PERSONAL STIMULUS-RESPONSE MODELING

Draft patent specification without claims

Project: HumanFB
Document date: August 15, 2026
Draft status: Claims intentionally omitted for this drafting pass
Drafting assumption: No relevant prior art is assumed for purposes of developing the disclosure; no prior-art search or patentability conclusion is included

TITLE OF THE INVENTION

[0001] Personal Stimulus-Response Modeling.

[0002] Not Applicable.

STATEMENT REGARDING FEDERALLY SPONSORED RESEARCH OR DEVELOPMENT

[0003] Not Applicable.

THE NAMES OF THE PARTIES TO A JOINT RESEARCH AGREEMENT

[0004] Not Applicable.

INCORPORATION-BY-REFERENCE OF MATERIAL SUBMITTED ELECTRONICALLY

[0005] Not Applicable.

STATEMENT REGARDING PRIOR DISCLOSURES BY AN INVENTOR OR JOINT INVENTOR

[0006] Not Applicable.

BACKGROUND OF THE INVENTION

Field of the invention

[0007] The present disclosure relates generally to computer-assisted acquisition and modeling of individual human responses to stimuli and interventions. More particularly, the disclosure relates to systems and methods for identifying a stimulus presented or applied to an individual, measuring or otherwise characterizing the presentation or delivery of the stimulus, acquiring an explicit or sensed response of the individual, associating the response with the stimulus in time and context, generating one or more machine-readable interpretations of the response, constructing a reusable personal response profile, and using the profile to select, retrieve, generate, recommend, configure, or control a later stimulus or intervention.

[0008] Embodiments may be applied to digital content, physical stimulation, environmental stimulation, ingested or administered products, self-administered activity, interactive systems, human or machine conversation, reconfigurable devices, robotic systems, and combinations thereof. Digital content may include text, images, audio, video, software interfaces, electronic messages, and conversational turns. Physical stimulation may include mechanical pressure, palpation, vibration, movement, temperature, light, sound, fragrance, flavor, electrical stimulation, or another physical input. Administered stimuli may include food, beverages, medicines, supplements, cannabis products, or other substances.

[0009] Recommendation and personalization systems may infer user interest from clicks, dwell time, purchases, playback, abandonment, or broad ratings. Such evidence can indicate that an item was encountered without establishing which component or interval elicited a response, what the user perceived, why the user responded, how the response depended on the user's starting condition, or whether the immediate response differed from a later evaluation.

[0010] A single source item may also be presented in materially different ways. A document may be excerpted, reordered, resized, summarized, or displayed in a particular visual context. A media item may be played at a selected volume, speed, crop, language, or starting position. An interactive conversation may branch and expose only part of its history. Associating a response only with a source identifier can therefore lose the presentation conditions that contributed to the response.

[0011] Physical interventions introduce an additional discrepancy between an intended stimulus and the stimulus actually delivered. A robot may be commanded to apply a nominal force at a nominal body location, while force, contact area, pose, compliance, slip, or localization cause a different physical input. A user may report ingesting an estimated quantity of a product without a directly measured dose. A reconfigurable device may be commanded to enter a state that differs from an observed state. A useful response record should preserve both intended and measured delivery and the confidence with which each is known.

[0012] Spoken comments can convey rich perceptual information that predefined rating controls do not anticipate. However, requiring a person to locate, press, or hold a recording control can interrupt an experience and may prevent spontaneous capture. Continuous remote processing of ambient audio can consume network or model resources even when no meaningful feedback occurs. A local trigger mechanism with pre-trigger buffering can reduce interaction friction while preserving the stimulus interval that preceded a spoken reaction.

[0013] Responses are also dynamic. A person may begin an episode in a particular mood, pain state, level of fatigue, hunger state, or motivational state. A stimulus may produce an immediate reaction, a delayed peak, adaptation, rebound, or a later reflective evaluation. Treating such a trajectory as one item score can lose information useful for predicting how another stimulus will affect the person under a different condition.

[0014] Multiple response channels may further have different timing and meaning. Speech may refer retrospectively to an earlier interval. A withdrawal movement may occur quickly. An electrodermal signal may occur after a physiological delay. A replay action may indicate interest, confusion, or verification. A system capable of retaining raw observations, latency information, stimulus anchors, explicit statements, and alternative causal attributions can support more accurate personal modeling than a system that stores only a final label.

[0015] There is accordingly a need for a technical framework that can associate arbitrary digital and physical stimuli with time-aligned, context-dependent, individual response evidence; preserve a distinction between observation and interpretation; reprocess historical evidence using later models; and compile a resulting personal response model into multiple downstream applications.

BRIEF SUMMARY OF THE INVENTION

[0016] In one aspect, a system obtains a representation of a stimulus associated with an exposure of an individual. The representation may identify a source resource, an immutable stimulus snapshot, one or more stimulus components, and an intended presentation or intervention protocol. The system obtains a presentation trace or delivered-stimulus trace representing at least part of what was actually presented or applied. The trace may include digital presentation events, physical measurements, device-state observations, administration information, or combinations thereof.

[0017] A response capture subsystem acquires response evidence associated with the exposure. The response evidence may include a voice recording, transcript, rating, pairwise comparison, selection, text entry, body-location indication, behavior, physiological signal, movement, facial signal, gaze signal, or later assessment. The system associates the response evidence with one or more anchors identifying a time interval, text span, image region, conversation turn, body location, physical phase, device state, product-effect phase, or transition.

[0018] An interpretation subsystem processes the response evidence to produce one or more versioned interpretation artifacts. The artifacts may represent semantic content, acoustic features, perceptual qualities, subjective axes, affective estimates, sensory states, concepts, comparisons, references, causal-attribution candidates, confidence values, embeddings, or open-vocabulary expressions. The original response evidence and each derived interpretation may remain separately addressable.

[0019] A profile-generation subsystem generates a personal response profile using a plurality of exposure records. The profile may include typed conditional parameters, learned embeddings, positive or negative exemplars, pairwise relationships, contextual effects, temporal effects, cross-domain relationships, exceptions, uncertainty values, or combinations thereof. A profile parameter may estimate how a feature of a stimulus or delivered intervention affects one or more response axes for the individual under a specified context or objective.

[0020] A compiler or decision subsystem applies a selected profile snapshot to a candidate stimulus, current state, objective, constraint, cost, or availability observation. The subsystem may generate a search query, retrieval representation, candidate ranking, recommendation, AI interaction context, content-generation specification, practitioner brief, experiment proposal, device configuration, or physical intervention proposal. The generated artifact may identify the profile version, evidence version, objective, uncertainty, and factors contributing to the output.

[0021] In another aspect, a local audio subsystem monitors an audio input using voice-activity detection, wake-word detection, speaker recognition, feedback-intent detection, or combinations thereof. The local subsystem maintains an audio and system-state ring buffer without requiring a remote language-model request. Upon detecting a trigger, the subsystem freezes pre-trigger data, samples a current stimulus and user-state context, records a voice memo, and associates the memo with the exposure and preceding stimulus interval.

[0022] In another aspect, a prompting subsystem detects an informative event and presents an audio prompt requesting a response. The event may include replay, pause, dwell, an observed state change, a prediction error, an uncertain comparison, an exposure boundary, or expected information gain. The prompt may be selected locally from stored audio or generated by local or remote speech synthesis. The prompt itself is recorded as a stimulus, thereby enabling separation of a response to the original stimulus from a response influenced by the prompt.

[0023] In another aspect, a physical-intervention system receives a requested intervention and applies a local capability or limit policy before actuation. A real-time controller operates an actuator and records a measured delivered-stimulus trace. The trace may include force, torque, pressure, displacement, contact area, position, velocity, vibration, temperature, or other physical measurements. A spoken or sensed response is associated with the trace and an anatomical anchor. A later proposal may be generated from the resulting personal model while remaining subject to the local capability or limit policy.

[0024] In another aspect, an administered-stimulus system records a product identity, batch identity, composition, administration route, amount, dose confidence, purchase information, baseline state, and administration time. The system acquires a series of immediate and delayed assessments and generates an individualized response trajectory. Candidate products may be compared using predicted desired effects, predicted undesired effects, price, availability, uncertainty, and acquisition cost as distinct decision dimensions.

[0025] The disclosed architecture permits the same evidence corpus to support multiple objective schemas without overwriting the evidence. For example, physical-pressure episodes may support pleasure, relief, mobility, pain-mapping, or diagnostic-information objectives. Product episodes may support relaxation, focus, pain relief, low undesired-effect, best-value, or availability objectives. Digital-media episodes may support general preference, learning, emotional regulation, search, or content-generation objectives.

[0026] Technical benefits of one or more embodiments may include reduced feedback-input friction, time alignment between a response and a preceding stimulus, differentiation of intended and delivered stimulation, reconstruction of exact presentation conditions, retention of uncertainty and provenance, reprocessing of historical evidence, individualized latency modeling, and reusable personal response parameters applicable across multiple downstream systems.

BRIEF DESCRIPTION OF THE SEVERAL VIEWS OF THE DRAWINGS

[0027] FIG. 1 is a block diagram of an example HumanFB system including stimulus sources, presentation or intervention subsystems, response-capture subsystems, data storage, interpretation subsystems, personal-model subsystems, compilers, and target systems.

[0028] FIG. 2 is a diagram of an example Personal Stimulus–Intervention–Response Graph linking a source resource, stimulus snapshot, protocol, measured exposure, context, response evidence, interpretation artifact, profile snapshot, decision trace, and outcome.

[0029] FIG. 3 is a timeline illustrating an intended digital presentation, an actual presentation trace, a pre-trigger buffer, a spoken response, one or more candidate stimulus anchors, and a later assessment.

[0030] FIG. 4 is a block diagram of a local ambient voice-capture subsystem including echo cancellation, voice-activity detection, speaker recognition, trigger detection, ring buffers, state sampling, recording, and a local spool.

[0031] FIG. 5 is a sequence diagram illustrating detection of an informative event, presentation of an audio prompt, recording of the prompt as an intervening stimulus, capture of a response, and generation of an interpretation artifact.

[0032] FIG. 6 is a diagram of an example dynamic subjective-state model having long-lived personal parameters, a slower disposition state, an instantaneous state, stimulus inputs, context inputs, observation channels, and response outputs.

[0033] FIG. 7 is a block diagram of a physical-intervention embodiment including an intervention planner, local capability validator, real-time actuator controller, delivery sensors, body-location subsystem, response capture, and an emergency or fast-release path.

[0034] FIG. 8 is an example anatomical map and synchronized force-response timeline for robotic palpation or massage.

[0035] FIG. 9 is a timeline of an administered-stimulus episode including a baseline assessment, administration record, onset assessments, peak assessment, declining-effect assessments, and a later reflective assessment.

[0036] FIG. 10 is a diagram of an example personal response profile including typed parameters, contextual conditions, embeddings, exemplars, exceptions, temporal epochs, and uncertainty values.

[0037] FIG. 11 is a diagram of a profile compiler producing different scoped outputs for web search, catalog search, private reranking, AI conversation, content generation, a practitioner, an experiment planner, and a physical device.

[0038] FIG. 12 is a diagram of a multi-objective decision process using predicted response axes, price, availability, acquisition burden, uncertainty, information value, and one or more objective weights.

[0039] FIG. 13 is a diagram of a digital-content embodiment in which a reaction is associated with a text span, conversation turn, media interval, image region, or interactive-application state.

[0040] FIG. 14 is a diagram of an active-learning process that selects a candidate exposure or comparison, receives local validation where applicable, obtains a response, and updates a personal model.

[0041] FIG. 15 is a block diagram of alternative local, distributed, and server-assisted implementations.

DETAILED DESCRIPTION OF THE INVENTION

Overview and terminology

[0042] The following description provides examples sufficient to enable implementation of the disclosed systems and methods. The examples are illustrative rather than limiting. Features described in connection with one embodiment may be used with another embodiment, and operations may be reordered, combined, divided, omitted, or performed concurrently unless a particular ordering is required by the nature of the operation.

[0043] As used herein, “stimulus” includes any digital, physical, chemical, environmental, conversational, cognitive, or composite input presented or applied to an individual and capable of being associated with a response. A stimulus need not be intentionally selected. An unplanned sound, contextual event, self-generated thought, or response of the individual may itself become a stimulus for a later response.

[0044] As used herein, “intervention” includes a stimulus having an intended delivery characteristic, delivery protocol, selected target, or controllable actuator. An intervention may be digitally presented, physically actuated, environmentally produced, ingested, administered, or self-administered.

[0045] As used herein, “response evidence” includes an explicit report or control input and also includes measured behavior or physiology. Response evidence may include speech, a voice memo, text, a rating, a comparison, a replay, an abandonment, a selection, movement, muscle activity, facial activity, gaze, pupil response, electrodermal activity, heart activity, respiration, temperature, neural activity, or another observable signal.

[0046] As used herein, “subjective state” includes a modeled sensory, affective, cognitive, attentional, motivational, evaluative, or physiological condition of an individual. A subjective state may include one or more axes such as valence, activation, engagement, pain, pleasure, relief, comfort, tension, clarity, cognitive load, novelty, attraction, arousal, fatigue, hunger, focus, desire to continue, desire to repeat, or reflective satisfaction. A subjective state is an estimate unless explicitly reported as such by the individual.

[0047] As used herein, “profile” includes a reusable representation of relationships among stimulus features, delivery features, contexts, states, objectives, and responses for an individual. A profile may include a vector, graph, table, probabilistic model, neural-network parameter set, rule set, collection of exemplars, embeddings, distributions, or a hybrid thereof. A profile need not purport to describe a fixed personality.

[0048] As used herein, “anchor” includes a reference connecting response evidence to one or more portions, components, positions, phases, or transitions of a stimulus or exposure. Anchors may be exact, approximate, probabilistic, or user-confirmed.

[0049] As used herein, “AI” or “machine-learning model” includes statistical, rule-based, probabilistic, neural, symbolic, embedding-based, generative, discriminative, or hybrid computational techniques. An embodiment need not use a remotely hosted generative model and may perform one or more operations using deterministic logic or local signal processing.

Example system architecture

[0050] Referring to FIG. 1, an example system 100 includes one or more stimulus sources 102, a source-capture subsystem 104, a presentation subsystem 106, an intervention subsystem 108, a timing and event subsystem 110, one or more response-input devices 120, one or more state sensors 122, a local voice subsystem 124, an event store 140, an object store 142, an interpretation subsystem 150, a state-estimation subsystem 152, a profile-generation subsystem 160, a compiler 170, and one or more target systems 180. A physical embodiment may further include an actuator system 190 and a local controller 192.

[0051] The components of system 100 may be implemented on one computing device or distributed across a wearable, mobile device, personal computer, embedded controller, robot, local server, and remote server. Communications may use wired or wireless connections and may be continuous, intermittent, or absent during an exposure. A local device may spool events and media for later transfer.

[0052] The source-capture subsystem 104 obtains a source-resource identity and optionally captures exact source bytes, normalized content, metadata, a transcript, one or more images, an audio or video object, a conversation branch, a physical configuration, a product label, a batch record, or another stimulus representation. The subsystem may calculate a cryptographic hash, perceptual hash, canonical identifier, or revision identifier.

[0053] The presentation subsystem 106 records presentation conditions such as a visible excerpt, viewport, page, scroll range, playback position, speed, volume, language, selected track, caption state, crop, resolution, display theme, font, recommendation position, surrounding items, conversation history, or interaction state. The presentation subsystem may receive events directly from an application or reconstruct them from screen, audio, or device capture.

[0054] The intervention subsystem 108 describes an intended physical, environmental, administered, or self-administered stimulus. In different embodiments, the intervention subsystem may generate a robot trajectory, device setting, environmental setting, administration schedule, or user instruction. The intervention subsystem may also obtain actual-delivery data from sensors or a user declaration.

[0055] The timing and event subsystem 110 provides identifiers, sequence numbers, monotonic timestamps, wall-clock estimates, synchronization observations, and timing uncertainty. Each recorder may issue a boot identifier and a monotonically increasing local event sequence. A server may preserve raw timing values and produce a corrected-time projection without modifying the source event.

[0056] The event store 140 receives append-only events. The object store 142 receives content-addressed media, source snapshots, and sensor traces. A normalized database projection may provide typed tables for efficient queries. An application role may be prevented from updating or deleting retained raw-evidence rows except through an explicit retention or erasure workflow.

[0057] The interpretation subsystem 150 produces versioned artifacts without replacing source evidence. A first analysis run may produce a transcript. A later run may resolve references in the transcript to stimulus anchors. Other runs may calculate acoustic features, subjective axes, domain concepts, embeddings, comparisons, or candidate causal attributions. Each run may identify inputs, implementation version, model version, prompt or template version, ontology version, parameters, timing, and confidence.

[0058] The state-estimation subsystem 152 estimates a baseline state, instantaneous state, response trajectory, or delayed state from explicit reports and sensor evidence. The subsystem may use filtering, smoothing, state-space models, hidden Markov models, Gaussian processes, Bayesian models, recurrent models, transformers, or other temporal models.

[0059] The profile-generation subsystem 160 selects evidence through a dataset snapshot and generates one or more personal-model artifacts. A profile snapshot may be immutable and may identify the evidence watermark, excluded observations, model version, feature schema, normalization, and objective scope.

[0060] The compiler 170 applies a profile snapshot to a current context, objective, and candidate set. The compiler may produce an output understandable by target system 180. The same profile may be compiled differently for a web search engine, local catalog, AI assistant, media service, robot, practitioner interface, product-selection system, or reconfigurable circuit.

Source and stimulus representation

[0061] Referring to FIG. 2, a source resource 202 may be associated with one or more stimulus snapshots 204. A snapshot 204 represents the source at a defined version or acquisition time. The snapshot may have one or more components 206 and relationships 208. For example, a video snapshot may include video, speech, music, caption, performer, and scene components; a conversation snapshot may include ordered turns and branches; a product snapshot may include a label, batch, composition, form, and package; and a physical-device snapshot may include a topology and component-state vector.

[0062] A snapshot may include source-declared metadata separately from computed metadata. A producer-provided product composition, a user-entered note, and a model-generated content tag may be stored in different namespaces with different provenance. This prevents a derived subjective characterization from being mistaken for an externally verified property.

[0063] An exposure episode 210 associates a snapshot 204 with a subject, context snapshot 212, intended protocol 214, measured presentation or delivery trace 216, and start and end conditions. A response packet 220 references the episode 210 and optionally one or more anchors 218. Interpretation artifacts 222 derive from the response packet 220 and associated stimulus data. Profile parameters 224 derive from a plurality of artifacts and evidence records.

[0064] In a digital embodiment, snapshot 204 may contain exact bytes or a hash and normalized representation. If exact source retention is not selected, the snapshot may contain a canonical locator, retrieval timestamp, revision, visible excerpt, strong hash when available, and enough context to distinguish a later source change.

[0065] In a physical embodiment, snapshot 204 may represent a device and proposed state, while measured trace 216 represents the state or energy actually delivered. A verification_mode field may distinguish direct sensor measurement, device readback, commanded-only state, user measurement, user estimate, user recall, and unknown delivery.

[0066] In a product embodiment, snapshot 204 may identify a product family and a particular batch. Separate snapshots may be used for products sharing a marketing or strain name but differing in producer, batch, composition, age, formulation, or package.

Stimulus anchors

[0067] An anchor 218 may use one or more coordinate systems. A media anchor may include time and frame ranges, a track, transcript span, or spatial region. A text anchor may include byte offsets, character offsets, a quotation, a quotation hash, a surrounding-context hash, a page, a paragraph, or a document-object-model selector. An image anchor may include a normalized box, polygon, or tracked object. A conversation anchor may include thread, branch, turn, and within-message span.

[0068] A physical anchor may include an anatomical region, laterality, coordinate on a body model, surface polygon, pose, orientation, depth estimate, applied contact region, and localization uncertainty. An applied-stimulus location and a referred-sensation location may be stored separately.

[0069] A transition anchor may reference a first component or state, a later component or state, and the interval of change. Transition anchors may identify a musical entrance, visual cut, onset of circuit bloom, crossing of a mechanical pain threshold, change in conversational tone, or onset of an administered effect.

[0070] Multiple anchors may be assigned to one response with respective probabilities. A user confirmation may select or modify an anchor while retaining earlier candidates. Anchors may be relocated after source reformatting using redundant location information.

Ambient voice-capture embodiment

[0071] Referring to FIG. 4, local voice subsystem 124 includes an audio-input interface 402, a playback-reference interface 404, an echo canceller 406, voice-activity detector 408, speaker detector 410, trigger detector 412, one or more ring buffers 414, state sampler 416, recorder 418, and local spool 420.

[0072] Audio-input interface 402 may receive one or more microphones. Playback-reference interface 404 may receive a digital copy of audio being presented by the system. Echo canceller 406 reduces classification of the presented audio as user speech. Speaker detector 410 may compare speech to a locally stored enrolled-speaker representation. In other embodiments, a close-field microphone, beamformer, turn-taking signal, or explicit wake phrase may be used.

[0073] Trigger detector 412 may detect a wake word, command, evaluative phrase, feedback intent, pain utterance, pleasure utterance, change report, or other speech pattern. Example phrases include “note that,” “remember this,” “more like this,” “less like this,” “right there,” “that hurts,” “that feels good,” or “something changed.” A trigger phrase may simultaneously constitute response evidence.

[0074] Ring buffers 414 retain a rolling interval of microphone audio and may also retain presented audio, video, application events, sensor measurements, physical-delivery measurements, and recent state estimates. Before a trigger, buffer contents may be overwritten. Upon a trigger, recorder 418 freezes at least part of the pre-trigger interval and continues recording until silence, a stop phrase, a maximum duration, or another termination condition.

[0075] State sampler 416 obtains an atomic or logically synchronized snapshot including the active source, stimulus snapshot, exposure episode, presentation state, physical state, body location, recent actions, context, baseline, current state estimate, and detector confidence. The record may distinguish probable utterance onset, wake-phrase onset, trigger-detection time, snapshot time, memo-content onset, and recording completion.

[0076] Local voice subsystem 124 may perform the monitoring and trigger operations without a remote language-model request. A local speech recognizer may optionally produce a preliminary transcript. Richer interpretation may occur locally or after encrypted transfer. Thus, the always-available monitor can operate with no per-token remote-model consumption until an event is committed for analysis.

[0077] The subsystem may expose several modes. In a voice-activated mode, the individual initiates capture. In a system-prompted mode, a prompt initiates capture. In a hybrid mode, both are available. In a deferred-recap mode, candidate moments are marked and presented later for annotation.

[0078] When the individual is participating in a conversation, a meta-commentary classifier may distinguish speech directed into the conversation from speech about the conversation. An explicit phrase such as “Recorder, note that” may open a feedback channel. The associated anchor may refer to one or more preceding conversational turns.

Prompted-feedback embodiment

[0079] Referring to FIG. 5, a prompt policy may receive interaction events, sensor features, model predictions, uncertainty values, exposure boundaries, and experiment objectives. The policy selects whether and when to prompt and may choose a prompt from stored audio, a text template, local speech synthesis, or a generated output.

[0080] A prompt event may be triggered by replay, pause, skip, dwell, abrupt termination, a detected state change, a disagreement between predicted and measured response, an uncertain pairwise comparison, or an expected reduction in model uncertainty. The prompt may request an overall evaluation, causal explanation, sensory location, comparison, desired direction, or distinction between content and presentation.

[0081] The exact prompt audio and metadata are stored as a stimulus component. Metadata may identify wording, voice, synthesis model, tone, pace, volume, policy version, trigger reason, and timing. A response packet may identify the original stimulus as a target while also identifying the prompt as an intervening stimulus.

[0082] The system may alternate prompt wording, tone, or timing to estimate prompt-induced bias. A model may condition interpretation on the prompt version. The system may also learn whether immediate prompting or deferred recap produces more consistent or useful evidence for the individual.

Response packet and thought-event embodiment

[0083] A response packet may group one or more voice recordings, text entries, ratings, comparisons, actions, body indications, and sensor-stream intervals that refer to a common exposure or anchor. The packet may include an objective, declared certainty, response latency, and immediate or later timing.

[0084] A transcript may be segmented into thought events. Example thought-event types include evaluation, sensory report, association, memory, comparison, prediction, desire, aversion, causal hypothesis, source interpretation, metacognition, and unresolved reaction. A thought event may include raw-audio references, transcript spans, candidate stimulus anchors, preceding and following state estimates, and attribution confidence.

[0085] The person's own response can itself become a stimulus. For example, speaking an evaluation aloud may strengthen, revise, or contradict the person's initial response. A prompt or displayed interpretation may likewise alter attention or state. The graph may represent recursive sequences in which source stimuli, system prompts, spoken thoughts, displayed interpretations, and subsequent responses are all ordered stimulus-response nodes.

[0086] A later user correction does not require replacement of the source voice recording or earlier interpretation. A new correction artifact may identify the earlier artifact, corrected concept, corrected polarity, corrected anchor, or personal-language rule. Dataset-generation policies may select the corrected artifact while retaining the historical chain.

Subjective axes and interpretation

[0087] Interpretation subsystem 150 may generate a sparse vector over general subjective axes. The axes may include valence, activation, engagement, intensity, novelty, comfort, pain, pleasure, relief, clarity, cognitive load, perceived credibility, agreement, usefulness, aesthetic fit, desire to continue, desire to repeat, reflective satisfaction, satiation, and certainty. A missing axis is distinguished from a neutral value.

[0088] Domain adapters may add axes. Text and conversation axes may include directness, warmth, concision, argument density, repetition, empathy, challenge, and precision. Media axes may include pacing, composition, framing, editing density, performance, narrative, groove, timbre, dynamics, melody, tension, vocal style, and production density.

[0089] Food axes may include aroma, sweetness, acidity, bitterness, heat, texture, richness, temperature, and aftertaste. Massage or palpation axes may include pressure, depth, pace, sharpness, guarding, referred sensation, release, mobility change, and delayed soreness. Adult-media axes may include attraction, arousal, pacing, tone, perceived chemistry, visual style, audio style, intensity, and desire to revisit. Amplifier axes may include attack, bloom, compression, brightness, fizz, tightness, string separation, touch response, and decay behavior.

[0090] A model may preserve an open-vocabulary phrase when no established concept fits. The artifact may contain the original phrase, evidence timing, an embedding, candidate concepts, and a review state. Later ontology versions may map the phrase to one or more concepts without removing the earlier artifact.

[0091] Semantic and acoustic interpretations may remain separate. An explicit statement may be assigned greater evidentiary weight than a conflicting prosodic estimate, or the conflict may trigger uncertainty or a clarification prompt. The system may learn individual language mappings, such as a recurring use of “interesting” to mean respected but not desired for repetition.

Dynamic personal-state embodiment

[0092] Referring to FIG. 6, a dynamic personal model may include relatively stable parameters \(\theta\), a slower disposition \(d_t\), an instantaneous state \(x_t\), a source stimulus representation \(s_t\), a measured presentation or delivery \(q_t\), a context \(c_t\), recent history \(h_t\), sensor observations \(y_t\), and explicit reports \(r_t\).

[0093] An example transition model is:

x_{t+1}=f_{\theta}\!\left(x_t,d_t,\phi(s_t,q_t),c_t,h_t\right)+\varepsilon_t.

The function may be implemented using a state-space model, Bayesian network, regression, neural network, rules, nearest-neighbor retrieval, or a hybrid. Different functions may be used for different domains or response timescales.

[0094] Sensor and report models may be represented as:

y_t=g_{\psi}(x_t)+\eta_t, \qquad r_t=h_{\omega}\!\left(x_{t-w:t},s_{t-w:t},L\right)+\zeta_t,

where \(L\) represents a personal language model or mapping. A report may summarize a state trajectory over a preceding window rather than only state at utterance time.

[0095] A response trajectory may identify onset latency, initial direction, peak magnitude, time to peak, adaptation, duration, rebound, aftereffect, and later evaluation. Different sampling schedules may be used for instantaneous digital stimuli and delayed administered stimuli.

[0096] Immediate attraction, enjoyment during exposure, desire to continue, immediate desire to repeat, later desire to repeat, and reflective satisfaction may be modeled as different outputs. A selected objective determines how the outputs contribute to a decision.

[0097] A counterfactual effect estimate may be represented as:

\Delta x_s(\tau)=x_{t+\tau}-\widehat{x}^{\,\mathrm{without}\;s}_{t+\tau}.

The estimated no-stimulus state may be obtained from a pre-exposure baseline, control period, repeated exposures, matched historical episodes, randomized order, or a learned trajectory. The system may label an effect as associative, quasi-experimental, or experimentally supported according to an evidence policy.

[0098] Carryover fields may identify exposure order, elapsed time, accumulated dose, residual state, mechanical load, tolerance, habituation, or model-selected exposure. Such fields may reduce attribution of a prior intervention's effect to a later stimulus.

Personal response profile

[0099] Referring to FIG. 10, a profile snapshot may include broad cross-domain sensitivities, modality parameters, domain attributes, objective-specific weights, contextual effects, source or product effects, interaction terms, exemplars, exceptions, embeddings, and time-indexed epochs.

[0100] A typed parameter may identify a scope, stimulus feature, response axis, effect estimate, uncertainty interval, evidence count, session count, stability value, dataset snapshot, and model version. For example, a parameter may estimate an association between a measured product component and relaxation under a defined baseline state, while explicitly identifying that the evidence is observational.

[0101] A parameter may be conditional on several features. Examples include pressure at a particular body region, a product under a particular objective, direct AI responses during coding, or slow pacing under a selected mood. Contradictory evidence in different contexts may generate separate parameters instead of one average.

[0102] Embeddings may represent source content, physical or chemical traces, spoken reaction language, state trajectories, or a personal preference space. Each embedding identifies its input and model. A compiler may use an exemplar or embedding directly even if no stable human-readable concept is available.

[0103] Cross-domain relationships may have learned transfer coefficients and uncertainty. A preference for slow musical development may weakly predict a preference for slow narrative development while having no established relationship to preferred email length. Transfer may be enabled only when supported by held-out performance or another criterion.

[0104] Profile snapshots may be generated for a historical date, recent time window, stable-preference view, named context, named objective, or particular application. A current-profile pointer may refer to a snapshot without deleting older snapshots.

Profile compiler and decision subsystem

[0105] Referring to FIG. 11, compiler 170 receives a scoped profile snapshot, current context or state, objective, target-system capabilities, and optionally a candidate set. It produces a target artifact and a decision trace identifying contributing parameters, model versions, and uncertainty.

[0106] For web search, the compiler may translate high-weight positive concepts, negative concepts, and source constraints into one or more editable query strings. It may produce narrow, balanced, and exploratory variants. A richer personal vector may be used locally to rerank results without transmitting the full vector to the search provider.

[0107] For an AI system, the compiler may generate a purpose-specific preference context, response-style parameters, examples, desired challenge level, preferred length, explanation density, or negative instructions. Different projections may be generated for coding, research, brainstorming, or another conversational objective.

[0108] For content generation, the compiler may produce a creative brief including tone, pacing, structure, sensory attributes, positive exemplars, and negative exemplars. For a practitioner, it may produce a natural-language response brief. For a physical system, it may produce a requested protocol subject to local capability validation.

[0109] A user may respond to the compiled artifact independently of the candidate stimuli it produces. For example, the user may state that a query overemphasized narrative or that an AI preference context made answers too terse. Such feedback trains the compiler mapping without necessarily changing the underlying stimulus preference.

[0110] Referring to FIG. 12, a multi-objective decision may combine expected response axes with price, availability, time, transition cost, uncertainty, and information value. An example utility is:

U(a\mid x,c,o)= \sum_k w_k(o)\,\mathbb{E}[R_k\mid a,x,c] -\lambda_p P(a)-\lambda_a A(a)-\lambda_t T(a) -\lambda_u \mathcal{U}(a)+\lambda_e \mathcal{I}(a).

The system may instead return a Pareto frontier without collapsing the dimensions. Price and availability observations may be time stamped and source specific so a historical decision can be reconstructed after market conditions change.

Physical-intervention embodiment

[0111] Referring to FIG. 7, a physical-intervention embodiment includes planner 702, capability validator 704, real-time controller 706, actuator 708, delivery sensors 710, location subsystem 712, response subsystem 714, and fast local response path 716.

[0112] Planner 702 proposes an intervention based on an objective, profile, current state, and candidate set. Capability validator 704 applies device, anatomical, force, speed, travel, temperature, energy, or other limits and may reject or modify the proposal. Real-time controller 706 controls actuator 708 using delivery sensors 710. Fast local response path 716 can reduce or stop actuation without waiting for a remote model.

[0113] A protocol may specify body location, pose, end-effector identity, shape, material, contact area, force, pressure, displacement, speed, dwell time, direction, path, vibration, temperature, ramp-up, release, and stopping conditions. The measured trace may include actual position, force, torque, pressure, displacement, compliance, velocity, acceleration, contact status, slip, saturation, and sensor uncertainty.

[0114] Location subsystem 712 may use a body model, image, depth sensor, landmarks, robot coordinates, wearable markers, user-selected map location, or combinations thereof. The system may transform coordinates among robot, camera, body, and session frames and store transformation uncertainty.

[0115] Referring to FIG. 8, a user utterance such as “right there” may be aligned with the pre-trigger force and position trace. The interpretation subsystem may extract sharp onset, delayed release, pain, pleasure, relief, preferred depth, or referred sensation. A later proposal may vary location, force, contact geometry, timing, or sequence within the capability validator's permitted space.

[0116] One evidence corpus may support different objective schemas, including pleasure, relief, mobility, pain-boundary mapping, response reproducibility, or diagnostic information gain. A diagnostic planner may choose a next permitted location or force intended to discriminate among hypotheses, while a massage planner may choose a trajectory intended to maximize predicted relief or pleasure.

[0117] In a reconfigurable electronic embodiment, actuator 708 may include switches, relays, analog controls, digital controls, or other state-changing components. The measured stimulus may include an observed circuit topology, component value, signal measurement, or output audio. A user's response is associated with the observed physical state rather than a requested preset alone.

Administered and self-administered embodiment

[0118] Referring to FIG. 9, an administered-stimulus episode may include product identification 902, baseline assessment 904, administration record 906, onset monitoring 908, peak monitoring 910, decline monitoring 912, and later assessment 914.

[0119] Delivery confidence may be classified as direct-sensor measurement, connected-device report, package scan, user measurement, user estimate, user recall, or unknown. A dose distribution may be stored instead of a point estimate.

[0120] A product snapshot may include vendor, SKU, producer, product family, batch, lot, form, composition, package quantity, dates, laboratory-report reference, price, discount, and acquisition location. A purchase observation and an availability observation may be separate time-stamped objects.

[0121] A monitoring schedule may be fixed or may adapt to a product form, route, previous onset curve, dose estimate, or individual history. A local audio prompt can ask for changes without requiring the user to operate a screen. The exact prompt remains part of the episode record.

Cannabis-product example

[0122] In one nonlimiting example, the product is a cannabis product. The source resource identifies a dispensary listing or package, while the stimulus snapshot identifies a producer, labeled strain, particular batch, product form, cannabinoid information, terpene information, package date, and other available composition data. Products sharing a strain label may remain separate snapshots.

[0123] An administration record may identify route, declared amount, estimated active amount, confidence, and time. Context may identify recent use, food, hydration, caffeine, other substances, starting relaxation, anxiety, pain, focus, fatigue, mood, appetite, setting, intended activity, and objective.

[0124] Response assessments may estimate onset, relaxation, anxiety, pain relief, focus, distractibility, sociability, sensory change, appetite, bodily heaviness, energy, fatigue, mental clarity, pleasure, duration, aftereffect, and immediate or later desire to repeat. A response axis may be omitted when not assessed.

[0125] A product-selection compiler may identify candidates predicted to satisfy an objective such as evening relaxation, pain relief with mental clarity, or best value among currently available products. The output may show predicted responses, undesired effects, price, availability, and confidence separately or under a selected objective function.

Digital-content and conversational embodiments

[0126] Referring to FIG. 13, a text embodiment records an exact article, message, email, AI response, or conversation branch and the portion visible to the user. A voice reaction may be anchored to a paragraph, phrase, sender, turn, or transition. Derived parameters may concern concision, argument density, warmth, directness, novelty, evidence, repetition, challenge, or another feature.

[0127] An AI-conversation embodiment records model identity, application settings, system context when available, branch identity, visible history, and turn range. A response such as “use this level of detail but stop repeating the caveat” may generate separate positive and negative parameters. A later compiler may provide those parameters to another AI system without providing the entire conversation history.

[0128] A media embodiment records a source version, playback timeline, tracks, transforms, user interaction, and one or more time or spatial anchors. A response such as “more like the transition at two minutes” can become an exemplar even before the system has named the controlling media feature.

[0129] An adult-media embodiment may associate timestamped reactions with pacing, tone, visual style, audio style, perceived chemistry, intensity, attraction, arousal, comfort, or desire to revisit. A compiler may create visible query variants and may privately rerank retrieved candidates using richer profile features than are transmitted in a public query.

[0130] A food embodiment may associate a response with a dish, component, bite, course, ingredient, preparation, temperature, texture, venue, or context. A compiler may produce recipe modifications, restaurant queries, or product comparisons using preference, price, distance, and availability.

Active-learning and experiment embodiments

[0131] Referring to FIG. 14, an experiment planner may select a next stimulus or pair based on expected preference, uncertainty, novelty, diversity, context relevance, transition cost, and expected information gain. A candidate may be a digital item, product, physical location, force trajectory, device state, or configuration.

[0132] The planner may deliberately repeat a reference stimulus, reverse presentation order, vary one feature, choose a nearby physical location, or select a product with a composition that discriminates among competing personal-model hypotheses. The proposal stores the candidate set, exclusions, acquisition values, model version, dataset snapshot, objective, and random seed.

[0133] For a physical system, the proposal is passed through local capability validator 704. For a self-administered system, the proposal may be displayed as an option rather than directly executed. The resulting exposure and response are linked to the proposal to measure whether the expected information or benefit was obtained.

Data provenance, versioning, and reprocessing

[0134] Referring again to FIG. 2, each derived entity may be linked by provenance relations such as used, generated_by, derived_from, supersedes, invalidated_by, or attributed_to. A decision explanation may traverse from the decision through profile parameters, model and dataset versions, interpretation artifacts, response packets, measured exposures, and source snapshots.

[0135] Dataset snapshots may identify a selection query, event watermark, included observation identifiers, exclusions and reasons, context policy, missing-data policy, train-test assignment, feature versions, ontology versions, and a manifest hash. A model version may identify algorithm, hyperparameters, code digest, container digest, seed, training metrics, calibration metrics, and artifact hash.

[0136] Reanalysis does not require modification of a historical transcript, vector, state estimate, or profile. A new artifact may supersede an older artifact for ordinary display while preserving both. External model services that cannot be reproduced bit-for-bit may be recorded with full replayable inputs and provider metadata.

Alternative implementations

[0137] Referring to FIG. 15, a local implementation may perform source capture, trigger detection, transcription, state estimation, profile generation, and compilation on one device. A distributed implementation may perform wake detection and buffering locally while performing semantic analysis and model training on a server. Another implementation may keep raw evidence local and transmit only selected derived parameters.

[0138] The event and object stores may use relational databases, document databases, graph databases, time-series databases, files, object stores, secure enclaves, or combinations thereof. Content addressing may use SHA-256 or another hash. Canonical serialization may use JSON, CBOR, protocol buffers, or another structured representation.

[0139] Sensors may be integrated in a phone, wearable, room, vehicle, robot, instrument, appliance, computer, or medical device. The response interface may use speech, text, buttons, gestures, gaze, body maps, neural signals, or combinations thereof.

[0140] A profile may be generated for one individual, multiple identified individuals, or a group while retaining subject identity for each response. Cross-user information may provide an initial prior, but personal evidence may update or override the prior. A profile can be exported as typed parameters, vectors, exemplars, or a compiled policy.

[0141] A stimulus may be generated by the same system that observes the response. For example, an AI system may generate text, speech, images, music, or a robot trajectory and then acquire a reaction. The generator version, prompt, seed, and output snapshot may be retained to enable later reproduction or variation.

[0142] An embodiment may operate prospectively by selecting stimuli, retrospectively by importing historical content and user annotations, or both. Historical email, chat, media, purchase, or device logs may be imported as source snapshots with available timing confidence.

[0143] A response may be associated with a stimulus immediately, after a delay, or during later review. A later memo may reference a prior episode by search, voice description, image, timeline, product package, body map, or exemplar. The resulting anchor may include lower confidence than a contemporaneous anchor.

[0144] The system may output a normalized score, probability distribution, predicted trajectory, explanation, ranked list, Pareto frontier, or experiment proposal. A normalized preference score may remain separate from confidence, evidence count, recency, price, and availability.

[0145] Presently contemplated practical implementation uses a local microphone front end with playback-reference echo cancellation, local voice-activity and trigger detection, several seconds of ring-buffer pre-roll, an append-only event journal, content-addressed media objects, a relational evidence store, versioned analysis workers, and a profile compiler. Physical embodiments additionally use a local deterministic capability validator and real-time controller that remain operational without a remote analysis service.

[0146] The foregoing embodiments demonstrate that the disclosed architecture is not limited to a particular stimulus modality, response sensor, subjective axis, machine-learning technique, database, search provider, product category, robot, medical task, entertainment category, or reconfigurable device. The architecture may be practiced with fewer than all described components and may be extended with later sensors, ontologies, models, or compilers while retaining historical evidence and provenance.

ABSTRACT OF THE DISCLOSURE

A system associates an individual's responses with digital or physical stimuli. A stimulus snapshot identifies content, a product, a device state, or an intended intervention. A presentation or delivery trace represents what was actually presented or applied. A response subsystem captures speech, ratings, behavior, physiology, sensory reports, or delayed assessments and associates the response with a temporal, content, anatomical, or transition anchor. Versioned analysis produces semantic, acoustic, perceptual, affective, and state artifacts while preserving source evidence. A personal model estimates context-dependent response trajectories and profile parameters from multiple exposures. A compiler applies a selected profile and objective to generate a search query, ranking, AI interaction context, recommendation, experiment, device configuration, or physical intervention proposal. Local voice triggering may freeze pre-trigger stimulus history without remote model use. Physical embodiments record measured delivery and subject later proposals to a local capability validator.