HumanFB: Personal Stimulus–Intervention–Response Architecture

Continuous multimodal capture, subjective-state modeling, reusable personal profiles, and closed-loop selection

Project: HumanFB
Document date: August 15, 2026
Status: Consolidated and expanded concept architecture
Scope: Digital stimuli, physical interventions, self-administered exposures, spoken thought capture, subjective-state estimation, preference learning, and multi-objective recommendation

Executive summary

HumanFB is a general system for recording what an individual experiences, how the experience is actually delivered, what the individual thinks, says, feels, or does in response, and what versioned models infer from those observations. It associates those response records with the exact stimulus, intervention, location, segment, context, and moment that elicited them. Over time, it builds a reusable, parameterized model of the individual's subjective responses rather than a collection of disconnected ratings.

The system began conceptually with Labamp: a reconfigurable analog amplifier whose exact physical circuit state could be paired with a player's rating or spoken reaction. The same architecture applies much more broadly. The stimulus may be a video, song, article, email, AI response, human conversation, meal, fragrance, room, robotic massage trajectory, diagnostic palpation, exercise, self-applied pressure, cannabis product, or another physical or chemical exposure. The response may be an explicit rating, comparison, voice memo, naturally spoken thought, sensory report, observed behavior, physiological signal, or later reflective assessment.

The broadest useful data structure is a Personal Stimulus–Intervention–Response Graph. The graph preserves several distinctions that ordinary recommendation systems collapse:

• A source is not the same as the exact version or segment presented.
• A requested physical intervention is not the same as the stimulus actually delivered.
• A measured signal is not the same as an inferred subjective state.
• A spoken thought is not the same as an AI interpretation of that thought.
• A momentary reaction is not the same as a stable preference.
• Immediate attraction or relief is not the same as later satisfaction.
• Preference is not the same as price, availability, time, uncertainty, or other decision criteria.

HumanFB therefore stores raw observations and measured stimulus traces as immutable evidence while retained. Transcripts, affective estimates, perceptual axes, concept mappings, preference parameters, predictions, and recommendations are versioned derived artifacts. New models can reinterpret old observations without overwriting what the user originally experienced or expressed.

The system can operate with very little interface ceremony. A local, always-available audio front end can run voice-activity and wake-intent detection without cloud tokens. The user can say “note that,” “more like this,” “right there,” or simply begin an evaluative memo. The trigger freezes several seconds of stimulus history, samples the system and estimated user state, and records until silence or a stop phrase. Alternatively, an AI or local policy can ask a short audio question when it detects an informative moment. Because the prompt itself changes the experience, it is also recorded as a stimulus.

For physical systems, HumanFB adds an intervention layer. A robot masseur or palpation system records commanded position and force, permitted trajectory, measured force and displacement, body location, contact geometry, timing, and response. A self-experiment logger records the identity, batch, dose estimate, administration time, and context of a product such as food, caffeine, or cannabis, then gathers repeated state assessments as effects develop. The same evidence framework supports pleasure, relief, pain, diagnostic information, sensory qualities, delayed effects, and user-defined objectives.

The ultimate output is not merely “the user likes X.” It is a personal state-transition model capable of statements such as:

Given this individual's current disposition, objective, recent history, and context, a stimulus with these measured features usually produces this immediate response trajectory, these delayed effects, and this later evaluation, with this uncertainty and these acquisition costs.

That model can compile into a web-search query, private reranker, AI preference context, inbox priority rule, creative brief, recipe variation, product shortlist, massage program, diagnostic experiment, or validated physical configuration. The profile remains a personal substrate; downstream applications receive only the projection they need.

1. Conceptual architecture

HumanFB contains four related systems that should remain separable.

The runtime loop is:

current disposition and context
            +
selected stimulus or intervention
            ↓
actual presentation or measured delivery
            ↓
sensory, emotional, cognitive, behavioral, and physiological response
            ↓
spoken thought, explicit judgment, and optional passive observations
            ↓
versioned interpretation and causal attribution
            ↓
personal model update
            ↓
query, ranking, prompt, product choice, or next intervention
            ↓
new exposure and outcome

1.1 Descriptive, predictive, and prescriptive layers

HumanFB can be useful at three levels:

1. Descriptive: Preserve what happened and make it searchable. “Show every time I described a sound as brittle.”
2. Predictive: Estimate likely responses. “Would this product probably relax me without producing much mental fog?”
3. Prescriptive: Select an experience or experiment. “Which available option best fits tonight's objective?” or “Where should the robot probe next to clarify the pain boundary?”

These layers can mature independently. A reliable descriptive record is valuable before the predictive model is sophisticated, and a predictive model should be validated before it directs physical interventions.

1.2 Evidence and interpretation

The controlling hierarchy is:

raw source or physical state
        ↓
presentation or delivery measurement
        ↓
raw response observation
        ↓
interpreted response axes and concepts
        ↓
profile parameters
        ↓
prediction or decision

Every downward step adds assumptions. The database must preserve the ability to traverse upward from a recommendation to the evidence and versions that produced it.

2. General stimulus and intervention model

HumanFB uses stimulus as the broad category: anything presented to or applied to the individual that may contribute to a response. An intervention is a stimulus whose delivery can be deliberately specified, measured, or controlled.

2.1 Digitally presented stimuli

• Text: articles, books, documents, email, messages, AI responses, human chats, captions, code, and interface copy.
• Images: photographs, illustrations, diagrams, advertisements, products, and interface states.
• Audio: music, voice, podcasts, soundscapes, prompts, alerts, and generated tones.
• Video: film, performance, tutorials, adult media, advertisements, and recorded interactions.
• Interactive experiences: websites, applications, games, AI conversations, branching narratives, and generative sessions.

2.2 Physical and environmental stimuli

• Mechanical pressure, palpation, vibration, stretch, motion, impact, and haptic patterns.
• Heat, cold, airflow, humidity, and environmental temperature.
• Light intensity, color, temporal modulation, and spatial distribution.
• Sound pressure, spectral content, direction, and room acoustics.
• Fragrance, flavor, texture, and other chemical or sensory exposures.
• Furniture geometry, posture, clothing, bedding, and spatial environments.

2.3 Ingested or administered stimuli

• Food and drink.
• Caffeine, alcohol, cannabis, supplements, and medication.
• Products with delayed, dose-dependent, and multi-phase effects.
• Combinations whose interactions and recent-use history may affect response.

2.4 Self-generated and self-administered stimuli

• Self-massage, stretching, exercise, posture changes, and breathing patterns.
• Substance ingestion or product use reported by the user.
• Memories, imagery, internal rehearsal, and self-talk.
• The user's own spoken reaction, which becomes an audible stimulus and may change subsequent thought.

2.5 Composite and interactive stimuli

Real experiences normally contain several simultaneous components. A video contains images, speech, music, editing, performers, captions, and narrative. A massage contains pressure, movement, temperature, body position, expectation, and conversation. A meal includes courses, ingredients, aromas, environment, company, and time.

HumanFB represents the experience as a bundle of synchronized, independently addressable stimulus components. Reactions may attach to the whole bundle, one component, a relationship between components, or a transition.

3. Source, intended stimulus, and measured delivery

The architecture distinguishes four objects that are often mistakenly treated as one.

3.1 Source resource

source_resource identifies an external, physical, or catalog origin:

• URL, file, document, message, thread, or conversation branch;
• media catalog item and edition;
• product SKU, batch, package, or laboratory report;
• recipe, venue, room, or physical object;
• device, robot, end effector, or Labamp hardware manifest.

3.2 Stimulus snapshot

stimulus_snapshot is an immutable version of the content or specified physical state at a particular time. Depending on the modality and capture policy, it may contain exact bytes, normalized text, media-object references, a transcript, structured turns, product composition, or a canonical physical-state vector.

Important fields include:

• source identifier and retrieval time;
• source revision or edition;
• canonical and perceptual content hashes;
• source-declared metadata;
• computed metadata with separate model provenance;
• components and relationships;
• availability and acquisition observations.

3.3 Intended presentation or delivery protocol

The intended object records what should happen:

• media start point, volume, speed, display transform, or visible passage;
• robot path, force, pressure, speed, dwell time, and body target;
• nominal product, dose, route, and administration schedule;
• requested circuit configuration or environmental setting.

3.4 Measured presentation or delivered stimulus

The measured object records what the individual probably received:

• actual playback intervals, volume, dropped frames, scroll ranges, and visible content;
• force–torque, pressure, contact area, displacement, speed, temperature, and vibration traces;
• observed device state rather than commanded state alone;
• measured dose or a confidence-qualified user estimate;
• interruptions, deviations, and delivery failures.

For physical intervention, the primary relationship is:

requested intervention
        -> capability-checked permitted intervention
        -> controller command
        -> measured delivered stimulus
        -> response trajectory

The difference between requested and delivered stimulus is itself useful evidence.

4. Identity, time, and synchronization

HumanFB correlates multiple streams, sometimes across several devices. Identity and timing are therefore foundational rather than administrative details.

4.1 Core identities

Use time-sortable UUIDs such as UUIDv7 while preserving explicit ordering fields.

4.2 Multiple clocks

Each recorder stores:

• device monotonic time;
• device wall-clock estimate;
• server ingest time;
• synchronization offset and uncertainty;
• boot identity and local sequence number.

Raw timestamps are never retroactively rewritten. Improved clock alignment produces a derived corrected-time projection with uncertainty.

4.3 Latency-aware alignment

Different response channels have different latency distributions. A withdrawal movement may be nearly immediate. Speech often refers to several preceding seconds. EDA may lag stimulus onset. A next-morning evaluation concerns a much larger episode.

The attribution system stores both observed time and estimated causal window:

response observed at t
candidate eliciting interval = [t - latency_max, t - latency_min]

Latency models are versioned by sensor, response type, and eventually by individual.

5. Anchoring reactions to stimuli and bodies

A response is substantially more useful when it identifies what it concerns.

5.1 Digital anchors

Use redundant coordinates. Exact offsets are precise but fragile; a quote plus surrounding-context hash can survive later formatting changes.

5.2 Body and physical anchors

Physical systems need a versioned body-coordinate model:

• anatomical region and laterality;
• surface coordinate or polygon on a body mesh;
• orientation and body pose;
• end-effector contact shape and area;
• estimated depth or tissue layer when available;
• reference landmarks and localization uncertainty;
• referred-sensation location distinct from applied-stimulus location.

For self-administered stimuli, body location may come from touch input on a body diagram, camera tracking, wearable location, or spoken description.

5.3 Transition anchors

Some reactions concern change rather than a stable segment:

• the moment a guitar tone blooms;
• a video cut or musical entrance;
• pressure crossing a pain threshold;
• movement from one body region to another;
• the onset or peak of an ingested effect;
• a conversational shift in tone.

transition_anchor references the prior state, later state, and transition interval.

6. Context and baseline state

The same stimulus can produce different effects depending on the starting condition. HumanFB records context and estimates a pre-exposure baseline.

6.1 Context snapshot

Possible context dimensions include:

• current objective or task;
• time, circadian phase, and recent sleep;
• location and environment;
• social setting;
• posture and activity;
• recent food, caffeine, medication, cannabis, exercise, or other exposures;
• device, renderer, speaker, display, robot, or end effector;
• prior stimuli in the session;
• self-reported mood, pain, hunger, fatigue, stress, libido, curiosity, or focus;
• sensor-quality and missing-data indicators.

Context schemas are extensible. Unknown must remain distinguishable from a measured neutral value.

6.2 Timescales of personal state

HumanFB separates:

1. Long-lived tendencies: comparatively stable sensitivities and preferences.
2. Disposition: state lasting minutes to hours, such as fatigue, hunger, anxiety, pain, or openness to novelty.
3. Instantaneous state: rapidly changing attention, valence, arousal, pleasure, tension, confusion, or sensory intensity.
4. Response trajectory: change during and after the exposure.
5. Reflective state: later judgment of whether the experience was worthwhile or desirable to repeat.

6.3 Baseline windows

For a controlled episode, record:

pre-exposure baseline
        -> onset response
        -> within-exposure trajectory
        -> immediate post-exposure state
        -> delayed follow-up state

The baseline may be a short sensor window, a spoken self-assessment, a previous stable period, or a model estimate with stated uncertainty.

7. Ambient voice capture and natural interaction

Voice is the principal low-friction channel for externalized thought. The user should not need to hold a button whenever a reaction occurs.

7.1 Token-free local monitoring

The always-available local path is:

microphone
  -> playback-reference echo cancellation
  -> local voice-activity detection
  -> local enrolled-speaker and wake-intent detection
  -> rolling audio buffer
  -> trigger
  -> freeze pre-roll and sample system state
  -> record until silence or stop intent
  -> encrypted local spool
  -> optional local or later server interpretation

This path uses DSP, a wake-word engine, or a small local model. It requires compute and power but no paid language-model tokens and no network request until a memo is committed or analyzed.

7.2 Natural trigger vocabulary

The initial implementation should support reliable explicit invocations:

• “Note that.”
• “Remember this.”
• “Feedback.”
• “More like this.”
• “Less like this.”
• “Right there.”
• “That hurts.”
• “That feels good.”
• “Something changed.”

The utterance may both trigger capture and carry semantic content. Later versions can recognize unconstrained evaluative speech without a fixed wake phrase.

7.3 Pre-roll

The local device continuously maintains short ring buffers for:

• user microphone audio;
• stimulus audio/video or application-state history;
• sensor signals;
• recent interaction events;
• state-estimate history.

On activation, it freezes the relevant seconds before the detected phrase. This allows “that” and “right there” to refer to an event that happened before the recorder knew a memo was beginning.

7.4 Atomic trigger snapshot

Activation records:

• probable utterance onset;
• trigger phrase onset and detection time;
• active source, stimulus snapshot, presentation, and anchors;
• playback position, visible range, robot position, body target, or product episode;
• recent pause, rewind, scroll, force, movement, or configuration changes;
• current context and recent state trajectory;
• detector and firmware versions;
• trigger confidence and alternative trigger interpretations.

7.5 Distinguishing the user from playback

For video, podcasts, calls, and voice conversations, the capture front end can combine:

• a direct playback reference for acoustic echo cancellation;
• directional or close-field microphones;
• locally enrolled speaker embeddings;
• turn-taking metadata;
• wake phrases restricted to the enrolled speaker;
• a classifier distinguishing conversation participation from commentary about the conversation.

An explicit meta-channel such as “Recorder, note that…” remains the reliable fallback.

8. AI-initiated audio prompts

The system may request a reaction instead of waiting for the user to initiate one.

8.1 Trigger policies

Candidate prompt triggers include:

• a rewind, replay, pause, skip, or abrupt abandonment;
• a large physiological or expressive change;
• sustained attention or unusual dwell time;
• a divergence between predicted and observed response;
• an uncertain comparison;
• an active-learning experiment needing clarification;
• a natural boundary such as the end of a track, paragraph, scene, course, massage phase, or product-effect interval.

8.2 Local prompt generation

Prompts need not consume language-model tokens. A local policy can select prerecorded audio or local text-to-speech templates:

• “What changed?”
• “What caught your attention?”
• “Was that pleasant, unpleasant, or simply intense?”
• “What would you want more or less of?”
• “Was your reaction about the content or how it was presented?”
• “How does that compare with the previous one?”
• “Where do you feel it?”

8.3 Prompt-as-stimulus recursion

The audio prompt changes attention and may change state. It is therefore recorded as an intervening stimulus with:

• exact text or audio hash;
• voice and synthesis model;
• tone and pacing parameters;
• playback level;
• selection policy and trigger reason;
• timing relative to the original stimulus;
• answer, nonanswer, interruption, or dismissal.

The sequence becomes:

original stimulus
  -> preliminary response
  -> system prompt
  -> response to prompt
  -> spoken explanation

Prompt wording and tone can be varied and logged to estimate prompt-induced bias.

8.4 Interaction modes

The hybrid mode is the recommended default architecture.

9. Response evidence and thought events

9.1 Response packet

A response_packet groups reactions to one exposure, anchor, transition, or prompt. It may contain:

• raw voice memo;
• typed comment or annotation;
• absolute ordinal rating;
• pairwise preference;
• save, skip, repeat, continue, stop, more, or less action;
• body-location indication;
• explicit sensory or emotional labels;
• declared certainty;
• physiological and behavioral stream references;
• immediate and later follow-up assessments.

9.2 Externalized thought events

Continuous or triggered speech can be segmented into thought_event objects such as:

• evaluation;
• sensory report;
• association or memory;
• comparison;
• prediction;
• desire or aversion;
• causal hypothesis;
• interpretation of another person or source;
• metacognitive explanation;
• unresolved reaction.

Example:

{
  "thought_event_id": "thought_01J...",
  "response_packet_id": "response_01J...",
  "onset_monotonic_us": 918200100,
  "duration_ms": 1830,
  "raw_audio_ref": "media_01J...",
  "transcript_artifact_ref": "artifact_01J...",
  "thought_types": ["SENSORY_REPORT", "COMPARISON"],
  "stimulus_anchor_ids": ["anchor_01J..."],
  "state_before_id": "state_01J...",
  "state_after_id": "state_01J...",
  "attribution_confidence": 0.74
}

In the more speculative thought experiment, a future direct neural or cognitive measurement is simply another raw observation stream. The data model does not need to pretend that current microphones or biosensors literally read unexpressed thoughts.

9.3 Measurement channels

Potential channels include:

• voice content, timing, and prosody;
• facial movement and facial EMG;
• gaze and pupil response;
• posture, gesture, withdrawal, and movement;
• EDA, heart rate, HRV, respiration, temperature, and peripheral blood flow;
• muscle activity and mechanical guarding;
• EEG or other neurophysiological signals;
• interaction behavior such as replay, dwell, scroll, selection, and abandonment;
• explicit ratings and comparisons.

Each channel stores sensor identity, calibration, sampling rate, clock quality, missingness, artifact estimates, and processing version.

10. Interpretation and data-vector abstraction

HumanFB never produces one monolithic “emotion score.” It creates separable versioned artifacts.

10.1 Literal artifacts

• raw audio or text;
• transcript with word timing;
• language and speaker segmentation;
• references resolved to stimulus anchors;
• evidence spans supporting extracted concepts.

10.2 Acoustic and behavioral artifacts

• speaking rate and pause structure;
• pitch and intensity distributions;
• timing relative to stimulus events;
• movement, gaze, withdrawal, replay, and dwell summaries;
• signal-quality and artifact estimates.

10.3 General subjective axes

Missing axes remain missing rather than being forced to zero.

10.4 Domain-specific axes

10.5 Open-vocabulary concepts

Every interpretation retains phrases that do not fit the current ontology. Store the original wording, evidence timing, embeddings, candidate mappings, user correction, and later ontology relationships. Personal vocabulary can become first-class profile parameters after repeated evidence.

11. Dynamic subjective-state model

HumanFB models the individual as a changing system rather than a static set of tastes.

Let:

• \(\theta\) represent relatively stable personal parameters;
• \(d_t\) represent slower disposition at time \(t\);
• \(x_t\) represent instantaneous subjective state;
• \(s_t\) represent the source stimulus;
• \(q_t\) represent the measured presentation or physical delivery;
• \(c_t\) represent context;
• \(h_t\) represent recent exposure history;
• \(y_t\) represent observed sensor signals;
• \(r_t\) represent explicit reports and spoken thoughts.

The state transition can be represented as:

Sensors and reports are observations of that latent state:

The system can therefore estimate both current state and how a stimulus changes it.

11.1 Immediate and delayed trajectories

For each response axis, the model may estimate:

• onset latency;
• initial direction and magnitude;
• peak value and time to peak;
• adaptation or habituation;
• duration;
• rebound or aftereffect;
• next-session or next-day evaluation.

11.2 Wanting, liking, and later satisfaction

Keep distinct prediction targets for:

• immediate attraction or craving;
• enjoyment during exposure;
• desire to continue;
• immediate desire to repeat;
• later desire to repeat;
• reflective satisfaction;
• fatigue, regret, pain, fog, or other delayed effects.

Different objective schemas can weight these targets differently.

11.3 Cross-domain response dimensions

The system may discover dimensions that transfer across domains: directness, warmth, novelty, slow buildup, high contrast, information density, rhythmic regularity, or visual simplicity. Transfer must be learned rather than assumed.

A profile parameter can state:

feature: slow buildup
source domain: music
target domain: narrative video
transfer coefficient: 0.43
credible interval: [0.11, 0.68]
supporting evidence: 27 paired observations

12. Temporal credit assignment and causal estimation

The central analytical difficulty is determining what produced a response when stimuli overlap and responses lag.

12.1 Candidate causes

A response may have been evoked by:

• the current content;
• something several seconds earlier;
• a transition between stimulus components;
• the user's own thought or memory;
• an AI prompt;
• accumulated exposure or fatigue;
• a contextual event not represented in the primary media;
• interaction among several components.

12.2 Attribution evidence

Attribution can use:

• explicit language such as “that sentence” or “when the bass entered”;
• gaze and selected region;
• exact robot position and force at utterance onset;
• pause, replay, rewind, skip, or touch behavior;
• latency models for each response channel;
• repeated exposures and reversed presentation order;
• controlled A/B comparisons;
• similar stimuli differing in one measured feature;
• user confirmation of candidate anchors.

Every attribution stores alternatives and confidence rather than only one asserted cause.

12.3 Counterfactual effect

The desired quantity is the change caused by the stimulus relative to what would probably have happened without it:

The counterfactual is estimated from baseline trajectories, controls, repetitions, matched historical episodes, randomization, and model uncertainty. HumanFB should label ordinary association separately from experimentally supported causal effect.

12.4 Carryover and sequence effects

Physical, emotional, and chemical interventions may affect later episodes. Record:

• prior exposure order;
• elapsed time and washout estimate;
• accumulated dose or mechanical load;
• tolerance or habituation;
• residual state;
• whether the later stimulus was selected because of the earlier response.

Without these fields, the system may attribute a carryover effect to the wrong item.

13. Robotic palpation and massage adapter

A robot masseur or physician-assistant palpation system is a particularly clear physical embodiment of HumanFB.

13.1 Intervention descriptor

The intended protocol includes:

• body location and localization uncertainty;
• pose and joint configuration;
• end-effector identity, shape, material, and contact area;
• force, pressure, displacement, speed, acceleration, and dwell targets;
• direction, path, and oscillation;
• temperature and vibration when relevant;
• ramp-up and release profiles;
• objective and stopping conditions.

13.2 Measured delivery

Record synchronized traces for:

• actual pose and end-effector coordinates;
• force and torque;
• calculated pressure and contact area;
• surface displacement and inferred compliance;
• speed, acceleration, and vibration;
• temperature;
• controller saturation, slip, and contact loss;
• calibration and sensor uncertainty.

13.3 Response channels

The user may respond through:

• “right there,” “too sharp,” “deeper,” “that releases it,” or other speech;
• explicit pain, pleasure, or relief ratings;
• withdrawal, guarding, muscle activation, or relaxation;
• facial response, respiration, HRV, EDA, or temperature;
• body-map indication of applied and referred sensation;
• range-of-motion or later-soreness assessment.

13.4 Fast control and slow learning

The physical control architecture separates:

long-horizon model and experiment planner
        -> requested intervention

local capability and limit validator
        -> permitted trajectory

real-time force/position controller
        -> measured delivery

fast local stop and release path
        -> immediate mechanical response

The learning model can choose among permitted actions but cannot bypass the local force, speed, travel, temperature, or device-state limits. This is an engineering boundary necessary for stable physical control.

13.5 Different objective schemas over one corpus

The same raw episodes can support:

• pleasure optimization;
• muscle-relief optimization;
• pain-boundary mapping;
• mobility improvement;
• reproducibility testing;
• diagnostic information gain;
• patient-preferred examination sequences.

The objective schema determines the decision, not the evidence retained.

14. Self-administered and ingested-stimulus adapter

When the user applies or ingests a stimulus, HumanFB records the best available estimate and its confidence.

14.1 Delivery-verification levels

DIRECT_SENSOR
CONNECTED_DEVICE_REPORTED
PACKAGE_OR_BARCODE_SCANNED
USER_MEASURED
USER_ESTIMATED
USER_RECALLED
UNKNOWN

A user report such as “about half a gummy” remains useful, but it is not stored as a laboratory-measured dose.

14.2 Episode structure

product or activity identification
  -> baseline assessment
  -> administration declaration or measurement
  -> onset monitoring
  -> scheduled or event-triggered check-ins
  -> peak assessment
  -> decline and aftereffect monitoring
  -> later reflective assessment

The appropriate follow-up schedule is adapter-specific and may be learned from previous onset and duration patterns.

15. Worked example: cannabis product logger

The cannabis logger illustrates delayed subjective effects, uncertain product naming, dose estimation, batch variation, price, and availability.

15.1 Product identity

Do not treat the strain name as sufficient identity. Store when available:

• dispensary and SKU;
• producer or grower;
• labeled strain and product name;
• batch or lot;
• harvest, package, purchase, and expiration dates;
• flower, edible, concentrate, vapor, tincture, or other form;
• labeled THC, CBD, minor cannabinoids, and terpene profile;
• laboratory-report reference and extraction version;
• package quantity and estimated consumed amount;
• route of administration;
• receipt price, discounts, taxes, and effective unit price.

Two products sharing a strain label may remain different stimulus snapshots because producer, batch, composition, age, and formulation differ.

15.2 Baseline and context

Capture:

• time since last cannabis use;
• recent-use history and estimated tolerance state;
• food, hydration, caffeine, alcohol, medication, and other relevant recent inputs;
• starting relaxation, anxiety, pain, focus, fatigue, mood, appetite, and sociability;
• setting, intended activity, and objective;
• self-estimated dose confidence.

15.3 Time-indexed prompts

A local scheduler may ask:

baseline:       “How are you feeling before taking it?”
early check:    “Anything noticeable yet?”
developing:     “What changed most?”
near peak:      “How would you characterize the strongest effects?”
declining:      “What remains?”
later:          “Would you choose that product again for the same purpose?”

The schedule can use fixed intervals initially and later adapt to the individual's historical onset curve, dose form, and administration route.

15.4 Response trajectory

Possible axes include:

• relaxation;
• anxiety or unease;
• pain relief;
• focus and distractibility;
• creativity or associative thought;
• sociability;
• sensory enhancement;
• appetite;
• bodily heaviness;
• energy and fatigue;
• mental clarity or fog;
• pleasure;
• onset, peak, duration, and aftereffects;
• immediate and later desire to repeat.

15.5 Example episode record

{
  "exposure_episode_id": "episode_01J...",
  "stimulus_snapshot_id": "product_batch_01J...",
  "administration": {
    "route": "ORAL",
    "declared_amount": 0.5,
    "declared_unit": "piece",
    "estimated_active_amount_mg": 5.0,
    "verification": "USER_ESTIMATED",
    "confidence": 0.62,
    "administered_at": "2026-08-15T19:30:00Z"
  },
  "baseline_state_id": "state_01J...",
  "context_snapshot_id": "context_01J...",
  "objective_id": "EVENING_RELAXATION_v2",
  "assessment_series_id": "series_01J...",
  "purchase_observation_id": "purchase_01J..."
}

16. Reusable profile representation

The personal profile is not one vector and not a permanent personality label. It is a layered collection of conditional, evidence-backed parameters.

16.1 Profile layers

• Broad cross-domain sensitivities.
• Modality-specific parameters.
• Domain-specific attributes.
• Objective-specific weights.
• Context-conditioned effects.
• Source-, product-, sender-, or device-specific effects.
• Interaction effects among stimulus features.
• Positive and negative exemplars.
• Exceptions to broader patterns.
• Time-indexed drift and preference epochs.

16.2 Typed profile parameter

{
  "profile_parameter_id": "parameter_01J...",
  "subject_id": "subject_local_1",
  "scope": {
    "domain": "cannabis_product",
    "objective": "evening_relaxation",
    "context_predicates": ["baseline_anxiety<0.4"]
  },
  "stimulus_feature": "product:terpene_linalool_fraction",
  "response_axis": "relaxation",
  "effect_mean": 0.38,
  "credible_interval": [0.09, 0.61],
  "evidence_count": 12,
  "distinct_batch_count": 5,
  "stability": 0.57,
  "dataset_snapshot_id": "dataset_01J...",
  "model_version_id": "model_01J..."
}

The example records an association unless the dataset and method justify a stronger causal designation.

16.3 Embeddings and exemplars

Embeddings support similarity and clustering where stable concepts do not yet exist. Keep separate embedding families for:

• stimulus content;
• physical or chemical stimulus traces;
• spoken reaction language;
• subjective-state trajectories;
• learned personal preference space.

Every vector names its input artifacts and model version. Exemplar anchors such as “more like this exact moment” can remain directly usable even when the model cannot explain them fully.

16.4 Contradictions and exceptions

HumanFB does not force contradictory observations into one average. It may learn:

• pressure at one location feels relieving while the same pressure elsewhere is painful;
• a cannabis product is preferred for sleep but avoided for focused work;
• direct AI answers are preferred for coding but not exploratory reflection;
• high information density is preferred in articles but disliked in visual interfaces;
• a musical or erotic pacing preference depends on disposition and objective.

These are conditional parameters rather than errors.

17. Database architecture

PostgreSQL is the recommended metadata and evidence store. Content-addressed object storage holds audio, video, images, high-rate sensor traces, source snapshots, and large model artifacts. A graph database may be added for specialized traversal but is not required.

17.1 Evidence tables

17.2 Derived tables

17.3 Raw, normalized, and interpreted layers

Maintain three visible layers:

1. Raw evidence: exact events, media, product records, and sensor streams.
2. Normalized observations: parsed exposures, measurements, ratings, utterances, anchors, and contexts linked to source evidence.
3. Derived interpretations: transcripts, vectors, state trajectories, model parameters, predictions, and decisions.

Normalized and derived projections should be rebuildable from the raw event stream plus versioned code and artifacts.

18. Generic event protocol

18.1 Event envelope

{
  "protocol": "humanfb.events/1.0",
  "event_id": "0198...",
  "event_type": "VOICE_MEMO_COMPLETED",
  "event_schema_version": 1,
  "recorder_id": "device_01J...",
  "boot_id": "boot_01J...",
  "device_seq": 4812,
  "device_monotonic_us": 912334455,
  "device_wall_time": "2026-08-15T20:14:31.552Z",
  "clock_uncertainty_ms": 18,
  "subject_id": "subject_local_1",
  "exposure_episode_id": "episode_01J...",
  "context_snapshot_id": "context_01J...",
  "payload": {},
  "software": {
    "build": "humanfb-recorder/0.4.0",
    "git_commit": "9ad4..."
  }
}

18.2 Core event catalog

Ingestion is idempotent. Device sequence and boot identity preserve local order. Unknown future event types are retained even when the current server cannot yet project them.

19. AI-analysis result schema

A general interpretation artifact may look like:

{
  "artifact_schema": "humanfb.response-interpretation/1.0",
  "response_packet_id": "response_01J...",
  "input_refs": [
    "media_01J...",
    "stimulus_trace_01J...",
    "context_01J..."
  ],
  "transcript_artifact_id": "artifact_transcript_01J...",
  "resolved_references": [
    {
      "phrase": "right there",
      "anchor_id": "body_anchor_01J...",
      "confidence": 0.91
    }
  ],
  "subjective_axes": {
    "pain": {"value": 0.64, "confidence": 0.84},
    "relief": {"value": 0.71, "confidence": 0.77},
    "pleasure": {"value": 0.42, "confidence": 0.58},
    "certainty": {"value": 0.81, "confidence": 0.73}
  },
  "domain_attributes": [
    {
      "concept_id": "palpation:sharp_onset",
      "strength": 0.83,
      "polarity": "NEGATIVE",
      "evidence_spans": [{"start_ms": 210, "end_ms": 880}]
    },
    {
      "concept_id": "palpation:delayed_release",
      "strength": 0.74,
      "polarity": "POSITIVE",
      "evidence_spans": [{"start_ms": 910, "end_ms": 2470}]
    }
  ],
  "open_vocabulary_mentions": [],
  "candidate_attributions": [
    {
      "stimulus_anchor_id": "body_anchor_01J...",
      "attribution_probability": 0.86,
      "latency_model_version": "speech-palpation/1.0"
    }
  ],
  "provenance": {
    "analysis_run_id": "run_01J...",
    "model_version_ids": ["model_stt_...", "model_semantic_..."],
    "prompt_digest": "sha256:17...",
    "ontology_version_id": "ontology_01J...",
    "implementation_digest": "oci:sha256:88..."
  }
}

Confidence fields are component-specific. Later calibration can compare model confidence with explicit corrections and repeated observations.

20. Objective schemas and multi-objective decisions

The same evidence corpus can support many definitions of “better.” HumanFB keeps response predictions, cost observations, availability, uncertainty, and acquisition burden separate, then combines them under a named objective.

20.1 Objective examples

• GENERAL_PREFERENCE_v3
• FOCUSED_WORK_MUSIC_v4
• EVENING_RELAXATION_v2
• PAIN_RELIEF_LOW_SORENESS_v1
• PALPATION_INFORMATION_GAIN_v2
• LABAMP_TOUCH_RESPONSE_v3
• PRIVATE_ADULT_SLOW_PACED_v2
• BEST_VALUE_AVAILABLE_NOW_v1

20.2 Decision utility

For candidate action or stimulus \(a\) under current state \(x\), context \(c\), and objective \(o\):

Here:

• \(R_k\) are desired and undesired response axes;
• \(P(a)\) is price or resource cost;
• \(A(a)\) is unavailability or acquisition burden;
• \(T(a)\) is time, transition, or administration cost;
• \(\mathcal{U}(a)\) is uncertainty when the objective is conservative;
• \(\mathcal{I}(a)\) is information value when the objective favors exploration.

The weights and transformations belong to the versioned objective schema.

20.3 Pareto-frontier presentation

Do not always collapse the result into one number. A decision UI can show:

The user or calling application can select among nondominated tradeoffs.

20.4 Availability observations

Availability is time-dependent evidence, not an intrinsic product property. Store:

• source or vendor;
• observed time;
• location or delivery region;
• stock status and quantity when available;
• price, unit, fees, and discounts;
• retrieval method and confidence;
• expiration or freshness estimate.

Historical recommendations remain reproducible even after the market changes.

21. Profile and intervention compiler

The compiler converts a profile snapshot and current objective into an artifact another system can execute.

21.1 Search-query compilation

The compiler maps private profile concepts into source-specific retrieval vocabulary:

profile target
  -> strong positive concepts
  -> negative concepts and exclusions
  -> source vocabulary and query dialect
  -> narrow, balanced, and exploratory query variants
  -> user or application selection
  -> candidate retrieval
  -> optional private reranking using the richer profile

A public query is a lossy projection. The full personal vector can remain local while the search provider receives only selected terms.

21.2 Compiler feedback

The user may critique the generated artifact:

• “The query captured pacing but overemphasized narrative.”
• “The AI response is now too terse.”
• “The massage plan got pressure right but missed the painful region.”
• “The product shortlist optimized relaxation but ignored price.”

These reactions train the mapping between personal parameters and each target's vocabulary or controls.

22. API surface

22.1 Capture and evidence APIs

22.2 Analysis and profile APIs

22.3 Device command boundary

Physical commands contain:

• expected device and capability manifest;
• expected current state;
• requested protocol ID;
• objective and proposal lineage;
• expiration and nonce;
• local validation requirements.

The device reports requested, accepted, applied, measured, completed, or failed states through events. A server proposal never substitutes for local validation and closed-loop control.

23. Provenance and reproducibility

Every derived artifact identifies:

• exact source and stimulus hashes;
• intended protocol and measured delivery trace;
• exposure, anchor, context, and baseline state;
• raw response and sensor artifacts;
• transcript, feature, ontology, and embedding versions;
• dataset selection, exclusions, and event watermark;
• code, model, prompt, parameter, and random-seed identities;
• objective and normalization schemas;
• compiler and capability-policy versions.

The provenance graph is:

source -> stimulus snapshot -> protocol -> measured exposure
                                      \-> context and baseline

measured exposure + response streams
        -> interpretation runs
        -> state and attribution artifacts
        -> dataset snapshot
        -> model and profile snapshot
        -> prediction or decision
        -> later exposure and outcome

Reprocessing produces new artifacts. A current-preferred pointer may change, but historical interpretations and decisions remain addressable.

24. Profile portability and scoped use

A reusable profile should have a vendor-neutral export that does not require exporting every raw stimulus or voice memo.

The portable form can contain:

• profile schema and snapshot identity;
• typed parameters with scope, effect, uncertainty, and stability;
• objective schemas;
• personal concept and language mappings;
• selected positive and negative exemplars;
• embedding references or locally packaged vectors with model identities;
• provenance summaries and evidence counts;
• compiler settings;
• a disclosure manifest describing which parameter families an application may request.

An application requests a scoped projection:

{
  "purpose": "product_selection",
  "domains": ["cannabis_product"],
  "objectives": ["evening_relaxation"],
  "requested_response_axes": [
    "relaxation",
    "anxiety",
    "mental_clarity",
    "desire_to_repeat"
  ],
  "requested_decision_axes": ["price", "availability"],
  "include_raw_evidence": false
}

This allows the same personal model to serve many applications without giving every consumer the full evidence corpus.

25. Operational failure and uncertainty behavior

26. Validation program

26.1 Event and storage validation

• Golden test vectors for event schemas, hashes, identifiers, and unit normalization.
• Power-loss tests at journal, media-finalization, and physical-command boundaries.
• Duplicate, delayed, reordered, and partial-batch ingestion tests.
• Projection rebuild from raw events.
• Historical decision reproduction from frozen datasets and models.

26.2 Timing and attribution validation

• Shared-clock tests across media, microphone, sensors, robot, and server.
• Known-latency stimulus-response fixtures.
• Voice references to known text, video, body, and transition anchors.
• Controlled A/B and reversed-order tests.
• Prompt-versus-no-prompt experiments to quantify observation reactivity.

26.3 Voice validation

• Wake detection under music, video, conversation, and robot noise.
• Playback rejection using reference audio.
• Enrolled-speaker and meta-commentary classification.
• Pre-roll completeness and trigger-latency reconstruction.
• Domain-language regression corpora.
• User-correction and reprocessing tests.

26.4 Physical-system validation

• Calibration of position, force, pressure, displacement, contact area, temperature, and timing.
• Requested-versus-delivered trace comparison.
• Local limit and stop-path testing independent of server connectivity.
• Body-coordinate repeatability across poses and sessions.
• Simulated sensor dropout, contact loss, slip, and controller saturation.
• Verification that learning proposals cannot bypass the local capability validator.

26.5 Personal-model validation

• Session-aware train/test splits.
• Calibration and credible-interval coverage.
• Repeated-reference episodes to measure consistency and drift.
• Ablations showing which sensors improve prediction beyond explicit voice and ratings.
• Held-out prediction across products, batches, contexts, and domains.
• Cross-domain transfer tests that compare learned transfer with a no-transfer baseline.
• Separate evaluation of immediate response, delayed effect, and reflective satisfaction.

26.6 Initial blinded music evaluation

The first concrete HumanFB evaluation uses 100 shuffled deep album tracks from prominent 1980s and 1990s heavy-metal and hard-rock artists. The working manifest encodes artist, song, and album names with a simple substitution cipher so that casually seeing the queue does not prime the listener. A playback component or facilitator resolves one row at a time; each track is paired with spontaneous or automatically prompted contemporaneous voice capture, compact explicit judgments, a session baseline, and exact playback timing.

The study separates three tasks that could otherwise be confused:

1. Reaction interpretation: derive a faithful, correctable response vector from the current memo.
2. Preference prediction: before a later track is heard, predict its likely response using only earlier evidence and stimulus-side features.
3. Discovery: rank or search for new tracks and compare profile-selected results with a blinded baseline.

Independent internet commentary may be ingested only after the user's response and first interpretation are frozen. It provides post-hoc descriptions and alternate vocabulary, not ground truth about the individual's reaction. Predictions are likewise frozen before exposure so the current voice memo cannot leak into the target being predicted.

The complete protocol, data contract, validity threats, and exit criteria are specified in humanfb_initial_music_evaluation.md. The encoded stimulus artifact is humanfb_80s_90s_metal_hard_rock_blinded_100.txt.

27. Phased implementation plan

Phase 0 — Evidence contract

Implement identifiers, clocks, append-only events, content-addressed objects, sources, stimulus snapshots, contexts, exposure episodes, anchors, and response packets.

Exit criterion: An exposure and reaction can be reconstructed exactly enough to identify what happened, when, under which context, and from which device.

Phase 1 — Local ambient voice capture

Implement local VAD, wake-intent detection, speaker/playback separation, pre-roll, atomic state sampling, voice spooling, transcription, and evidence-linked semantic extraction.

Exit criterion: A natural spoken reaction produces an anchored memo without a button and without losing the stimulus interval preceding the trigger.

Phase 2 — Digital-stimulus adapters

Add text, conversation, audio, and video snapshots; scroll, viewport, turn, and timeline capture; generic anchors; and search/profile compilers.

Exit criterion: The same profile can personalize a web query and an AI response style from evidence collected across several digital modalities.

Phase 3 — Dynamic state and delayed assessment

Add baseline estimates, repeated check-ins, state trajectories, immediate-versus-reflective targets, and latency-aware attribution.

Exit criterion: The system predicts not only overall preference but onset, peak, duration, and later evaluation with calibrated uncertainty.

Phase 4 — Self-administered product logger

Add product, batch, dose-confidence, administration, purchase, price, availability, and delayed-effect schemas. Implement the cannabis logger as a demanding worked adapter.

Exit criterion: A query can compare historical product episodes under a named objective and show predicted response, unwanted effects, price, availability, and confidence separately.

Phase 5 — Physical intervention adapter

Add body coordinates, intervention protocols, measured delivery traces, robot capabilities, local validation, response synchronization, and active experiment proposals.

Exit criterion: A spoken “right there” is correlated with exact robot position and force, produces an anchored pain/relief interpretation, and informs a later permitted intervention proposal.

Phase 6 — Reusable profile platform

Add frozen profile snapshots, scoped portability, cross-domain parameters, compiler feedback, provenance explanations, and multiple downstream applications.

Exit criterion: A new application can request a narrow profile projection and generate an explainable decision without accessing unrelated raw evidence.

28. Open research and design questions

State representation

• Which subjective axes are universal enough for the core schema?
• When should a recurring personal phrase become a first-class concept?
• How should the model distinguish no observation from a neutral response?
• Which parameters should be modeled continuously and which categorically?

Temporal modeling

• How much pre-roll is sufficient for natural spoken references?
• Which latency kernels can be shared across users and which must be personalized?
• How should the system represent multi-peak and rebound effects?
• How should carryover and tolerance be estimated from sparse observations?

Prompting

• When is an AI prompt informative enough to justify interrupting the ongoing exposure?
• Which questions elicit causal explanations rather than post-hoc rationalization?
• How strongly do voice, wording, and timing change the reported response?
• When should clarification happen immediately versus during deferred recap?

Physical intervention

• What anatomical coordinate representation is repeatable across pose changes?
• How should contact geometry and tissue compliance enter the personal model?
• Which response channels provide useful information beyond speech and explicit ratings?
• How should diagnostic information objectives differ from comfort, relief, and pleasure objectives?

Product and market data

• How reliable are product labels relative to producer, batch, and measured composition?
• How should uncertain dose and missing composition propagate into prediction intervals?
• Which price and availability sources can be snapshotted reproducibly?
• When does a batch-specific preference generalize to a product family?

Profile portability

• Which profile parameters transfer reliably across domains?
• How can a scoped profile remain interpretable across different model families?
• When should a downstream application receive typed parameters, embeddings, exemplars, or a compiled policy?
• How can the user inspect the evidence behind a profile parameter without reviewing an overwhelming raw history?

29. Product thesis

Most personalization systems observe consumption behavior and infer preference indirectly. HumanFB records subjective meaning explicitly and connects it to a synchronized account of what the person actually experienced.

The unit of personalization is not:

user consumed item X

It is:

this exact source, version, component, or physical protocol
+ this exact presentation or measured delivery
+ this baseline, context, and recent history
+ this immediate and delayed response trajectory
+ these spoken thoughts and explicit judgments
+ these versioned interpretations and uncertainties
+ this objective-specific profile update

That record supports a qualitatively different class of systems. A search engine receives an editable preference projection. An AI assistant receives a purpose-specific interaction style. A product selector considers subjective effect, price, and availability. A robot receives a constrained intervention proposal. A researcher or user can inspect which evidence produced any conclusion.

The long-term system is a personal model of subjective dynamics: how this individual tends to change in response to stimuli, interventions, context, and their own thoughts. Likes and dislikes are one projection of that model. Relief, learning, focus, arousal, pain, pleasure, curiosity, clarity, and later satisfaction are others.

Conclusion

HumanFB generalizes human-in-the-feedback-loop preference learning into a Personal Stimulus–Intervention–Response Graph. It records arbitrary digital and physical stimuli, exact presentation or measured delivery, naturally expressed thoughts, explicit judgments, physiological and behavioral observations, instantaneous disposition, response trajectories, and delayed assessments. It then builds versioned, contextual personal models that can be reused across search, recommendation, AI interaction, product selection, experiment design, and physical control.

The strongest architectural rule is to preserve the chain of evidence at full resolution. A robot command is not a measured force. A product name is not a batch composition. A heart-rate change is not an emotion. A voice memo is not its semantic vector. A state change is not automatically a stable preference. A preference estimate is not a purchasing decision. Each becomes a separate object linked through time, anchors, provenance, and uncertainty.

With that separation, HumanFB can begin as a practical voice-enabled logger and grow into a general model of personal subjective response without discarding or corrupting its earliest observations. The same system can remember what the user experienced, explain what the user appeared to care about, predict how a new stimulus may affect them, and choose what to present or apply next under whatever objective the user specifies.