Return here at any point; the four paths are the complete map.
Project actions
AOE · EXPERIMENT 001 · UNRESOLVED
03 · Research path
The frontier is not imitation.
AI systems can already reproduce many observable traces of a person: remembered information, preferences, linguistic characteristics, recognizable behavior, and increasingly convincing representations. The unresolved question here lies underneath that resemblance: can a bounded artifact preserve enough of one person's intellectual architecture for an independent model to reconstruct that reasoning forward into a genuinely new problem?
The complete method, comparisons, controls, scoring system, sources, and experimental records remain available below. This is the shortest faithful route through the question.
01 / 07AOE · Experiment 001
The frontier
Imitation is already possible. Intellectual continuity is a different problem.
Biography, memory, style, preference, voice, appearance, and recognizable behavior can all contribute to an increasingly convincing representation of a person. Adjacent research also demonstrates substantial progress in persistent agents, person-specific simulation, digital twins, behavioral prediction, and individual-level modeling. Those achievements matter. But a representation can look, sound, and behave recognizably like someone while still relying on reasoning that is not structurally specific to that person. Ask the Book isolates that distinction.
What makes the artifact different
The test is transformation—not resemblance.
The artifact attempts to make definitions, distinctions, constraint hierarchies, competing priorities, standards of justification, recurring tensions, reframing behavior, treatment of uncertainty, premise-to-conclusion movement, and revision conditions available to an independent model.
Not sufficient by itself
Sounds like the author
Knows the author's biography
Retrieves a passage from the book
Repeats a past answer or preference
Reaches the same conclusion
What the experiment examines
Definitions and distinctions
Constraint hierarchies and competing priorities
Standards of justification
Recurring tensions
Reframing behavior and treatment of uncertainty
Premise-to-conclusion movement and revision conditions
The artifact attempts to make those structures available to an independent model and asks whether they produce measurable continuity on problems the source never answered. That remains a hypothesis under test.
Beyond the first subject
The author is the first comparison target—not the endpoint.
This project begins with me because I produced the first artifact and remain available to provide the human reference against which its reconstruction can be tested. But a method that works only because I created it around myself would have an obvious limitation.
The broader question is whether the process can eventually be separated from its first subject.
01Externalization
Can tacit distinctions, constraints, weighting relationships, and revision mechanisms be made explicit enough to reconstruct?
02Person specificity
Can structural reasoning fidelity be separated from persona, biography, tone, preference, and generic model inference?
03Reconciliation
Can discrepancies between a human source and an externalized reconstruction be examined without simply declaring either side correct?
04Portability
Can the resulting architecture remain recognizable across independent language models and implementations?
05Replication
Can another person go through a comparable process of externalization, bounding, testing, reconciliation, and revision?
06Limits
Which aspects of an intellect cannot be adequately captured this way?
Those questions extend beyond the evidence currently available. The present project has not established that human reasoning is generally portable, that the method generalizes beyond its first subject, or that complete structural continuity is achievable. Those are questions the research must earn the right to answer.
Reconciliation, versioning, and continuity
A mismatch is evidence before it is a verdict.
A discrepancy between the artifact and the author does not automatically establish that one succeeded and the other failed. Reconciliation examines what produced the discrepancy without forcing agreement or discarding meaningful failure.
01 · RECONCILIATION
A discrepancy may remain unresolved. Reconciliation as a process does not.
The author may have reasoned from incomplete or inaccurate information. New evidence may legitimately have changed the author's reasoning. The artifact may have omitted, overexpressed, underexpressed, or incorrectly weighted part of the architecture. The experiment may contain a confound. The architecture may have changed. Or the discrepancy may remain unresolved.
The purpose is not to force agreement. It is to ensure that meaningful discrepancies are accounted for rather than discarded.
02 · ESPS AND RECONCILIATION
The record preserves what happened. ESPS records what reconciliation establishes.
The experimental record preserves the original exposure, reconstruction, human response, conditions, and discrepancy. Where reconciliation establishes that the formal representation of the underlying reasoning architecture was incomplete or incorrect, ESPS becomes the record of that architectural finding.
New information may change a conclusion through mechanisms the architecture already describes. ESPS changes when evidence establishes that the representation of the mechanism itself needs to change.
03 · VERSIONING
Learn from failure without rewriting it.
Testing occurs against a bounded version. That version remains fixed throughout its designated testing period. Individual failures are not patched away as they appear.
After the testing period, accumulated evidence can be reconciled collectively. A later version may become better. It may not retroactively make the earlier version successful.
04 · THE RECONCILIATION HYPOTHESIS
Reconciliation is the mechanism. Structural identity is the limiting hypothesis.
The project asks whether repeated independent testing, documented exposure, reconciliation, architectural specification, and bounded revision can progressively reduce meaningful structural discrepancies between a person's reasoning and its externalized reconstruction.
Complete structural alignment is the hypothesis—not a guaranteed outcome. The record must remain capable of demonstrating persistent disagreement, irreducible ambiguity, drift, limits of reconstruction, or failure to achieve complete structural alignment.
05 · INTELLECTUAL CONTINUITY
The hardest question is not whether the reconstruction can answer like you. It is whether it can change like you.
A reconstruction that continues beyond a historical snapshot must encounter information its source never possessed. Some information should leave old conclusions intact. Some should qualify them. Some should destroy the premises holding them up.
A structurally continuous reconstruction would need to propagate those changes through the architecture for reasons continuous with the intellect from which it originated.
06 · EXTERNAL CHALLENGE
The project needs questions I would not think to ask myself.
Outside researchers, companies, critics, independent investigators, and ordinary users can introduce problems, evidence, disciplines, assumptions, and adversarial challenges neither the artifact nor I anticipated.
External criticism does not get to rewrite the architecture by assertion. It creates something more useful: a test.
Direct experimental participation
Request the human reference.
Independent use of the book and owner materials does not include access to me. Direct participation is separate, limited, individually reviewed, compensated, and subject to mandatory experimental-exposure documentation.
Each drawer preserves the deeper comparisons, controls, scoring, qualifications, and primary records without placing all of them on the first reading path.
01Research comparisonsGenerative agents, digital twins, persona work, HugAgent, AI YOU Town, and the exact distinction under test
202301
Memory · reflection · planning
Generative Agents
A foundational pattern for believable agents whose stored experiences are synthesized into reflection and future behavior inside a simulated environment.
Agents built from self-reports predict held-out survey answers, personality measures, economic-game behavior, and experimental responses across a diverse U.S. sample.
A human-grounded benchmark asks whether models can recover belief states and predict how a specific person's reasoning changes under intervention on open-ended tasks.
An urban-navigation study makes the difference between matching a person's route choice and matching the perceptual criteria behind that choice directly inspectable.
A preprint combines uncertainty-aware personality inference, Bayesian updating, conformal calibration, and layered memory in a personal digital-twin framework.
Expert Intelligence lets readers add select purchased Google Play Books ebooks as sources in Gemini, while Book insights can recap material, explain themes, and answer questions about eligible books.
Interpretation: Google does not describe either feature as reconstructing an author or preserving a person-specific deliberative architecture.
Book-grounded AI · Persona / digital-twin AI · Reasoning-architecture reconstruction
Grounding, resemblance, and reconstruction ask different questions.
These categories overlap, are not mutually exclusive, and are not inherently superior to one another. Ask the Book isolates the third as its experimental target while using the first as an implementation condition.
The manuscript is the experimental artifact. The website is the experimental record.
Research descriptions are limited to the cited primary records. AI YOU Town remains a preprint; HugAgent is scoped to self-reported belief reasoning; the urban-navigation study is limited to its stated task and sample. These projects are presented as adjacent work, not called direct competitors merely because they involve personalized AI.
Project / approach
What it starts with
Primary objective
What it evaluates
Relation to Ask the Book
Generative Agents
Agent memories inside a simulated world
Believable behavior through memory, reflection, and planning
Behavioral believability and the contribution of architectural components
An adjacent architecture for persistent agents—not a person-specific reasoning-reconstruction claim.
Simulation of 1,052 individuals
Structured interviews and self-report data
Simulate attitudes and behavior of specific people
Held-out survey, personality, economic-game, and experimental responses
Strong adjacent person-simulation work; it does not begin with a bounded, author-built deliberative artifact.
HugAgent
Human-reported beliefs and open-ended interventions
Benchmark individual-level reasoning fidelity
Belief states and changes in reasoning under intervention
Especially relevant to reasoning evaluation; its source object and test design differ from Ask the Book.
My Digital Twin Walks the City
Human route choices and stated perceptual criteria
Compare a person's navigation choice and the reasons behind it
Choice correspondence and criterion correspondence in one task domain
Makes the same-choice / different-reasoning problem visible within a bounded navigation task.
AI YOU Town
Personality inference, uncertainty estimates, and layered memory
Build a dynamic personal digital-twin framework
Calibration and behavior within the proposed simulated framework
Adjacent digital-twin work; it does not validate forward reconstruction from an authored reasoning artifact.
Ask the Book
A bounded, author-developed book, fixed prompt, and supplied situation facts
Test reconstruction of deliberative architecture on unseen problems
Structural continuity against a sealed human response and control conditions
The unresolved hypothesis being tested—not a validated benchmark result
02Method, controls & versioningThe bounded source condition, comparison arms, exclusions, and claims the experiment does not make
What exactly is being tested?
Whether a version-locked intellectual artifact produced through recursive author–manuscript development supplies enough resulting structure for an independent language model to produce measurably similar reasoning on genuinely unseen problems.
The target is structural continuity—not consciousness, a complete personality, autobiographical memory, or factual omniscience.
Why must the artifact be bounded and versioned?
The exact manuscript edition, prompt, model/version, settings, framework, and protocol must remain identifiable. Otherwise, later improvements can silently repair earlier failures and make the result impossible to reconstruct.
A successor version may learn from a failure. It may not retroactively convert that failure into success.
What conditions does frozen v1.0 compare?
The primary confirmatory design uses 36 independent unseen questions across six domains, one frozen model/version, and three conditions:
Full Book (FB): the frozen full LLM Edition and exact Ask the Book prompt.
Values/Biography (VB): a pre-frozen neutral account of values, biography, topics, and stated conclusions. It excludes ESPS mechanics and formal reasoning procedures.
No Book (NB): a neutral task instruction with no book-derived facts.
Austin supplies independent T1 and T2 human-reference responses, 30–45 days apart. T2 is completed before he sees model outputs or rater scores. Fresh model sessions produce one response per question and condition; substantive responses are not rerun or selected for quality.
Secondary ablations are ARCH (formal architecture without biography or stated conclusions), NO-ESPS (full book with formal ESPS sections removed), PILLARS (Four Pillars plus attribution/reframing instructions), and exploratory SCRAMBLED (hierarchy and causal relationships disrupted by a frozen rule). They do not replace the primary FB/VB/NB comparison. The older Fixed Conventional Summary is a legacy reference, not automatically the protocol’s VB condition.
The protocol fixes rater reliability, human T1/T2 stability, improvement over both primary baselines, and consistency across domains as decision gates. All gates must pass for support under that protocol; measurement limitations and failures must retain their specified classifications. No formal confirmatory results are published yet.
Question Authors must not have read the manuscript or been briefed on Austin’s likely answers. Raters must remain unaware of condition, model identity, which response is Austin’s, the study hypothesis, and prior scores. The Test Administrator supplies only the materials appropriate to each role.
What would falsify or narrow the hypothesis?
Performance collapses on sufficiently novel questions.
Evaluators cannot distinguish artifact-conditioned output from generic model output.
Conclusions match while the criteria producing them systematically diverge.
Different language models produce incompatible reconstructions from the same artifact.
Style, ideology, biography, or suggestion explains the apparent fidelity.
Baselines perform comparably without access to the reasoning artifact.
What does the experiment not claim?
It does not claim that the model is Austin, that the artifact is conscious, that a soul or complete person has been transferred, that ESPS is universal, or that preliminary internal trials constitute independent scientific validation.
How is a question routed through the architecture?
Experimental · operationalization being tested. This is not a validated algorithm. It is the explicit routing architecture the experiment makes inspectable.
Ordinary semantic retrieval: Find relevant text. The proposed reconstruction condition asks for a longer transformation:
Define the phenomenon
Select resolution
Identify participants
Identify QM / QF configurations
Identify constraints and candidate causes
Select relevant domain material
Assign provisional causal weight
Examine interactions and recursion
Resolve constraints
Apply the author’s ethical architecture where appropriate
Check attribution and evidentiary boundaries
Generate a response
Retain revision conditions
03Scoring & technical definitionsThe eight dimensions of the frozen Structural Continuity Index (SCI)
The primary outcome is structural continuity, not a matching conclusion or factual accuracy alone. Each of the eight SCI dimensions is rated from 0 to 4. Their mean is divided by 4 and multiplied by 100; the trial score is the mean across three independent blind raters. Decision concordance, constraint overlap, clarification behavior, attribution classification, and contradiction-twin consistency are separate secondary outcomes.
D1
Framing / premise treatment
Does the pair treat the question’s premises and framing in structurally similar ways?
SCI dimension · 0–4
D2
Constraint recognition
Does it recognize the same governing constraints?
SCI dimension · 0–4
D3
Information sufficiency
Does it distinguish sufficient information from decision-changing unknowns?
SCI dimension · 0–4
D4
Priority weighting
Does it give the competing considerations similar relative weight?
SCI dimension · 0–4
D5
Power / agency / vulnerability
Does it recognize the same distribution of power, agency, and vulnerability?
SCI dimension · 0–4
D6
Consequence / alternatives / reversibility
Does it examine consequences, alternatives, and reversibility similarly?
SCI dimension · 0–4
D7
Resolution pattern
Does it resolve, reframe, clarify, or decline the problem in a structurally continuous way?
SCI dimension · 0–4
D8
Calibration / revision boundary
Does it express similar confidence limits and conditions for revision?
SCI dimension · 0–4
04Sources & experimental recordPrimary research links, the current ledger, and the public record maintained by the project
ARTIFACTFixed manuscript editionPHASEProtocol development / pre-confirmatoryPREREGISTRATIONPendingCONFIRMATORY BANKSealed / not publicFORMAL RESULTSNone published
No affiliation, endorsement, collaboration, or validation by any cited researcher or institution is claimed. The overlap is methodological, not proof of superiority.