines nothing inherits from its parent.

A new context can specialise an existing one, inheriting indicators, thresholds, evidence and audience profiles and overriding only what differs. Sport is built as a specialisation of Wellness on exactly this mechanism.

• Which indicators a domain uses — Requires: configuration
• Thresholds and bands — Requires: configuration
• Interpretive cases — Requires: configuration
• Scientific evidence bound to a context — Requires: ingestion run
• Audience profiles and narrative style — Requires: configuration
• A new model or a new capture signal — Requires: software release

The line is drawn at the signal. How a signal is interpreted is data; the signal itself is code. A context built from existing signals ships without a release.

Ambiguity is rejected at load time

A domain in which two interpretive cases could match the same evidence does not load. Determinism is enforced when configuration is authored, not hoped for at runtime, and any state a domain does not declare resolves to an explicit "uncatalogued" result.

05 · Konstellation · inference

The language model writes; it does not decide

Inference runs as a graph of stages. The property that matters for audit is which stages may use a language model and which may not.

The determinism boundary

Case resolution never calls a language model. It is a deterministic function of the measurement and the domain configuration: the same input gives the same case, every time, and the case can be recomputed offline from stored data with no model provider available.

The language model does two things only: it retrieves supporting evidence and it writes the narrative for a given audience. It never decides what happened, so the operational or clinical decision surface is not stochastic.

Two halves of every result

• Machine-readable, deterministic (structured) — Indicators with reliability, the resolved case, and an emotional reading. This is what automated systems should consume.
• Human-facing, audience-profiled (interpretation) — A narrative with citations to the evidence that grounds it. Always derived from the structured half, never the reverse.
• Provenance (metadata) — Protocol and domain versions, calibration applied, degraded steps and latency.

Three service profiles

• Full — Response time: ≈ 2 min; Includes: Case, evidence, narrative; Intended consumer: Reports, back office
• Streaming — Response time: progressive; Includes: Same, emitted stage by stage; Intended consumer: Live sessions
• Lite — Response time: < 5 s; Includes: Case resolution only; Intended consumer: Software agents and adaptive systems in a request loop

The lite profile exists because the consumer is increasingly not a person. An agent that needs human context inside its own turn cannot wait two minutes and does not need prose. Under load it returns "uncatalogued" rather than a guessed case.

Audience profiles

The same measurement can be narrated differently without recomputation: coach and athlete, interviewer and interviewee, clinician and patient. Each profile sets register, level of detail and what is withheld.

06 · Evidence grounding

Every explanation cites its science

When the system says why, it cites. Retrieval is built so that those citations are specific to the situation and auditable.

Specific evidence first

Evidence is bound at three levels: this situation, then this analytical domain, then the context. Retrieval descends in that order, and general literature never displaces specific literature. A reading in sport is grounded in sport science, not in generic texts that happen to match.

Search is hybrid and multilingual, because the scientific corpus is not monolingual and purely semantic search silently under-retrieves minority-language sources.

Grounding belongs to the context, not the indicator

The same indicator means different things in different contexts, so it has no literature of its own. A context without a bound corpus produces structured results with an empty citation list: a visible state, never a silent fallback to generic sources.

07 · Proprietary models

Five models, each with its figure and its scope

Five proprietary models estimate affective and cognitive magnitudes from facial geometry. Each is reported with what it measures, what it was validated against, and the figure that supports it.

• Emo — Measures: Displayed emotion, 7 classes; Ground truth family: Externally perceived emotion; Headline figure: macro-F1 ≈ 0.43; Evaluated on: Identity-disjoint, three held-out sets (> 100k training frames)
• Valence — Measures: Perceived valence, −100…+100; Ground truth family: Human judgement; Headline figure: r ≈ 0.72 · AUC ≈ 0.98; Evaluated on: ~50k human-voted frames outside training data
• Arousal — Measures: Expressive intensity, 0–100; Ground truth family: Multi-corpus reference; Headline figure: CCC ≈ 0.84; Evaluated on: Unseen corpus (≈ 0.73 identity-disjoint)
• FAA — Measures: Motivational direction (approach / withdrawal); Ground truth family: Laboratory EEG; Headline figure: AUC ≈ 0.81; Evaluated on: Within subject, out of sample, ~460 subjects, ~50 studies, 4 EEG systems
• Workload — Measures: Cognitive load proxy, 0–100; Ground truth family: Laboratory EEG; Headline figure: r ≈ 0.48 pooled; Evaluated on: Out of sample; within-subject median ≈ 0.33

Pooled figures absorb differences between people; within-subject figures match the production use case of tracking one person over a session. Where both exist, both are given so neither has to be inferred.

Selected successor · cognitive load from in-house EEG

A successor Workload model trained entirely on in-house EEG has been selected. Its figures are within-subject and out of sample throughout, and it holds at production frame rate without changes to capture.

• Within-subject r: ≈ +0.18 (leave-one-study-out and subject-disjoint agree)
• 95 % CI: [+0.15, +0.21] (no permutation replicate at or below zero)
• Consistency: 13 / 13 (studies positive · ~80 % of subjects)

The signal sits in the facial regions where the cognitive-load literature places it, which supports a physiological reading rather than a dataset artefact.

Limits: How each model may be described

Emo measures displayed, not felt, emotion. Valence is observer-perceived and absolute levels are not comparable across studies. Arousal measures expressive intensity, not autonomic activation. FAA supports within-subject ranking and extreme detection, not absolute values, and must not be presented as valence. Workload is a facial estimate of an EEG proxy, not measured cognitive load, and ranks better than it levels. All camera models perform best on single-person, passive-viewing material.

08 · Validation programme

Sensor validity and construct validity, tested separately

Beneath the five models sits a layer of facial-action channels that compose higher-order constructs. Every shipped channel goes through a structured validation pass against independent ground truth.

The distinction the method is built on

An emotion-labelled dataset can answer only one question: does the channel detect the facial movement it is built from. That is sensor validity. When a channel's target is a physiological or psychological state, only physiological ground truth can show whether it predicts that state. That is construct validity. Both are tested, separately, and a validation claim that does not say which one it refers to is not reported.

How a channel is validated

• Independent ground truth (reference) — Chosen for what the channel needs: human labels, eye-tracker gaze and fixations, skin conductance or EEG. Never the channel itself.
• Acted and spontaneous material (replication) — Results are replicated across posed and candid footage to separate a measurement gap from a performance artefact.
• Nulls traced upstream (root cause) — A weak result is traced back to its input before it is attributed to a weak effect. This is how silent input defects are found.
• Quality gates applied retroactively (gates) — Minimum face size and maximum head angle are applied to every pass, and results that do not survive them are downgraded to inconclusive.

Head pose accuracy

Head pose is reported in degrees by both the platform and an independent reference system, which yields true accuracy figures rather than correlations.

• Roll: ≈ 2.5° (mean absolute error)
• Pitch: ≈ 4.2° (mean absolute error)
• Yaw: ≈ 7.7° (mean absolute error)

Limits: Open research lines

Validation status differs by channel. The five models in §07 are validated against their own ground truth. Several action-layer channels, and the higher-order composites built on them, have not yet shown construct validity against physiological reference and are treated as open research lines. Neurologyca does not present those composites as validated measures, and the platform's reliability and "uncatalogued" mechanisms keep them from producing unsupported results at runtime.

09 · Ground truth and laboratory

Derived physiological magnitudes, not annotations

Ground truth has a specific and falsifiable definition here: objective neurophysiological magnitudes, derived deterministically from clinical-grade biosignals and normalised against each participant's own neutral baseline.

• EEG — Origin: Direct recording from contact sensors in the laboratory; What the model learns: Physiology
• Electrodermal — Origin: Skin-conductance response in the laboratory; What the model learns: Physiology
• Perceived emotion — Origin: Consensus over externally annotated corpora; What the model learns: Externally perceived emotion

Both kinds of ground truth are legitimate and measure different things. The family is declared per model, so a downstream claim such as "this person felt X" can be checked against what the producing model can support.

Magnitudes derived this way include frontal alpha asymmetry, activation, cognitive load, attention, acute stress and fatigue. Labels are synchronised frame by frame with video, and reference recordings span four different EEG systems, which forces models to generalise across equipment as well as across people.

Laboratory repository

• Sessions: > 7,000
• Unique subjects: ≈ 7,000
• Studies: > 300
• Trials: > 90,000
• EEG: > 5,500 sessions
• Electrodermal: > 5,500 sessions
• Eye tracking: > 3,800 sessions
• All four modalities: ≈ 2,900 (≈ 40 % of the repository)

The four-modality subset allows cross-channel validation on the same person at the same instant: checking that a facial estimate of load follows that subject's frontal EEG in that second, not a group average.

10 · Indicator catalogue

Three tiers, thirty registered indicators

Every indicator can be traced down to the raw signal that produced it, which makes an output auditable rather than merely plausible.

1. HCI — Human Context Indicator — what a domain reads to resolve a case
2. MCI — meta-indicator — intermediate composition
3. RCI — raw indicator — the signal as captured

Nine categories organise the catalogue, including Attention, Mood, Emotions, Cognition and Voice, plus higher-order composite indices. The thirty registered Human Context Indicators are:

Attention

• Attention
• Effective Active Attention
• Active Audience

Cognition

• Workload
• Coherence
• Comprehension
• Confusion
• Mental Fatigue Load
• Alert

Emotions

• Joy · Sadness · Surprise
• Anger · Disgust · Fear
• Neutral
• Valence
• Intensity

Mood and regulation

• Interest
• Frustration
• Rejection
• Stress intensity
• Motivation
• Emotional Mental Control
• Facial Muscle Tension

Camera physiology

• Breathing
• Heart rate
• Beats per minute
• Inter-beat interval
• Heart rate variability

No context uses all thirty: Hiring selects fifteen, Education seven, Driving four. The catalogue is the superset, and each domain's selection is configuration. Being in the catalogue does not mean every indicator has the same validation maturity; see §07 and §08.

11 · Active contexts

What is configured today, and how mature it is

Seven contexts are configured, each with its own indicators, resolution mechanism and evidence corpus. Maturity is stated per context, because "supported" and "deployed" are different claims.

• Wellness · parent context — Subject: Individual user; Scope: Five wellbeing subcontexts with longitudinal follow-up, in a declared pre-diagnostic frame; Maturity: configured
• Sport · specialises Wellness — Subject: Athlete; Scope: Daily check-in, pre-training and press-conference modules, with more than 200 declared cases; Maturity: configured deepest corpus
• Hiring — Subject: The interview as a process, not the candidate; Scope: Fifteen indicators, including indicators specific to interview dynamics; Maturity: configured
• Education — Subject: Student, after a lesson; Scope: Biometric and verbal readings resolving to a learning profile and an intervention path; Maturity: configured
• Market Research · neuroqualitative — Subject: The stimulus; the participant is the instrument; Scope: Integrated stimulus and speech situations, each with a research-usability verdict; Maturity: configured phase 1 validated on packaging
• Gambling · Care — Subject: Player, over time; Scope: Passive behavioural telemetry instead of session-based capture; Maturity: in definition
• Driving — Subject: Driver; Scope: Driver-state scenarios over a small indicator set; Maturity: demonstration not deployed

A configured context has its indicators, thresholds, cases and evidence defined and loadable. That alone does not mean every composite it uses has cleared validation.

The marginal cost of a new context

A new context needs an indicator selection, thresholds for that context, interpretive cases, an evidence corpus and an audience profile. All five are configuration. It needs no new model, capture signal or engine, so adding a context costs the work of writing its science, not of building a platform.

12 · Context Maker

From a client brief to a loadable context

Context Maker turns a client's brief, in the client's own words, into a ready-to-load context grounded in scientific literature and limited to what the platform can actually measure.

• Structure the need (understand) — An agent interviews the brief until it becomes a set of analytical questions.
• Read the science (research) — Literature is planned, retrieved and read in full text, and every claim keeps its source.
• Check the inventory (match) — The need is matched against the live catalogue of indicators, signals and models.
• Emit configuration (build) — Indicators, thresholds, cases, evidence and audience profile, ready to load.

Why it is more than a research assistant

The platform's own inventory is consulted live while the context is assembled, so a generated context cannot ask for an indicator that does not exist, a signal the capture does not emit, or a model that was never trained. It is provider-agnostic and can be embedded behind a partner's own interface through a single versioned API.

13 · Interpretive mechanisms

Three ways to resolve a case on one engine

The strongest evidence that the platform is generic is not the number of contexts. It is that one engine runs three structurally different case-resolution mechanisms, none of them hard-coded.

• Graded matrix — Principle: Indicators crossed in graded bands, then refined by modulators; Used by: Sport
• Binary matrix — Principle: High/low readings combined into a compact case key; Used by: Education, Market Research
• Pattern matching — Principle: Levels of several indicators matched against declared patterns; Used by: Wellness

All three share the same triage, the same emotional model and the same inference graph. What differs between them is configuration.

The matrix decides; the emotional model names

Neurologyca maintains a deterministic registry of nearly a hundred emotions, each placed on valence and intensity and described in cognitive, emotional, behavioural and adaptive terms. It is used inside an already-resolved case to supply precise vocabulary: a matrix cell cannot tell frustration from disappointment from impatience, and the registry can. An emotional label is never allowed to decide the case.

Readings that fall close to a threshold are resolved and flagged as marginal, so the consumer knows when a value sat on a boundary.

14 · Declared limits and failure modes

What the system does when it does not know

An applied-neuroscience system that cannot say "I don't know" is not deployable. These are the states in which the platform declines to produce a result, and what it returns instead.

• No interpretive case matches — Behaviour: Returns an explicit "uncatalogued" result
• Input requirements not met — Behaviour: The channel is not emitted at all (fail-closed)
• Low signal quality — Behaviour: Value returned with reduced reliability; composites inherit the weakest input
• A step could not be measured as designed — Behaviour: Recorded as degraded; the session continues
• No personal baseline yet — Behaviour: Population calibration only, declared in the result
• Value estimated rather than measured — Behaviour: Estimation method declared
• Lite response time exceeded — Behaviour: "Uncatalogued", never a guessed case
• No evidence corpus for the context — Behaviour: Structured result with no citations

Each of these is a distinct, visible state in the result. None is a silent fallback, and none produces a fabricated result.

Limits: Comparability across devices and studies

Absolute levels from camera-derived channels are not comparable across recording setups: camera angle and distance shift readings materially even when correlation with the reference stays high. Protocols that need level comparison constrain the setup (supported device, stable distance, angle gate); protocols that only need within-person ranking do not.

For camera-based heart signals, skin pigmentation affects signal quality. Performance is audited per skin-tone group using an objective colour measure rather than self-reported phototype.

Scope boundaries that are product decisions

• No intent inference. Contexts describe states and behavioural outcomes. No case resolves to an intention.
• No deception detection. Coherence measures internal consistency of speech. Low coherence is most often cognitive load, and high coherence is compatible with rehearsed delivery. Neither supports a claim about honesty.
• No diagnosis. One context works in a declared pre-diagnostic frame with its own recipient and entry route. No other context makes clinical claims.
• Process, not traits. In assessment contexts, indicators describe the state of the process, not traits of the person: "this interview produced rejection at minute twenty", never "this candidate is a person who rejects".

15 · Engineering discipline

The parts that do not appear in a demo

Indicators of whether a system is maintainable, not only whether it works today.

• Automated tests: > 2,000
• Versioned schema migrations: > 100
• Regression harnesses: interpretive drift (stored sessions replayed against candidate configuration)
• Model parity: verified per model (training and serving outputs match)

In a configuration-driven system the main risk is not broken code. It is an edited threshold silently changing which cases resolve. The regression harnesses replay stored sessions against a candidate configuration and compare the resolved cases, which turns interpretive drift into something a reviewer can inspect.

Reproducibility guarantee

Given a stored measurement and a domain version, case resolution is exactly reproducible offline, with no model provider, network or retrieval. Only the narrative needs a language model, and the narrative never feeds a decision. An auditor can verify the decision surface without access to the inference infrastructure.