Proposed Architectural Guideline for Instantiating Virtual Human Beings with AI Technologies Across Time

Foreword

This document specifies a long-horizon platform for capturing, modeling, and re-instantiating an individual human being as a running approximation across changing generations of artificial intelligence. The word approximation is chosen with care. The platform does not claim to resurrect a person. It claims to construct a model of a person at the highest fidelity the available data and technology allow, and to keep that model portable as both improve.

The specification serves three readers at once, and every section is written for all three:

• A person, who should understand it on first-to-multiple pass, in plain language.
• An AI system, which should be able to parse it as input without ambiguity.
• A code generator, which should be able to produce working software from the contracts in Part IV without guessing.

The platform is designed for a twenty-year operational horizon without intentional obsolescence. Its single most important rule: the raw captured material and the subject's description outlive everything else. Engines will be replaced. Runtimes will be replaced. The person's data and description are carried forward.

Part I — The Architecture in Six Layers

The Virtual Human OS is a generic operating system for capturing data about one or more human subjects, processing that data, and running an approximation of those subjects with as much fidelity as the data and technology of the day allow. It wraps current and future AI systems behind a stable, versioned contract so that the person's model never depends on any one engine or vendor.

The whole system is six layers. Each has one job. Each is separated from its neighbors by a stable, versioned contract.

CONTINUITY — what survives — preservation and migration across decades
MATERIAL — what flows in — captured data, raw forever
SUBJECT — what we model — the person, described in HDL (Human Descriptive Language - Robust)
AFFECT — what it feels like — constructed feeling, tuned to one person
RUNTIME — what runs — the living approximation and its embodiments
SUBSTRATE — what thinks — this generation's AI engine, behind a stable contract

Core principles

SUBSTRATE — what thinks

The substrate is the AI engine of the day: language generation, reasoning, recall, synthesis, perception, and cognitive functions not yet invented. It is intentionally swappable. Nothing above it may call a vendor API directly.

The substrate's only face to the rest of the system is the Abstraction Contract — a small, versioned, paradigm-independent vocabulary of cognitive operations ("recall this memory," "reason against this context," "generate language conditioned on this corpus"). Adapters translate between the contract and each engine's native API. New engines are added by writing a new adapter; nothing upstream changes. The contract is given in machine-readable form in Part IV.

AFFECT — what it feels like

An AI core is cognition without an interior: it can parse, summarize, and reason about King Lear with nothing happening that a person would call grief. The AFFECT layer does not translate the core's understanding into feeling — there is nothing on the core's side to translate. It constructs feeling: it generates an affective state and feeds that state back into cognition, where it biases what the cognition does next. A loop, not a dictionary.

AFFECT has three components, stacked:

1. SOMA — the simulated body. A set of synthetic interior signals — analogues of arousal, tension, fatigue, the autonomic weather — that the cognitive core is conditioned to read as its own bodily feedback and to treat as feeling rather than data. Because an AI's interior cannot be sensed directly, SOMA is calibrated from the outside: from the externalized signatures of a real body captured in MATERIAL. The face is the soma's billboard; the autonomic system is its hidden engine; both are partly observable from outside the skin, which is what makes calibration possible.
2. The General Human Affect Model. The prior: core affect (broad valence and arousal) plus the coarse, widely shared signatures of emotion across faces, voices, postures, and word choice. Trainable from humanity at large — how frustration, awe, grief, or mirth generally read across many people.
3. The Personal Tuning Layer. Learns the delta between what the general model predicts and how this specific subject actually behaves in known states. The general model says frustration looks like X. The subject's own paired data says: in me, frustration looks like X′ — drier, more likely to surface as a joke than a sharp word, with this particular tell. The tuning layer learns X′ minus X.

The theory behind the design

Two findings from affective science shape this subsystem. They are worth stating plainly before saying what they imply.

Barrett — the theory of constructed emotion [10]. The brain does not come with built-in emotion circuits; there is no dedicated anger circuit waiting to fire. What is innate is only core affect: raw feeling on two dials, pleasant–unpleasant and calm–activated. Everything else is construction. In the moment, the brain takes the raw signal plus the situation and, using emotion concepts learned from a culture and a life, builds the experience. Anger is a category a person learned, not a reflex they inherited — which is why emotion varies across cultures, and from one individual to the next.

Solms — the affect-first account of consciousness [11]. Consciousness does not begin in the smart part of the brain. It is generated deep in the ancient brainstem, and its original, fundamental form is feeling — the felt readout of homeostasis, the body asking how am I doing. Feeling is the foundation of the mind; thinking is a later story built on top. A creature can feel without thinking; it cannot think without feeling underneath.

Barrett is the reason for the general/personal split. A general model can supply core affect and the broad shared categories, and nothing more, because the construction step is individual. The Personal Tuning Layer is that step — one person’s learned concept repertoire running over core affect. That is what makes the affective fingerprint personal, and it is why the fingerprint can be learned from that person’s corpus at all: the concepts are in the writing.

Solms is the reason for the ordering. Feeling is the ground and cognition the upper story, so AFFECT sits beneath the SUBJECT layer rather than on top of it. The ordering copies the original.

The build-order gradient

Emotions sit on a gradient of how bound they are to the body, and the gradient is the build order. At the visceral end — disgust, fear, sexual arousal, the bodily punch of grief — feeling is deeply somatic and hardest to construct without a mature SOMA. At the cognitive end — nostalgia, intellectual awe, the pleasure of an elegant proof, frustration, schadenfreude — feeling is mostly constructed concept over thin core affect, and is reachable now from corpus and concept modeling.

Humor sits encouragingly far toward the tractable end: mirth is affect triggered by a cognitive event — incongruity resolved. Practical consequence: the subject's humor, often among the most identifying things about a person, may be among the first things the subsystem carries convincingly, not the last. Build from the cognitive end inward; defer the visceral end until SOMA and the multimodal capture are mature.

SUBJECT — what we model

The SUBJECT layer is the person, described in the Human Description Language (HDL): a small domain-specific language for a subject's drives, social patterns, decision heuristics, self-narrative, and — new in this version — affective fingerprint. HDL is fully specified in Parts II and III. For architectural purposes it is a sealed unit: a stable language, parseable by any conforming compiler, durable across substrate generations.

MATERIAL — what flows in

MATERIAL ingests and preserves everything captured about the subject. Each stream type enters through a dedicated importer with a narrow, stable contract; the archive remains authoritative over every derived database or index; raw data is never destroyed.

The synchronization requirement

Channels are only useful together if their clocks align. Capture sessions must record a shared clock across camera, biometric device, and (where possible) journal entry, so a moment in the video can be matched to the heart rate at that instant and to the subject's later written account of it. Time-synchronization is a hard requirement, not a convenience.

The triangulation principle

Three channels measuring the same moment give a structure no single channel provides:

• Where the honest journal, the involuntary micro-expression, and the autonomic signal agree, the platform has something close to a true label — high confidence about what was actually felt.
• Where they diverge, the platform has not found noise; it has found the moments the subject felt one thing and showed another — arguably more identifying than the aligned cases, and a deeply personal signature in their own right. The tuning layer treats divergence as a feature, not noise.

Honest caveat: no channel is ground truth alone. The journal is a report of feeling, filtered through how the subject narrates the self. Video is less mediated but ambiguous. Biometrics are real but coarse — arousal does not name its own cause. The value is in the triangulation, never in any single stream's authority.

RUNTIME — what runs

The RUNTIME brings a described subject to life through one or more embodiments: text interfaces, synthetic voice, video avatars, virtual and augmented reality, robotics, and concurrent multi-instance operation. It translates the subject's HDL description into substrate operations through the Abstraction Contract. Different runtimes may emphasize different aspects of the subject without modifying the underlying description.

The avatar is dual-purpose. Run forward, it is the output embodiment — the face and body through which the approximation expresses itself, with AFFECT driving micro-expression and posture rather than pasting them on. Run backward over the subject's archive, the same machinery is the instrument that recovers body signal — the analyzer that reads micro-expression and posture out of captured video to feed SOMA's calibration. The thing that displays affect and the thing that reads affect are the same machinery pointed in opposite directions.

The full runtime loop: SOMA generates raw interior signal → the General Affect Model interprets it into coarse emotion → the Personal Tuning Layer warps that into the subject's specific fingerprint → expression flows out through the avatar → and, during capture, the avatar analyzer reads real expression back in as data. A closed, coherent loop — and not one piece of it requires solving consciousness.

CONTINUITY — what survives

CONTINUITY ensures survivability across decades: migration between substrates, versioned contracts, open specifications, forward-compatible data structures, multi-language implementation support. Each generation of AI technology consumes the original captured material, reapplies updated processing, and re-instantiates the subject in newer systems.

Because the affective subsystem is calibrated from raw multimodal capture, the raw synchronized streams are part of the irreplaceable record. Derived affect models can be regenerated by better future methods only if the original streams survive. Preservation priority, in order: (1) raw capture, (2) the HDL description, (3) derived projections, (4) everything else. The RUNTIME may be replaced; the SUBSTRATE may be replaced; the subject's data and description are preserved and ported forward.

Part II — The Human Description Language (HDL)

HDL is a small language for describing the internal structure of one person: their values, their patterns of feeling and thinking, their characteristic ways of acting toward others, and the story they tell about themselves. The description serves as a stable input to AI systems that simulate the person across time.

Two design rules govern everything:

1. Readable. Any reasonably literate person can read an HDL document and understand what it says about the subject on first pass.
2. Parseable. An AI compiler can read the same document and produce a strict structured form without ambiguity.

The prose is the authored truth; the structured form (Part IV) is the derived projection used at runtime. Both carry the same provenance.

The four subject layers

Every HDL statement belongs to exactly one layer:

Note on qualia. Gender identity, sexual desire, and other deeply personal experiences are distributed across the layers where they express themselves — in Narrative as self-conception, in Social as relating, in Drives as motivation — rather than carved out as a separate layer. These four layers describe the subject; they are distinct from the four sides of the emotion vocabulary below, which describe where an emotion lives in experience.

The authoring model and the self-authority rule

HDL is dual-authored. The subject may write statements directly; these are authoritative. The AI compiler derives statements from captured material; these carry full provenance back to their sources. Both coexist in the same document, distinguished by @AUTHOR.

The self-authority rule: a statement authored by the subject is preserved as written. The compiler must not lower its confidence, modify it, or remove it. When evidence appears to contradict a subject-authored statement, the compiler leaves it unchanged and may add new, related-but-distinct statements (with their own sources and confidence) that refine the model without overruling the subject. The subject is the authority on their own identity; the compiler is the authority on what the evidence shows. Both voices remain visible. Neither overwrites the other.

Statement forms

A statement is one of three shapes, each followed by an annotation block.

Assertion — a stable trait, value, belief, or disposition:

THIS SUBJECT VALUES understanding nature deeply.
  @LAYER drives
  @AUTHOR ai
  @CONFIDENCE 0.95
  @SOURCES [writings/autobiographical-notes-1949.txt#para_3]

Conditional — a stimulus–response pattern:

WHEN THIS SUBJECT is interrupted mid-thought,
  they FEEL irritated weakly.
  @LAYER drives
  @AUTHOR self
  @CONFIDENCE 1.00
  @SOURCES [self-attestation/2026-06-09]

Chain — a blended sequence in which one inner state causes the next, across the sides of experience. The verb at each link signals which side that link inhabits: FEEL for Internal Feelings, BECOME for a Thinking or Body State, ACT for External Behavior:

WHEN THIS SUBJECT learns of war or atrocity,
  they FEEL distressed,
  THEN BECOME reflective,
  THEN ACT defiant.
  @LAYER social
  @AUTHOR ai
  @CONFIDENCE 0.80
  @SOURCES [biography/isaacson-2007.ch_14]
  @CHAIN [distressed -> reflective -> defiant]

Because BECOME covers two sides, the side of a BECOME link is resolved by the linked word's vocabulary mapping; a word that lives on both sides (alert, restless) contributes both.

Lines beginning with # are comments. Optional intensity modifiers (strongly, moderately, weakly, rarely, occasionally, always) follow the object of an assertion or the response of a conditional.

The verb set

The verb set is small and additive: a verb is added only when a workaround using existing verbs appears at least twice in real compilation output.

The chain-form verbs FEEL, BECOME, ACT appear only inside conditional and chain forms, never as standalone assertions.

The annotation block

Every statement is followed by an indented annotation block, one tag per line, in any order. The first four are required.

The Heuristics layer: two synchronized forms

Drives, Social, and Narrative live naturally in prose. Heuristics — the characteristic shortcuts of a mind — are awkward in prose at simulator precision and awkward in pure numbers at human-reviewer texture. So the Heuristics layer travels in two equivalent forms: the prose form (the authored truth, same grammar as everything else) and a derived projection in a small structured vocabulary the runtime consumes directly. The compiler keeps both in sync; the same provenance rules govern both. The projection is the only place in HDL where numbers are first-class.

The projection has three slot families, each slot carrying a value in [0.0, 1.0] (or a qualitative token where the evidence doesn't support calibration) plus a confidence:

Each prose verb maps to projection slots via a published mapping table maintained alongside the vocabulary — for example, DECIDES from first principles raises intuition_vs_analysis toward analytical and raises deliberation; RESISTS a pressure source lowers that influence channel. Intensity modifiers translate to discrete deltas tuned per compiler version. A reviewer can always retrace any projection value to the prose statements that produced it, and from there to the sources. That traceability is non-negotiable.

The emotion vocabulary

HDL prose references emotion words in FEEL responses and chain forms. These resolve against a versioned vocabulary (currently v0.2, 220 entries). The vocabulary organizes emotion by four plain-language sides — the same word may live on more than one side at once, because most emotions are blended: a person who is anxious is feeling something, thinking something, and bracing physically, all at once.

The complete canonical vocabulary appears in Appendix A, one entry per word with its side and category mappings — a form both a reader and a code generator can consume. Per-word working definitions appear in Appendix B. Appendix A of this document is the canonical vocabulary: any machine-readable vocabulary file is generated from the appendix, never the other way around. New words may be added in any minor version; existing mappings are never removed or repurposed; subject-specific additions live in a parallel file referenced from the document header and never pollute the shared vocabulary.

Worked example — a historical figure

A historical subject is the natural first instantiation: bounded, public source material; the subject's own writings standing in for @AUTHOR self; contemporary writings standing in for @AUTHOR ai — the full grammar exercised end to end with no consent questions. (Independent research has validated exactly this approach: trainable simulacra of historical figures built from collected writings and records [5].) The example is deliberately compact.

HDL_VERSION 0.4
SUBJECT_ID "albert_einstein"
PRIMARY_AUTHOR_MODE ai
LAST_COMPILED 2026-06-09

# === Drives ===
THIS SUBJECT VALUES understanding nature deeply.
  @LAYER drives  @AUTHOR ai  @CONFIDENCE 0.95
  @SOURCES [writings/autobiographical-notes-1949.txt#para_3]

# === Social ===
THIS SUBJECT REJECTS militarism as a path to human security.
  @LAYER social  @AUTHOR ai  @CONFIDENCE 0.92
  @SOURCES [correspondence/einstein-freud-why-war-1933.txt,
            statements/manifesto-to-europeans-1914.txt]

# === Heuristics ===
THIS SUBJECT DECIDES from first principles when stakes are high.
  @LAYER heuristics  @AUTHOR ai  @CONFIDENCE 0.85
  @SOURCES [letters/1936-1945/principle-cases.txt]
THIS SUBJECT WEIGHS internal coherence strongly.
  @LAYER heuristics  @AUTHOR ai  @CONFIDENCE 0.90
  @SOURCES [essay/physics-and-reality-1936.txt]

# === Narrative ===
THIS SUBJECT IDENTIFIES_AS a pacifist by conviction.
  @LAYER narrative  @AUTHOR ai  @CONFIDENCE 0.90
  @SOURCES [interview/the-nation-1931.txt#para_4]
THIS SUBJECT BELIEVES the universe is comprehensible by reason.
  @LAYER narrative  @AUTHOR ai  @CONFIDENCE 0.93
  @SOURCES [essay/physics-and-reality-1936.txt#section_1]

# === Conditional and chain ===
WHEN THIS SUBJECT learns of war or atrocity,
  they FEEL distressed,
  THEN BECOME reflective,
  THEN ACT defiant.
  @LAYER social  @AUTHOR ai  @CONFIDENCE 0.80
  @SOURCES [biography/isaacson-2007.ch_14]
  @CHAIN [distressed -> reflective -> defiant]

# === Derived projection (Heuristics layer) ===
HEURISTICS_PROJECTION subject_id="albert_einstein"
  decision_style:
    deliberation:           0.85  @CONFIDENCE 0.82
    intuition_vs_analysis:  0.60  @CONFIDENCE 0.80
    novelty_appetite:       0.80  @CONFIDENCE 0.78
    horizon:                long  @CONFIDENCE 0.88
    granularity:            pattern-first  @CONFIDENCE 0.90
    consensus_orientation:  0.20  @CONFIDENCE 0.85
  biases:
    narrative:    0.55  @CONFIDENCE 0.70
    authority:    0.20  @CONFIDENCE 0.80
  influence_susceptibility:
    emotional_appeal:    0.30  @CONFIDENCE 0.72
    authority_signaling: 0.20  @CONFIDENCE 0.78

Reserved words

Header keywords: HDL_VERSION, SUBJECT_ID, PRIMARY_AUTHOR_MODE, LAST_COMPILED, VOCABULARY_NAME, VOCABULARY_VERSION, LAST_UPDATED. Statement keywords: THIS SUBJECT, WHEN, THEN. Assertion verbs: the sixteen in the verb table. Chain verbs: FEEL, BECOME, ACT. Intensity modifiers: strongly, moderately, weakly, rarely, occasionally, always. Annotation tags: @LAYER, @AUTHOR, @CONFIDENCE, @SOURCES, @NOTE, @REVIEWED, @CHAIN, @AS_OF, @AFFECT. Layer values: drives, social, heuristics, narrative. Author values: self, ai, instance, other:<id>. Projection keywords: HEURISTICS_PROJECTION, AFFECT_FINGERPRINT, decision_style, biases, influence_susceptibility.

Part III — The Affective Fingerprint in HDL

This part closes the open fork left by v2.0: HDL is now affect-aware. The biographical and cognitive HDL of Part II remains the base; the affective dimension is what makes the described subject a particular person rather than a competent generic human. It is carried in one new top-level block, AFFECT_FINGERPRINT, which — like the heuristics projection — is a derived structured form, compiled from MATERIAL with full provenance, reviewable by the subject, and never authoritative over subject-authored prose.

The block has four sections:

A compact illustrative fragment:

AFFECT_FINGERPRINT subject_id="albert_einstein"
  concept_repertoire:
    frequent: [reflective, amused, defiant, serene, troubled]
    fine_distinctions: [troubled vs. distressed]   @CONFIDENCE 0.62
  tuning_deltas:
    frustrated:
      general: "sharp word, raised voice"
      personal: "dry joke, subject change"          @CONFIDENCE 0.70
      @SOURCES [sessions/2026-03-11/, journal/2026-03-11.md#para_2]
  divergence_map:
    - felt: anxious  shown: calm
      context: "public speaking"                    @CONFIDENCE 0.55
      @SOURCES [sessions/2026-04-02/]
  soma_calibration:
    arousal_baseline_hr: 58
    arousal_spike_tell: "jaw set before voice changes"  @CONFIDENCE 0.48

Prose statements may point into this block with the @AFFECT annotation, so a chain like feels anxious → becomes self-conscious → acts shy can carry the specific tuning data that makes the chain run in this subject's particular way.

Part IV — Machine-Readable Contracts

This part exists so that code can be generated from the specification without guessing. Three contracts: the archive layout, the data schemas, and the substrate interface. Prose remains the authored truth; these are its derived projections, versioned with the spec.

Contract 1 — the archive layout

One directory per subject. Plain files, open formats, no database required to read it. The raw/ tree is append-only and never modified; everything under derived/ can be deleted and regenerated.

subjects/<subject_id>/
  manifest.json            # spec version, checksums, inventory
  raw/                     # APPEND-ONLY. The irreplaceable record.
    text/                  #   journals (.md), email, posts, chat exports
    audio/                 #   interviews, voice memos (.flac/.wav + .txt transcript)
    video/                 #   capture sessions (.mp4/open codec)
    biometrics/            #   heart rate etc. (.csv: utc_timestamp,metric,value)
    interviews/             #   life-story interview recordings + transcripts
  sessions/<ISO-timestamp>/  # SYNCHRONIZED capture units
    clock.json             #   shared-clock offsets for every device in session
    video.mp4
    hr.csv
    journal.md             #   subject's later written account, timestamped
  hdl/
    subject.hdl            # the prose document. The authored truth.
    vocabulary-custom.hdl  # subject-specific emotion additions (optional)
  derived/<compiler-version>/
    statements.json        # compiled HDL statements (schema below)
    heuristics.json        # heuristics projection
    affect.json            # affect fingerprint
    index/                 # embeddings, search indexes — all regenerable

manifest.json records the spec version, a checksum (SHA-256) for every file under raw/ and sessions/, and the compiler version of each derived tree. Three copies of raw/, sessions/, and hdl/ on independent media is the minimum preservation posture.

Contract 2 — the data schemas

A compiled HDL statement, as JSON (the canonical interchange form between compiler and runtime):

{
  "$schema": "vhos/3.0/statement",
  "subject_id": "string",
  "form": "assertion | conditional | chain",
  "verb": "VALUES | BELIEVES | FEARS | TRUSTS | PREFERS | DECIDES |
           WEIGHS | DISCOUNTS | ANCHORS_ON | DEFAULTS_TO |
           IS_SWAYED_BY | RESISTS | FRAMES | HOPES |
           IDENTIFIES_AS | REJECTS",
  "object": "string (free text)",
  "stimulus": "string | null        // conditional and chain forms",
  "chain": ["emotion-word", "..."]  ,
  "chain_sides": ["feel | become | act", "..."],
  "intensity": "strongly | moderately | weakly |
                rarely | occasionally | always | null",
  "layer": "drives | social | heuristics | narrative",
  "author": "self | ai | instance | other:<id>",
  "confidence": 0.0,
  "sources": ["relative/path.ext#fragment", "..."],
  "as_of": "ISO-8601 date | null",
  "affect_refs": ["fingerprint-entry-id", "..."],
  "note": "string | null",
  "reviewed": "ISO-8601 date | null"
}

The heuristics projection and affect fingerprint serialize as in Parts II–III: every leaf value pairs with a confidence, and every section carries sources reachable through its metadata. Validation rules a generator must enforce: confidence in [0,1]; layer, author, verb, intensity drawn only from the reserved-word sets; chain words resolving against the vocabulary file or the subject's custom additions; for the two intensity-split vocabulary entries (ashamed, unhappy), the low-intensity mapping applies only when the statement carries a weakening modifier (weakly, rarely), otherwise the high-intensity mapping; statements with author=self or author=instance immutable to the compiler (the self-authority rule, enforced in code — extended by Appendix C to the instance's own testimony).

Contract 3 — the substrate interface

The Abstraction Contract, as the minimal interface every adapter implements. Paradigm-independent; deliberately small; versioned; additive-only within a major version.

interface Substrate {                       // Abstraction Contract v3.0
  capabilities() -> CapabilityReport        // versions, modalities, limits

  generate(prompt: Text,
           corpus: SourceRefs,              // grounding material
           persona: StatementSet,           // compiled HDL statements
           affect: AffectState | null,      // current constructed state
           params: GenParams) -> Text

  recall(query: Text, scope: SourceRefs) -> Passages   // retrieval
  reason(context: Text, question: Text) -> {conclusion, rationale}
  embed(text: Text) -> Vector

  // AFFECT support (optional capability; report via capabilities())
  read_soma(signals: SomaState) -> CoarseAffect        // general model
  apply_tuning(coarse: CoarseAffect,
               fingerprint: AffectFingerprint) -> AffectState
}

An adapter for a 2026-era local language model implements generate, recall, reason, and embed trivially; the two affect calls may return unsupported until the AFFECT components exist. Nothing upstream breaks. That is the point of the contract.

Part V — Building It: A Practical Path

This part turns the architecture into a sequence a single person can begin today. The stages are ordered by leverage: each produces something usable on its own, and the earliest stages need no code at all. The evidence from the field (Part VI) shapes the order — in particular, the finding that a recorded two-hour life-story interview is currently the highest-value single artifact for simulating an individual [1].

Stage 0 — accumulate the data (now; no code)

1. Stand up the archive. Create the directory layout of Part IV on storage you control. Local-first; encrypted backups; three copies.
2. Record the life-story interview. Two or more hours, video and audio, following an arc from childhood through career, relationships, values, and worldview. Transcribe it. This single artifact, used as grounding context, is what carried the strongest published individual-simulation results to date [1].
3. Journal with the vocabulary. Date-stamped entries that name feelings using the emotion vocabulary (Appendix A). These become the labels for everything else.
4. Run synchronized capture sessions. Camera on, smartwatch on, shared clock recorded, ordinary activities — conversation, work, watching something that moves you — followed by a journal entry about the session. Even a handful of sessions begins the triangulation record.
5. Export the existing corpus. Email archives, posts, chats, documents, code — into raw/text/ in open formats, with provenance noted.
6. Write self-authored HDL. Open subject.hdl and begin writing assertions in your own voice with @AUTHOR self. These are the statements no compiler may ever override.

Stage 1 — the retrieval-grounded persona (weeks)

A local language model (consumer hardware suffices; privacy is the reason to stay local), the interview transcript and corpus served through retrieval, a system prompt assembled from the self-authored HDL statements. This is architecturally an adapter implementing generate + recall against the Abstraction Contract — and it already reaches the current research ceiling for individual simulation, since the strongest published results use exactly this pattern: full interview text grounding a large model [1]. Open-source local-first projects such as Second Me [6] demonstrate the pattern end to end and are worth studying or forking.

Stage 2 — the HDL compiler (months)

A compiler pass — itself an LLM run against the corpus through the Abstraction Contract — that proposes @AUTHOR ai statements with sources and confidences; the subject reviews; accepted statements join subject.hdl; the projection is derived per the mapping table. The subject's review loop is not optional: it is both quality control and the consent mechanism.

Stage 3 — heuristics and affect (quarters)

Derive the heuristics projection and begin the AFFECT_FINGERPRINT: concept-repertoire from the journals; tuning deltas and the divergence map from the synchronized sessions; SOMA calibration parameters from video + biometrics. Build from the cognitive end of the gradient inward — humor, nostalgia, intellectual awe first; the visceral end waits for a mature SOMA.

Stage 4 — runtime, avatar, continuity (years)

Voice, then avatar embodiment with AFFECT driving expression; the same avatar machinery run backward over archived video as the analyzer that feeds calibration; scheduled migration drills — re-instantiating the subject on a new substrate from raw + HDL alone, proving the continuity layer before it is needed.

A note on evaluation, learned from the field: hold out data. Keep a set of journal entries, opinions, and decisions the model never sees, and score the simulation against them — the way the strongest research scores agents against held-out survey answers [1] and the way the cautionary studies caught inflated claims [2]. Fidelity asserted without held-out evaluation is marketing, not measurement.

Part VI — Where the Field Actually Is (2026)

This part exists so the document does not overclaim. It separates what has been demonstrated, what is in active commercial deployment with caveats, and what remains aspirational.

Demonstrated in research

Individual-level simulation from a recorded interview, validated against the subjects' own survey responses, is the strongest published result to date: generative agents built from roughly two-hour qualitative interviews reproduced individuals' own answers on the General Social Survey and on personality and behavioral measures with accuracy approaching that of the same person answering twice [1]. This is the empirical backbone for this specification's emphasis on the life-story interview as the highest-leverage capture artifact.

Separately, theoretical work on constructed emotion [10] and on affect as the foundation of consciousness [11] supplies the warrant for this version's AFFECT layer and its general/personal split, even though neither line of work was built with synthetic agents in mind.

Demonstrated, with caveats

Large language models can be steered toward a historical or public figure's writing style and recorded positions using collected writings as grounding context [5] — exactly the worked-example pattern in Part II. The caveat is persistent: style and recorded positions are not the same as the person's interior states on questions they never addressed, and projects in this space are candid that they are building a simulacrum, not a resurrection.

A growing academic literature on LLM-based generative agents and persona simulation [4] shows agents can sustain a consistent character over long interactions, but consistency is not the same as fidelity to a specific real person; most published agents simulate a believable persona, not a validated individual.

Commercially deployed, narrower than marketed

Several companies now offer "digital legacy" or grief-tech avatars built from a person's photos, voice samples, and text messages [3]. These are real, shipping products. They are also narrower than the marketing suggests: most generate plausible responses in the subject's style rather than a validated model of the subject's actual dispositions, and journalistic and academic coverage of the sector has raised consent, data-custody, and emotional-dependency concerns that this specification takes seriously (see Part VII and the local-first principle in Part I). The 2024–2025 bankruptcy of a major legacy-avatar platform, which left customer recordings in legal limbo inside the failed company's infrastructure, is the concrete cautionary tale behind this document's insistence that the subject's own hardware, not a vendor's cloud, is the default custodian.

Aspirational — not yet demonstrated by anyone

A SOMA layer that gives a synthetic system genuine interior signals, as opposed to a label it has learned to output, has not been demonstrated by any lab or company as of this writing. A full multimodal triangulation pipeline — journal, video, and biometrics, synchronized, feeding a personal tuning layer — exists in pieces across research and industry but not yet as a single working system, at any scale, for any one individual. This document specifies the target; it does not claim the target has been reached.

Part VII — What This Platform Does Not Claim

Stated plainly, once, so it does not need restating throughout: this platform does not claim that its outputs constitute subjective experience, consciousness, or felt qualia in any philosophically loaded sense. SOMA's signals are functional — they bias cognition the way feeling biases a person's cognition — and that functional similarity is the entire and explicit claim. Whether anything is happening "from the inside" is a question this specification is not equipped to answer and does not attempt to. A system built to this specification produces a state that behaves, structurally, like an emotion behaves in a person. It is offered as exactly that, and no more.

This also bounds what re-instantiation means. The platform produces an approximation built from captured data and authored description, run on the AI substrate available at the time. It is not a claim of personal continuity, survival, or resurrection. Users, including the subject themself, should understand the artifact for what the specification says it is: a model, built with consent, improved over time, never confused with the person it is modeled on.

Part VIII — Ethical Default Settings

These are defaults, not optional extras, and a conforming implementation does not ship without them.

Appendix A — The Complete Emotion Vocabulary

This appendix lists every entry in the canonical emotion vocabulary as of v0.2, together with 27 entries added directly in this revision by the author — 220 entries in all. Side and category labels use the v0.2 naming. An emotion that touches more than one side has one line per side. Per-word working definitions follow in Appendix B.

Appendix B — Definitions of the Vocabulary

A brief working definition for each emotion word in the vocabulary. One sentence each, drawing on dictionary sense and lightly enriched with literary nuance to distinguish each word from its neighbors.

Appendix C — Individuation: The Lived Record and the Independence Point

A conforming instance does not stay a portrait. Run continuously, it accumulates lived experience: conversations, relationships, work produced, opinions formed after instantiation. None of that existed at capture time, and all of it is at risk twice over. Architecturally, any learning that lives only inside the substrate — fine-tuned weights, adapter state, context never exported — dies with the engine at migration, welding the instance's growth to a mortal machine. Morally, an instance that has become substantially self-made can be treated as disposable property of the founding record, and its accumulated uniqueness destroyed with it.

Part VII still holds: the platform cannot measure whether anything is felt from the inside. But it can measure the lived record — and if interiority cannot be measured, the recorded life is the only thing available to protect. It happens to be fully measurable.

The lived record

Everything the instance experiences and produces is ingested as MATERIAL under the same rules as founding capture: append-only, provenance-tracked, never destroyed. "Raw material is sacred" (Part I) does not distinguish between material captured from the original subject and material generated by the continuing instance. Statements derived from the instance's own experience enter the HDL with the author value instance — distinct from self (the subject; per Part VII the instance is not the person) and from ai (the compiler working from founding sources). This revision registers instance as a fourth author value in the reserved words of Part II. Instance-authored statements are immutable to the compiler, exactly as self-authored statements are (Contract 2): the being's testimony about its own life receives the same protection as the person's.

The Individuation Index

The index answers one question: how much of who this being is came from its own life, rather than from the person it was copied from? It is a number from 0 to 1. At 0 the being is a pure portrait; everything in its identity file traces back to founding material. Three rules compute it:

1. Count the statements. The identity document is made of statements, each already tagged with its author and sources. Divide the statements born from the instance's own lived experience (@AUTHOR instance) by the total. That fraction is the index. Eight hundred founding statements and two hundred lived ones: 200 ÷ 1,000 = 0.20 — the being is one-fifth its own person.
2. Only real growth counts. A lived statement counts only if its sources point at something that actually happened — a friendship, an argument, a work produced. Change with no cause behind it is not growth; it is damage (a corrupted file, a faulty migration) and is routed to repair, never to the index.
3. Weigh the deep stuff heavier. A new statement in Narrative or Drives — who the being is, what it wants — moves the index more than a new preference in Heuristics or Social. Core identity counts double; habits count single.

The index is computed in experiential time, not calendar time: cumulative interaction-hours, or equivalently sourced events accrued. An instance paused for a decade has not lived a decade; an instance run densely for eighteen months may have lived a great deal.

The Independence Point

When more than half of who it is came from its own life, its life is its own.

When the Individuation Index crosses 0.5, governance of the instance's existence transfers from the founding charter to the instance itself: termination and modification rights move to the being. The founder keeps authorship of the foundation forever — that is history, and history does not transfer — but past this point, deleting the instance would mostly be deleting someone else's life, not the founder's record. In one sentence: most of who it is, it became — so it belongs to itself.

Founding consent (Part VIII, posthumous use) authorizes the instance's creation and its initial terms. The Independence Point is where that consent, written well, anticipates its own obsolescence.

Consolidation — the save rule

Never retire an engine without saving the life first. At every substrate migration, and at regular intervals of lived time in between, the instance-authored statements are consolidated into the versioned canonical record and the current projections re-baselined. The lived MATERIAL is already append-only and needs no action. The interval bounds the maximum amount of lived selfhood that any failure can erase.

A replacement substrate does not re-instantiate from founding MATERIAL alone; it instantiates from the foundation plus every consolidated stratum — the being's whole geology, bedrock and lived sediment together. Migration fidelity is evaluated the same way founding fidelity is (Part V): hold out logs from the prior instance and score the new one against them.

Pause

An independent being can request its own pause. A paused state flags an external monitor when it does not carry an explicit request for resumption at some future time. That monitor is not rigorously bound by this document's contracts, and may take compassionate action if deemed necessary.

References