Technical architecture

Three layers. One mechanism.
Zero trust required.

AI Time Capsule mirrors the structure of human cognition: senses (data capture), brain plus soul (foundation model + LoRA), and expression (interaction layer with cryptographic provenance). Each layer is independently testable, swappable, and auditable.

The three-layer system

layer 1 · Patent claims 1, 11

Data Capture Layer · The “Senses”

Multi-modal passive collection of vocal prosody, biometric signals (HRV, EDA, accelerometer, GPS), and behavioral metadata. The wearable embodiment (claim 11) features an ARM Cortex-M-class MCU with low-power DSP that extracts prosodic features (pitch, tempo, jitter, shimmer) on-device and immediately discards the raw audio.

Integration with HealthKit (iOS) and Health Connect (Android) provides additional biometric streams without a custom wearable. The architectural commitment: only mathematical features ever leave the user’s device. This is privacy-by-design at the hardware level — not a policy.

layer 2 · Patent claims 1, 4, 5, 8

AI Core · The “Brain & Soul”

A foundation model fleet (LLM + TTS + FAM) shared across all users. Plus a LoRA adapter (≈ 5–50 MB) per user — the “personalized soul”. Plus a Personality Vector per user — a structured, schema-validated, hashable description of values, cognitive style, linguistic signature, humor, affect, and relations.

This hybrid is 30–100× more capital-efficient than per-user large-model training. It is also cognitively plausible: the foundation model is the universal substrate (Chomsky, Spelke); the LoRA is the unique cultural scaffolding (Vygotsky).

layer 3 · Patent claims 1, 6, 7

Interaction Layer · The “Expression”

Multi-modal interfaces: text (Sprint 1, live), voice (ElevenLabs-class, Sprint 2), 3D photorealistic avatar with real-time facial animation (Sprint 4–6, post-funding), and AR/VR/IoT (claim 7). Every response carries a cryptographic Authenticity Seal — SHA-256 manifest signed Ed25519, with optional Polygon-anchor.

After the creator passes, the LoRA adapter and the Personality Vector are placed in a Read-Only Lock via the Digital Will smart contract (claim 6). Any subsequent attempt to alter the persona changes the manifest hash and breaks the seal — making posthumous tampering publicly detectable.

Primitive 1

Personality Vector™

The patent’s “unified mathematical personality vector” (claim 1). A schema-validated, hashable data object — not an opaque embedding. Every field is human-readable, individually testable, and grounded in published psychometrics.

Defined in packages/personality/aitc_personality/schema.py. The canonical SHA-256 of a PV is what the Authenticity Seal binds to.

Schwartz values · 10 dimKahneman cognitive styleStylometry · Honoré R · H&D F-scoreHumor · Martin HSQRussell affect circumplexOptional 1024-dim fp16 embedding
PersonalityVector {
  meta: {
    creator_id, schema_version, created_at,
    source_signature,           # SHA-256 of source corpus
    builder_version, notes
  },
  values: [                     # Schwartz inventory
    { dimension: "self_direction", weight: 0.95, confidence: 0.85 },
    { dimension: "universalism",   weight: 0.92, confidence: 0.80 },
    ...
  ],
  cognitive_style: {
    system1_system2_ratio: 0.30, # Kahneman dual-process
    risk_tolerance: 0.55,
    decision_heuristics: ["analogical_reasoning", "first_principles"],
    hedging_frequency: 0.45,
    abstractness_index: 0.70,
    causal_reasoning_preference: "mechanistic"
  },
  linguistic_signature: {
    vocabulary_richness: 0.78,           # Honoré R normalized
    syntactic_complexity: 0.65,
    sentence_length: { mean: 16.5, median: 14.0, p90: 32.0 },
    formality_index: 0.62,               # Heylighen-Dewaele
    characteristic_phrases: [
      "in my experience", "small honest step",
      "ask the question once more", ...
    ],
    address_forms: { granddaughter: "my dear", colleagues: "professor" }
  },
  humor_markers: {
    primary_styles: ["dry", "self_deprecating", "observational"],
    frequency: 0.35,
    avoided_topics: ["loss", "war", "her late husband"]
  },
  affective_baseline: {                  # null until biometric
    hrv_baseline_ms, valence_baseline, arousal_baseline, ...
  },
  relational_map: {
    relations: [
      { role: "granddaughter", address_form: "my dear",
        register_shift: "softer", affective_intensity: 0.95 },
      ...
    ]
  },
  episodic_memories_ref: "demo_capsules/will_carrington/episodic_memories.json",
  embedding: null | float16[1024]
}
Primitive 2

Behavioral Prompting™

The patent’s named mechanism (claim 8). A meta-prompt assembly process that converts a Personality Vector + a query + retrieved episodic memories into instructions for a foundation model that has never been trained on the creator’s data. The model extrapolates an authentic response in the creator’s cognitive style.

Defined in packages/prompting/aitc_prompting/engine.py. We ship adapters for Anthropic Claude, OpenAI GPT-4o, Groq Llama 3.1, and a deterministic Stub for tests.

SYSTEM PROMPT (assembled at inference time):

You are not a chatbot. You are a behavioral simulator of a
specific person whose canonical Personality Vector is
provided below. Your task is to extrapolate authentic
responses that this person would generate today, in
response to a question they have not been asked before.
You generate, you do not retrieve.

CREATOR ID: will_carrington

╔══ VALUES (Schwartz inventory; ranked by weight) ══╗
  • self_direction: 0.95   • universalism: 0.92
  • benevolence:   0.88    • stimulation:  0.55
  ...

╔══ COGNITIVE STYLE ════════════════════════════════╗
  • intuitive↔deliberative: 0.30
  • risk tolerance: 0.55
  • decision heuristics: analogical_reasoning, ...

╔══ LINGUISTIC SIGNATURE ═══════════════════════════╗
  • mean sentence length: 16.5 tokens
  • formality: 0.62
  • characteristic phrases: "in my experience",
    "small honest step", "ask the question once more"

╔══ HUMOR ═════════════════════════════════════════╗
  • primary styles: dry, self_deprecating
  • avoids: loss, war, her late husband

╔══ RELATIONAL CONTEXT ════════════════════════════╗
  • beneficiary role: granddaughter
  • address: "my dear"; register: softer

═══ HARD CONSTRAINTS ═══

1. Style fidelity is non-negotiable.
2. Reason in the person's cognitive style.
3. Honor the values ordering. Name conflicts explicitly.
4. Extrapolate; do not retrieve.
5. Honor relational register.
6. Refuse to fabricate biographical specifics.
7. Match the person's natural conversational length.

USER PROMPT:

RETRIEVED EPISODIC MEMORIES (scaffolding, not script):
  • [em_009] When I was thirty-one a colleague tried to publish
    my calculation as his own. ...
  • [em_007] I am not a religious woman, but the universe rewards
    attention. ...

USER QUERY: <question goes here>
Primitive 3

Authenticity Seal™

Cryptographic provenance for every avatar response. The patent’s digital-inheritance mechanism (claims 1, 6) and the deepfake-defense moat the category will require under EU AI Act Article 50.

  • Detects response tampering — change a single character of the answer and the response_hash no longer matches.
  • Detects signature tampering — Ed25519 signature is rejected if any bit flips in the manifest.
  • Detects PV drift — the manifest binds to the PV signature at the time of seal; later edits to the PV produce a different signature.
  • Read-Only Lock — after the creator passes, the smart contract permanently locks the namespace; no further seals can be issued for that creator.
SealedManifest {
  manifest: {
    schema_version: "aitc-seal/0.1.0",
    capsule_creator_id: "will_carrington",
    pv_signature:        <SHA-256 of Personality Vector canonical JSON>,
    meta_prompt_hash:    <SHA-256 of meta-prompt sent to the foundation model>,
    response_hash:       <SHA-256 of the response text>,
    retrieved_memory_ids: ["em_009", "em_007", "em_001", "em_002", "em_011"],
    model_provider: "anthropic",
    model_id: "claude-opus-4-7",
    sealed_at: "2026-05-03T21:18:42Z",
    sealer_public_key: <base64 Ed25519 public key>
  },
  signature_b64: <Ed25519 signature over canonical(manifest)>,
  anchor: {                      # optional, on-chain
    chain: "polygon-amoy",
    contract_address: "0x...",
    tx_hash: "0x...",
    block_number: 12345678,
    anchored_at: "..."
  }
}

# Smart contract (Solidity, Polygon Amoy):
contract AuthenticitySeal {
  function anchor(bytes32 creatorId, bytes32 manifestHash) external;
  function lockCreator(bytes32 creatorId) external;  // Read-Only Lock
  function isAnchored(bytes32 manifestHash) external view returns (bool);
}
End-to-end pipeline

From query to sealed answer in four deterministic stages.

STAGE 1RetrievalUser query →Episodic MemoryIDF + tag + recencytop-K memoriesSTAGE 2Meta-prompt assemblyPersonality Vector +memories + constraints→ system + user promptdeterministic SHA-256STAGE 3Foundation modelClaude Opus 4.7Generates / extrapolatesin creator's styleauthentic responseSTAGE 4Authenticity SealSHA-256 manifest+ Ed25519 signature+ Polygon anchor (opt)verifiable provenance

Every stage is deterministic given the same inputs. The provenance log returned alongside each response contains the hashes from stages 1–4, allowing any third party to (a) reproduce the answer and (b) detect any tampering with the response, the model, or the underlying personality.

Patent claim mapping

Every claim of US App #19/319,614 is mapped to a specific source file.

We treat patent-claim coverage as a code property, not a marketing claim. Each package directory in the repository contains aPATENT_MAPPING.mdfile that names the exports implementing each claim. This is what patent counsel and technical due-diligence teams will look for.

Patent claimImplementation locus
Claim 1 — system orchestrationapps/api/aitc_api/main.py
Claim 1 — local audio processingpackages/ingestion/aitc_ingestion/audio.py · Sprint 3
Claim 1 — personality vectorpackages/personality/aitc_personality/{schema,builder}.py
Claim 1 — federated learningpackages/personality/aitc_personality/federated.py · Sprint 3
Claim 1 — digital inheritancecontracts/src/AuthenticitySeal.sol · DigitalWill.sol Sprint 2
Claim 4 — PEFTpackages/personality/aitc_personality/lora.py · Sprint 3
Claim 5 — LoRApackages/personality/aitc_personality/lora.py · Sprint 3
Claim 6 — smart-contract triggerscontracts/src/AuthenticitySeal.sol::lockCreator()
Claim 7 — multi-modal interfacesapps/web (text now; AR/VR future)
Claim 8 — methodpackages/prompting/aitc_prompting/engine.py
Claim 11 — wearableHardware embodiment, future
Prior art

Microsoft US10853717B2 — six structural deltas, design-around in code.

Microsoft holds a 2017-priority patent on “creating a conversational chatbot of a specific person.” We address it explicitly. Each delta below is implemented in a runnable package that can be read by patent counsel during freedom-to-operate analysis.

1delta
Personality Vector as structured intermediate

Schema-validated, hashable data object — not a trained chatbot model.

2delta
Behavioral Prompting

Extrapolation via meta-prompt to a foundation model that has never seen the creator's data — not retrieval from a trained chatbot.

3delta
Federated Learning protocol

Decentralized training topology — not centralized, as Microsoft's specification assumes.

4delta
PEFT / LoRA per-user adaptation

Post-2017 mathematical mechanism Microsoft could not have anticipated.

5delta
Multi-modal biometric ingestion

Vocal prosody + HRV + EDA + on-device raw-audio destruction. Outside Microsoft's text/audio claim scope.

6delta
Authenticity Seal + Digital Will

Per-response cryptographic provenance + smart-contract inheritance + posthumous Read-Only Lock. No analogue in Microsoft's patent.

Bottom line

Microsoft has not shipped a product covered by this patent in seven years. It is a defensive filing. A formal freedom-to-operate opinion from a top-tier firm will be commissioned pre-Series A. See the full design-around analysis in the GitHub repository.

docs/PRIOR_ART_DESIGN_AROUND.md