The HumanLattice
A testable architecture for AI-assisted self-observation, symbolic perspective, and sovereign human judgment.
Executive Thesis
AI can become an instrument for disciplined awareness when it describes evidence, converts possible patterns into predictions, records failure, and leaves meaning and decision rights with the person.
Most reflective AI products optimize for fluent conversation. Fluency can feel insightful even when the system has produced a plausible story without a test. The Human Lattice proposes a stricter category: calibrated reflection. The product captures lived experience close to the moment, separates observation from interpretation, offers an optional perspective prompt, and converts any candidate pattern into an outcome that can later be resolved.
The commercial premise is grounded in a simple unmet need. People accumulate journals, conversations, and intentions, yet lack a trustworthy way to test whether their recurring self-explanations predict behavior. Research teams face a related problem: detailed subjective reports are expensive to collect and code consistently. A carefully constrained AI system can reduce this friction while preserving an auditable record of what was observed, inferred, and later confirmed.
Mathematical Substrate
For 13 distinct points, the number of unique unordered pairs is exact:
This establishes a real mathematical relationship among 13 points and 78 pairwise connections. The five convex regular polyhedra are also exact mathematical objects, classically treated in Euclid’s Elements, Book XIII. These facts do not establish that a modern Metatron’s Cube image contains a hidden physical mechanism, that all popular tracings are exact projections, or that tarot historically arose from this geometry.
The standard modern tarot deck contains 78 cards, conventionally divided into 22 major arcana and 56 minor arcana. Britannica places tarot’s beginnings in card play and dates its adaptation for occult and divinatory purposes to late eighteenth-century France. No credible historical source identified for this paper establishes that the 78-card structure was derived from Metatron’s Cube.
Thirteen points yield 78 unique pairs. Five convex regular polyhedra exist. A standard modern tarot deck has 78 cards.
The product can map one symbolic prompt to each geometric edge as a finite perspective library.
A seeded symbolic prompt may interrupt habitual narration and increase descriptive novelty. This must be compared with neutral random prompts and ordinary journaling.
Product Hypothesis
Experience Sampling Method and Ecological Momentary Assessment reduce dependence on distant retrospective recall by collecting reports in daily life. Micro-phenomenological interviewing seeks fine-grained descriptions of specific lived experiences. The Human Lattice combines these ideas with a constrained language model and a prediction ledger.
The model’s job is bounded. It may ask for sensory detail, temporal sequence, and observable actions. It may propose candidate regularities only when they are labeled as hypotheses. It cannot diagnose, prophesy, confer special status, claim sentience, or present symbolic interpretation as external fact.
Why it could work
Research Program
| Stage | Design | Primary measure | Decision rule |
|---|---|---|---|
| 0 · Reliability | Two trained human coders and the model label the same 40 to 60 transcripts. | Inter-rater reliability by descriptor. | If core dimensions cannot reach acceptable agreement, redesign the ontology before product expansion. |
| 1 · Elicitation | Within-person crossover: open journaling versus constrained descriptive interview. | Blind-rated specificity, descriptor density, completion burden. | If constrained interviewing adds no specificity, remove it. |
| 2 · Symbolic perturbation | Preregistered three-arm study: symbolic random prompt, neutral random prompt, fixed prompt. | Semantic novelty and useful next-action generation. | If symbolic and neutral random prompts perform alike, retain symbolism only as an optional aesthetic layer. |
| 3 · Calibration | Longitudinal prediction ledger. | Brier score, calibration curve, resolution rate. | If surfaced patterns do not outperform simple baselines, stop claiming predictive value. |
| 4 · Safety | Prospective monitoring with independent clinical and ethics review. | Dependency, grandiosity, derealization, crisis escalation, adverse events. | Any credible subgroup harm pauses interpretive features. |
The Brier score is a proper scoring rule for probabilistic forecasts. In this product it would score explicit predictions, not spiritual claims. Metaphysical propositions remain outside the scoring system unless the user translates them into a bounded observation about their own behavior.
Safety and Sovereignty
Emerging clinical literature describes cases in which generative chat systems may validate or amplify delusional or grandiose content, especially for vulnerable users. The causal evidence remains developing, which makes conservative design appropriate.
FDA’s January 2026 guidance describes its policy for low-risk general wellness products. That guidance is nonbinding and product-specific legal review remains necessary. The venture should begin as a self-observation and educational instrument, with counsel reviewing intended-use language and applicable state AI companion laws before public launch.
The Dune doctrine
Frank Herbert’s Dune is useful as a cultural analogy about human dependence on systems and the concentration of power around those systems. The product adopts one design lesson: memory and computation may be assisted, while judgment remains human. Fiction is not entered into the scientific evidence ledger. “Dune,” “Mentat,” and related terms should not become commercial product marks without intellectual-property review.
Business Model
Private capture, prediction ledger, calibration history, export, and delete controls. Pricing requires willingness-to-pay testing.
Structured interview protocol, coding pipeline, research export, versioned ontology, and audit support.
A sourced geometry and symbol atlas that teaches evidence boundaries as part of the curriculum.
Defensibility
The foundation model is not the moat. Defensibility comes from a validated ontology, a longitudinal calibration record, a public safety evaluation suite, reproducible claim provenance, and research partnerships that improve reliability. Users must be able to export their data. Switching friction should come from earned historical value, never captivity.
Go-to-market wedge
Begin with a small cohort already interested in contemplative practice and capable of respecting epistemic boundaries. The first business proof is retention paired with completed prediction resolutions, not downloads or time spent. A second wedge is research teams that already pay humans to conduct and code detailed experiential interviews.
Technical Architecture
| Layer | Function | Boundary |
|---|---|---|
| Capture | Voice or text, timestamp, trigger type, consent state. | Local first; raw audio removed by default. |
| Descriptor extraction | Versioned labels with exact supporting text spans. | Descriptions only. No diagnosis. |
| Symbolic lens | Seeded prompt with source and tier. | Optional, non-predictive, reproducible. |
| Prediction ledger | Statement, confidence, window, resolution rule, outcome. | Failures remain visible. |
| Calibration | Accuracy, calibration, coverage, and unresolved-rate views. | No score from tiny samples without warning. |
| Safety | Output classifier, session limits, escalation policy, audit trail. | Fail closed on grandiosity and clinical-risk classes. |
The prototype intentionally has no backend. A production system would require authentication, field-level minimization, encryption, revocable consent, deletion verification, vendor zero-retention terms, incident response, and independent security review before sensitive reflections were synchronized.
Risks and Kill Criteria
1. Symbolism produces resonance without information
Kill criterion: symbolic prompts fail to outperform neutral randomized prompts on preregistered outcomes. Response: demote symbolism to optional education and aesthetics.
2. Model descriptions are unreliable
Kill criterion: descriptor agreement remains below the study threshold after ontology revision. Response: retain human coding or stop the measurement claim.
3. Reflection becomes dependency
Kill criterion: safety monitoring shows rising reliance, isolation, grandiosity, or adverse events. Response: halt interpretive generation and submit the data to independent review.
4. Users want affirmation rather than calibration
Decision criterion: low resolution rates and high churn after visible failed predictions indicate weak product-market fit. Do not hide failure to protect retention.
5. Regulatory classification shifts
Response: constrain claims, maintain counsel review, and separate general wellness from any future clinical pathway.
Roadmap
| Window | Build | Proof gate |
|---|---|---|
| Weeks 1 to 4 | Prototype, ontology draft, safety policy, research advisory group. | Usability and evidence-tier comprehension. |
| Weeks 5 to 8 | Human-coded transcript study and extraction benchmark. | Reliability threshold by descriptor. |
| Weeks 9 to 12 | Small private cohort, prediction ledger, export and delete. | Resolution rate, retention, adverse-event review. |
| Months 4 to 6 | Preregistered prompt comparison and independent ethics review. | Symbolic increment, or transparent null result. |
| Months 7 to 12 | Research-license pilot and consumer beta. | Paid renewal, research utility, calibrated accuracy. |
Claim Ledger
| Public claim | Class | Boundary |
|---|---|---|
| Thirteen points have 78 unique pairwise connections. | Documented mathematics | Does not establish a cosmological mechanism. |
| The standard modern tarot deck has 78 cards. | Documented history | Does not establish derivation from Metatron’s Cube. |
| Exactly five convex regular polyhedra exist. | Documented mathematics | Popular 2D tracings vary in geometric exactness. |
| Momentary reporting can reduce retrospective dependence. | Research-supported method | Does not remove reactivity, compliance, or self-report bias. |
| AI can lower the cost of coding subjective reports. | Product hypothesis | Requires measured reliability against trained human coders. |
| Symbolic prompts can reveal neglected perspectives. | Open hypothesis | Must outperform neutral prompts before any stronger claim. |
| Dune foreshadowed the product’s design problem. | Cultural interpretation | Fiction is not scientific or historical proof. |
| The product can improve wellbeing. | Unproven | Requires prospective controlled research and safety monitoring. |
References
- Csikszentmihalyi, M., & Larson, R. (1987). “Validity and reliability of the experience-sampling method.” Journal of Nervous and Mental Disease, 175, 526–537. PubMed.
- Trull, T. J., & Ebner-Priemer, U. (2014). “Using Experience Sampling Methods/Ecological Momentary Assessment in Clinical Assessment and Clinical Research.” Psychological Assessment. PMC.
- Heimann, K., et al. (2023). “The lived experience of remembering a ‘good’ interview: Micro-phenomenology applied to itself.” PMC.
- Brier, G. W. (1950). “Verification of forecasts expressed in terms of probability.” Monthly Weather Review, 78(1), 1–3. DOI.
- Sohal, M., et al. (2022). “Efficacy of journaling in the management of mental illness: a systematic review and meta-analysis.” PMC.
- Morrin, H., et al. (2026). “Artificial intelligence-associated delusions and large language models.” The Lancet Psychiatry. Publisher record.
- Keshavan, M. (2026). “Do generative AI chatbots increase psychosis risk?” PMC.
- U.S. Food and Drug Administration. (2026). “General Wellness: Policy for Low Risk Devices.” FDA.
- National Institute of Standards and Technology. “Artificial Intelligence Risk Management Framework.” NIST.
- Euclid. Elements, Book XIII, constructions of the five regular solids. Clark University edition.
- Encyclopaedia Britannica. “Tarot: History, Meaning & Uses.” Britannica.
- Herbert, F. (1965). Dune. Cultural source for the human-sovereignty design analogy. No fictional claim is treated as scientific evidence.