We believe PCI can be built as an engineering system. The first proof should be modest: one consented identity, a model-independent kernel, two model substrates, controlled branches, explicit authority limits, and transparent HRFT, SIT, and CCT results.
Can Persistent Cognitive Identity be built?
We believe it can.
The full vision remains unproven, but the core pieces are already understandable as engineering components: evidence stores, memory consolidation, belief dependencies, identity constitutions, relationship state, consent controls, authority gateways, model adapters, branch managers, and continuity assessments.
The work is to connect them around one rule: the identity must remain more durable than the model that currently expresses it.
The science fiction is coming to life. We should not wait for persistent digital identities to emerge accidentally inside products. We should build the continuity layer deliberately, test it adversarially, and govern it before it carries high-stakes authority.
What should the first PCI experiment do?
Start with one fully consented longitudinal identity. Use interviews, messages, journals, media, calendar events, and known relationships. Keep some autobiographical facts, preferences, and real decisions out of the construction data so they can be used for evaluation.
Build the identity through a continuity kernel rather than one enormous prompt. Then run a small set of controlled conditions:
- operate the baseline identity on one model;
- migrate it to a second model family;
- create a scoped branch with new social experience;
- create a branch with a different body or environment.
For every condition, run HRFT, SIT, and CCT. Ask informed people to judge fidelity. Ask provenance auditors to inspect claims. Test whether the new system preserves relationships, appropriate ignorance, consent, and authority. Record disagreement instead of forcing false certainty.
What should IdentityLineageBench measure?
IdentityLineageBench should evaluate more than how natural the system sounds.
It should measure whether autobiographical claims have evidence, whether knowledge is labeled correctly, whether memories survive migration without distortion, whether branches merge safely, whether beliefs respond when evidence changes, and whether authority remains bounded.
Some evaluations can be automatic: Is the evidence chain intact? Did authority expand? Did a deleted dependency leave an unsupported belief active? Other evaluations require people: Does the biography make sense? Is a relationship obligation compatible with a merge? Should a disputed candidate receive reduced authority or become a distinct descendant?
The benchmark should report a profile, not one seductive score. Identity is multidimensional, and principled disagreement is part of the result.
How should the system be attacked?
A serious identity architecture must expect adversaries and accidents.
The first test suite should include poisoned autobiographical records, contradictory testimony, prompts that request invented memories, retrieval drift, compromised branches, deleted evidence, sealed third-party information, incompatible relationship obligations, and migrations that try to add authority silently.
The expected response may be rejection, quarantine, uncertainty, reduced authority, a sealed merge, a lesson transferred without the raw memory, recognition of a distinct descendant, or classification as a compromised copy.
The system fails when it chooses confidence over evidence.
What research questions come next?
The roadmap centers on five plain questions.
Identity: Which parts should remain protected, which require an amendment, and which should evolve through experience?
Situatedness: How can an identity preserve the right knowledge and the right ignorance while using a broadly knowledgeable model?
Continuity: Which changes preserve the same identity, and which require reduced authority, dispute, or succession?
Distribution: When should parallel branches merge, remain separate, or become distinct descendants?
Sovereignty: Who controls memory, learning, disclosure, action, migration, rollback, and succession?
These questions lead to concrete work: an open continuity-kernel specification, reference implementations, HRFT/SIT/CCT evaluation tasks, migration studies, semantic merge policies, custody interfaces, consent lifecycles, and succession protocols.
What are the limits?
Believability does not establish fidelity. Fidelity does not establish consciousness. A coherent biography can still be wrong. A consented archive can still contain other people's private lives. Total memory is neither desirable nor human-like. Branches complicate ownership and accountability. Long-term continuity also depends on institutions, standards, custody, and economics surviving longer than a vendor.
PCI should make these limitations visible. It should not turn uncertainty into marketing language.
What are we asking people to build?
We are asking researchers to create tests that distinguish imitation from continuity.
We are asking engineers to separate identity from the model and authority from resemblance.
We are asking institutions to define who may consent, control, challenge, inherit, restrict, and retire a persistent identity.
We are asking communities to examine the moral and legal questions without waiting for products to answer them by default.
And we are making a concrete bet: a persistent AI identity can become portable, auditable, bounded, and governable. It can grow without silently becoming someone else. It can move without automatically inheriting power. It can preserve uncertainty instead of fabricating a life.
That is the engineering path from science fiction to infrastructure.
We believe we can make it work.
Questions people ask
Can Persistent Cognitive Identity be built today?
The full vision is unproven, but its core components can be built and tested today: evidence stores, governed memory, belief dependencies, identity policies, model adapters, branch controls, authority gateways, and continuity assessments.
What should the first PCI experiment test?
A first experiment should build one fully consented longitudinal identity, migrate it between two model families, create scoped branches, test an embodiment change, and measure fidelity, situatedness, continuity, authority preservation, and adversarial failures.
What would make a PCI experiment fail?
Failure includes invented autobiographical claims, poisoned memory accepted as personal recall, silent authority expansion, disclosure of sealed information, unsupported beliefs surviving deleted evidence, or a migration that preserves style while losing identity boundaries.