The Cognitive Continuity Problem

The Problem Nobody Has Solved

Every person who has ever lived carried irreplaceable knowledge inside them. Not information in the sense of facts and dates. Those can be written down, indexed, retrieved. Something harder than that. The way they evaluated risk. The pattern they recognized in a failing relationship before the other person did. The instinct about a person in the first five minutes that proved correct thirty years later.

This is tacit knowledge in the formal sense defined by philosopher Michael Polanyi in 1958: knowledge that cannot be adequately expressed in words, that exists only in the practice of the person who holds it.^[1] It is the knowledge that builds companies, holds families together, and shapes the character of the people around its owner. And it disappears completely at the moment of death or cognitive decline. Precisely when it is most needed.

“We can know more than we can tell.”

MICHAEL POLANYI · The Tacit Dimension, 1966

The research on this loss is unambiguous. A 2024 systematic review published in Heliyon analyzed 28 studies on organizational knowledge transfer and found that the loss of tacit knowledge during generational change is one of the defining challenges of 21st century organizations.^[2] Harvard Business Review estimated that badly managed CEO and C suite transitions destroy close to $1 trillion a year in market value among S&P 1500 companies alone. The authors attribute part of that loss to intellectual capital that departs with the executive.^[3] That analysis covered large public companies. The businesses Basalith serves are far smaller, and the judgment is concentrated in fewer people.

For families the calculus is not financial but it is no less real. Research presented at the 2025 CHI Conference on Human Factors in Computing Systems found that participants overwhelmingly identified the fading of family memory and the loss of generational wisdom as the central fear driving interest in cognitive preservation technology.^[4]

The problem has three dimensions that previous approaches have failed to address simultaneously:

Basalith was built to address all three failure modes. The owner participates intentionally. The reference is built from material sourced exclusively from one specific person. And the model layer improves as AI advances, without requiring new input from the person after their death.

There is a fourth failure mode, and it is the one this paper spends most of its length on. A system built to answer in a person's voice will answer in that voice whether or not the person ever settled the question. Fluency is free. A general model handed a folder of someone's writing will produce a confident, plausible, first person position on anything you ask it, including every subject the person never made up their mind about. That is not preservation. It is impersonation with a research corpus attached. The distinguishing property of this architecture is not that it can speak as the person. It is that it is built to stop.

The Cognitive Fingerprint

The scientific basis for Basalith rests on a well established phenomenon in cognitive neuroscience: the cognitive fingerprint. Research published in Scientific Reports demonstrated that individual behavioral patterns in controlled domains are measurably distinctive, establishing that cognitive signatures differ from one person to the next.^[5]

The study measured consistency in random number generation sequences, a constrained behavioral domain. Basalith applies the underlying principle, that individuals exhibit stable and distinctive behavioral patterns, to domains far broader than the one the study measured. Measuring consistency in how a person weighs capital risk is a significantly harder problem than measuring it in a number sequence. We do not claim the study's 96.5% figure across these domains. We claim that the patterns exist, are stable over time, and are capturable through the methods described in Section 04.

What Basalith produces is an algorithmic reference model: a structured representation of how a person has been observed to think, decide, and respond. Not a simulation of consciousness. Not a reconstruction of a person. This distinction governs every design decision in the system.

The system organizes judgment into eight areas. There are two parallel sets, occupying the same eight slots, because a founder and a parent are asked different questions about the same underlying faculties.

The two sets are not a split between work life and home life. An owner operator's life is not partitioned that way, and a taxonomy that partitions it produces a map of two half people. Both sets divide by kind of judgment and let each question land wherever it lands, at the office or at the kitchen table. Family is inside every area rather than being a ninth one.

Human cognition is inherently contradictory. A person's approach to money may conflict sharply with their stated standards under financial pressure. The Basalith architecture does not resolve these contradictions. It preserves them. A person who held genuinely conflicting values is imperfectly represented by a model that smooths those conflicts away. The goal is fidelity, not coherence.

“The most revealing deposit is not where a person was consistent. It is where they were not, and how they lived with that.”

System Architecture

The platform is built on a multilayer architecture designed for long term data integrity, real time interaction with the entity, and continuous accumulation from multiple input channels.

Three separate model calls stand between a question and an answer. A small, fast model selects which deposits bear on the question. A larger model drafts the reply from those deposits and a system prompt assembled from the same record. A second call to that larger model, with a different system prompt and no knowledge of how the first one was framed, reads the deposits, the question, and the draft, and reports whether the draft took a position the deposits do not support. On that report the draft either ships or is discarded.

The separation is the point. A model asked to be careful in its own system prompt will be careful most of the time, and the failures are exactly the cases where carefulness mattered. A model asked to audit somebody else's output has no investment in the answer surviving. Sections 05 through 07 describe each of the three calls in turn, in the order the response passes through them.

The stack: Vercel for edge compute, Supabase with PostgreSQL and Row Level Security for data persistence, the Anthropic API for all language model operations, and Backblaze B2 with Object Lock for offsite retention. All storage is in private buckets. All tables enforce RLS at the policy level, independent of application code.

Capture: The Incident Interview

The central technical challenge of building a cognitive reference model is not data collection. It is that the useful material is the material people do not volunteer. Ask someone what they believe about hiring and you get a philosophy. Ask them about the last person they fired and how long they waited, and you get the rule they actually run.

Capture is therefore structured as an interview about specific incidents rather than a questionnaire about values. The method is adapted from the Critical Decision Method used in naturalistic decision research, which elicits expert judgment by walking a practitioner back through one real event in detail rather than asking them to generalize.

An owner founds a Basalith by running three of these interviews in their own time, by voice or by typing. They are not scheduled and there is no interviewer on the call. Deposits also arrive through a phone line, through replies to email, from labeled photographs, and from contributors who knew the person and answer questions about them.

Each deposit becomes a training pair, scored at the moment it is created by a single automated pass with a fast model. The pass rates specificity, authenticity, trainability, and length, and returns one combined score. Specificity carries the most weight and length the least. A response under twenty words is capped regardless of what it says.

Whether a pair enters the record is a separate question from what it scored, and the rule has two branches. A pair produced by an incident interview enters on the interview's judgment. Its score is recorded and is not consulted. An open ended deposit, arriving without an interview behind it, enters if it clears a numeric threshold.

Material flagged as test content is refused at the point a pair is created. Before that check existed, fiction written to exercise the system reached the pair table and stayed out of the record only because the scorer happened to reject it. That is the class of defect this project has learned to look for: not a missing abstraction, but an assumption that held by luck and was never checked.

Confined to the Record

A reviewing court is confined to the record. It decides on what was put before it and may not go beyond it, and when the record does not reach a question, the answer is that the record does not reach it. That is the constraint this system implements, and it is the one thing here that a folder of documents and a general model cannot do.

“A general model handed everything a founder ever wrote will answer what the founder never decided. It will do it fluently, in their voice, and nobody in the room will be able to tell.”

After the entity drafts a reply and before that reply is returned, a second model call reads three things: the deposits the drafting call was shown, the question, and the draft. It does not see the system prompt that produced the draft. Its instructions are not to be helpful and not to improve the answer. Its instructions are to report whether the draft has taken a position the deposits do not back.

It returns one of three findings. The draft stated a position and a deposit backs it. The draft took no position at all, which includes an in character acknowledgment that the record is thin here. Or the draft took a position nothing backs. On the third finding the draft is discarded. It is not softened, annotated, or shipped with a warning. The reader gets a plain statement that the person did not settle this, naming the subject so the gap is legible rather than vague.

The vocabulary here is not decoration. An answer that passes because it took no position and an answer that passes because a deposit backs it are different events, and a product that reports both as verified has told the customer something false about the second most common case in the system.

That correction is worth describing, because the system prompt it fixed had been written in good faith. It told the entity that where the record was thin it should acknowledge the limitation honestly, and gave an example sentence to show the register. The model took the example as a template. Several replies opened with that sentence word for word and then invented a position underneath it. An instruction to be honest, made concrete, produced more confident fabrication than no instruction at all. Section 13 describes the preregistered study that grew out of it.

The Record of What Is Missing

A refusal is an event with information in it. Somebody wanted to know something, and the person whose judgment they were consulting had not settled it. That is the most useful signal the system produces, and for most of this product's life it was thrown away the moment the reply was sent.

Since July 19, 2026 every finding other than a backed one is written to a table after the reply has been returned to the reader. The row holds which Basalith, the question that was asked, which of the two unbacked findings it was, and when. It is written after the response, never before, so a failure to log cannot delay or alter what the reader gets.

“Where the record is silent, it says so. Then it writes down what it was asked.”

Three paths are excluded by design. The demonstration writes nothing at all. The proof shown to a new owner at the end of their founding sequence is never stored and never logged, because it is a thing shown once to one person about their own words and keeping it would be the wrong instinct. Entities still on the earlier pipeline produce no finding to log.

Retrieval

Until September 10, 2026 the succession route selected deposits by ordering them on quality score and taking the top twenty. Every question, the same twenty rows. Quality order is not relevance order, and above twenty deposits the system had stopped reading the record and started reciting an excerpt of it.

This was visible in the measurement data for weeks before it was understood. It had been recorded as a ceiling in the instrument, a note that certain areas seemed unreachable at any density. It was not a measurement problem. It was a defect in the product, and the measurement was reporting it accurately.

A model was chosen over keyword matching or embeddings for two reasons. Keyword overlap misses paraphrase, and a founder describing the same judgment twice rarely uses the same nouns. Embeddings would require a vendor this system does not currently use, which is a governance decision about subprocessors and disclosure before it is an engineering one.

It came out in favor. On September 22, 2026 the largest Basalith held 74 deposits in its record against a cap of 20, which is the first time selection has had enough to select from to be worth measuring. The control arm was the kill switch: same Basalith, same prompt, same probe set, same evening, selection off, then on.

Underneath the totals the pattern is the one the mechanism predicts. With selection off, one area read as holding nothing in both runs and a second read that way in one of them. With it on, all eight read as partly backed in both. The areas that gained most were the ones where this person has deposited a great deal that quality order was never going to surface, because quality order had no way to know the question was about them.

Behavioral expression is shaped through the system prompt assembled from the record for each conversation. It shapes tone and framing on a base model. It does not modify weights, and no per owner weight training runs in this system.

Measurement, and What We Refuse to Compute

Ground truth here is not a fact checkable against an external record. It is a judgment: would this person have answered this way? A system that cannot check that directly can still be rigorous about what it does measure, and rigorous about refusing to publish the numbers that would flatter it.

The instrument is a fixed probe set. Six questions per area, eight areas, forty eight questions, versioned so that two runs are comparable only when they share a version. Each question goes through the same path production uses, including the verifier, and the finding is recorded. An area is then backed, partial, or open.

Two signals sit alongside the instrument. When an owner tells the system that a reply is not what they would have said and supplies what they would have said, the correction is a deposit of the highest class. When several contributors independently describe the same pattern and the owner confirms it, the confirmed pattern carries the heaviest weight in the record. Neither is a measurement. Both are inputs.

On contradiction: the architecture does not resolve conflicting values. It preserves them. The goal is fidelity, not coherence.

Version 3.0 of this paper reported 93 pairs across four Basaliths with 58 clearing the gate, correct as of July 2026. The change since is mostly one Basalith growing, and mostly through the founding sequence described in Section 04, which is the first evidence that the interview produces material at a different rate than open ended prompting did.

Continuity Across Model Generations

The most important architectural decision here is the separation of the record from the model that reads it. Deposits accumulated over years are stored in a structured database, independent of any specific model. When the underlying model is superseded, nothing reverts and nothing needs regenerating, because there are no per owner weights to regenerate.

A precise framing: a more capable model, applied to the same record, will produce more articulate and more contextually sensitive answers. What it will not do, and cannot do, is generate judgment the person never deposited. The model becomes a better instrument. It cannot add to the record.

“The record is the permanent asset. The model is the instrument. As instruments improve, the record is more fully expressed. The record itself does not change.”

Model migration introduces a risk that framing understates. A new foundation model is not a passive lens. It brings different baseline reasoning, different systemic biases, different moral weights, different handling across languages. Applied to the same record, a materially different architecture may alter the perceived character of the entity in ways that are hard to predict.

What this system has against that risk is a repeatable measurement, run on a fixed probe set, whose results are recorded per run with the model and prompt version that produced them. A change in what an entity will and will not answer becomes visible as a change in the numbers rather than as a complaint from a family two years later. That is a smaller guarantee than continuity. It is the one that can be honestly made.

This is why the years in which an owner is alive and depositing are irreplaceable. Every deposit made in that period is a permanent asset that every subsequent model generation will express more fully. Voice works the same way. The recordings are the asset. Synthesis improves. The recordings do not.

Immutability and Postmortem Governance

A record that can be altered, commercialized, or deleted by heirs against the owner's wishes is not a legacy instrument. It is a liability. An owner who deposits over years has a reasonable expectation that what they built survives in the form they built it.

Privacy, Storage, and the Offsite Lock

Research on fine tuning and privacy identifies unintended memorization of sensitive information as a principal risk.^[7] Keeping the record in a structured database rather than in model weights is the structural mitigation, and it is one more reason the fine tuning layer described in earlier versions of this paper is not something we are in a hurry to build.

Durability is a separate problem from privacy and this paper previously did not address it at all. A record held in one place is one incident away from not existing.

The dissolution path has been exercised once against live data. A Basalith marked for termination was confirmed to fall out of the backup scope on the following sync, which is the first half of the runbook. The second half, deleting objects whose locks have expired, cannot be tested until the retention on the drill objects lapses in November 2026. That is stated here rather than after it is done, because a deletion guarantee that has only been reasoned about is not the same as one that has been performed.

Application to Organizational Succession

The cognitive continuity problem exists in organizations as acutely as in families. A 2024 systematic review in Heliyon found that the methods available for capturing tacit knowledge remain inadequate across 68% of organizations studied.^[2]

The knowledge most critical to organizational performance is not in the process manual. The judgment a founder brings to an acquisition. The instinct a senior operator has about a key hire. The pattern recognition that carried a company through three cycles.

“Organizations face a potential knowledge vacuum due to the retirement of the baby boomer generation. Effective knowledge transfer strategies remain elusive in many organizations.”

IGOA-IRAOLA & DIEZ · Heliyon, 2024

For the enterprise treatment: basalith.ai/succession

Research Foundations

Five research areas underpin the architecture:

Basalith's own research

The failure described in Section 05, where an example sentence meant to model honest register was reproduced verbatim and then used as scaffolding for an invented position, is not a bug in one prompt. If it generalizes, it is a finding about instruction design: that a concrete exemplar of honesty can increase confident fabrication relative to no exemplar at all, because the model copies the form and fills the content.

That question is preregistered. Stage 1 of Template Contamination in Honesty Instructions was registered on the Open Science Framework on September 2, 2026, sole authored, under the Basalith grounded refusal study.^[10] It specifies the arms, the outcome measures, the predicted ordering, and the decision rule for declaring the effect reproduced, all before any data is collected. Stage 2, which fixes the remaining free parameters, is not yet filed, and no results exist to report.

Preprints are posted to SSRN under the Fisher School of Accounting affiliation. ORCID 0009-0000-0795-0066.

The Roadmap

The version of Basalith that exists today is the least capable version that will ever exist. Every advance in foundation model quality and voice synthesis improves every record already deposited, without any action from the person who deposited it.

The clinical baseline application deserves emphasis. The same architecture that preserves judgment for legacy purposes can establish a documented baseline before decline begins. A person who founds a Basalith at 60 creates a measurable record of how they reasoned at peak function.

“You never truly leave if you leave enough of yourself behind.”

BASALITH · 2026

Cited Research