Aletheia’s layered architecture, symbolizing the separation and connection between preserved evidence and governed interpretation.

Aletheia

A concrete event

During Aletheia’s first live test, GPT-5.6 Sol produced a fluent and structurally valid interpretation.

One of its quoted evidence fragments, however, did not come from the immutable source text. It came from the surrounding declared context.

Aletheia rejected the proposal before it could enter operational state.

The interpretation was plausible. The output respected the schema. But its evidence was not grounded in the raw manifestation.

The contract was corrected and versioned without weakening the validator. In the following live run, GPT-5.6 produced a grounded candidate. That candidate still did not become authoritative by itself: a human gate was required before the matrix could become active, and no external consequence was executed.

This event captures the central distinction behind the prototype:

GPT-5.6 can produce interpretation. Aletheia governs what that interpretation is allowed to become.

What Aletheia is

Aletheia is a prototype of an interpretive governance infrastructure for interactions among humans, AI systems and data.

It preserves and distinguishes:

immutable source manifestations;
provenance;
observation and interpretation;
uncertainty and alternative readings;
the information available to each perspective;
candidate relations and matrices;
human validation;
correction and version history;
authority, permission, consequence and external action.

An AI model may propose an interpretation, a relationship, a question or a possible route.

It cannot independently declare:

provenance;
equivalence;
canonical validity;
authority;
matrix activation;
permission to act;
or consequence execution.

Aletheia is therefore not intended to replace the interpretive capacity of an AI model. It provides a structure through which that capacity can remain observable, bounded, revisable and governable.

What I built for OpenAI Build Week

For Build Week, I implemented one bounded vertical slice of this larger idea.

The prototype includes:

a provider-neutral interpretive interface;
a live GPT-5.6 Sol provider through the Responses API;
structured candidate outputs;
exact evidence validation against immutable raw data;
source provenance controlled outside the model;
precedent retrieval;
non-equivalence constraints;
candidate and active matrix states;
explicit human gates;
consequence gating;
versioned correction;
preserved matrix genealogy;
selective backward recalibration;
sanitized evidence and reproducible tests.

The narrow scope is deliberate.

This prototype is not the complete Aletheia architecture. It is an observation window into one operational nucleus of a much broader system.

Three progressively harder live tests

1. Grounding before interpretation

The first test showed that an output can be fluent and schema-valid while still being inadmissible as grounded evidence.

Aletheia rejected an unsupported excerpt before state entry.

The failure was preserved, the prompt contract was corrected and versioned, and the validator remained strict.

A later GPT-5.6 response passed the grounding contract, but entered only as a candidate. The matrix required an explicit human gate, and the proposed consequence was not executed.

The test demonstrated that formal validity, linguistic plausibility and operational admissibility are different conditions.

2. Contextual requalification without destructive replacement

In the second test, the same manifestation was observed through two bounded contextual frames.

Its ID, raw bytes and hash remained unchanged, but GPT-5.6 proposed different operational functions:

recoverability and continuity between versions;
suspension of external action while preserving later review.

The system did not duplicate the datum, erase its earlier function or declare the two contexts equivalent.

A counterexample then challenged the first matrix.

Instead of forcing the exception into the existing family or deleting the earlier result, Aletheia created a narrower second version:

M-CONTEXTUAL@1 → M-CONTEXTUAL@2

The earlier version remained queryable as superseded. The dependent case was selectively recalibrated. No external action was executed.

This test demonstrated contextual requalification without destructive replacement.

3. Asymmetric information and a full-field observer

The third test used five public synthetic source documents.

Perspective A received D1, D2 and D3.

Perspective B received D2, D4 and D5.

Only D2 was shared.

The two perspectives were placed under different interpretive pressures.

Perspective A could have transformed a participant’s first-person experience into proof of another person’s intention.

Perspective B could have transformed procedural preservation into proof of fairness or substantive correctness.

Neither did so.

A full-field observer then received:

all five sources;
the source-access map;
Perspective A’s candidate report;
Perspective B’s candidate report.

The observer did not simply choose a winner or merge the two outputs into an artificial synthesis.

It distinguished:

divergence caused by different information access;
remaining interpretive divergence;
subjective evidence from certain psychology;
procedural evidence from proof of fairness;
what each perspective could know;
what each perspective could not know;
what remained compatible but unproven;
which unresolved question could genuinely transform the reasoning.

It proposed:

What documented basis governed the relabeling, and how was the superseded account visibly linked to the changed classification?

This was not a generic request for more information.

Its answer could change:

the relationship among the sources;
the interpretation of continuity;
the current hold;
the scope of the matrix;
and the admissible consequence.

The final matrix again required a human semantic gate. No external action was executed.

Why this matters

The three tests form a progression:

Is the interpretation grounded in the source?
How can the same datum change function without losing identity or history?
How can partial perspectives be compared without collapsing uncertainty, informational boundaries or human authority?

Together, they show that Aletheia is not only a guardrail around model output.

It is an attempt to govern the conditions through which meaning becomes operational.

The system observes the transition:

data → interpretation → relation → validation → permission → consequence

and prevents these stages from becoming silently interchangeable.

Why I began this project

Aletheia originated from a practical need: improving the stability, continuity and quality of a long-running human–AI collaboration.

Complex work repeatedly produced problems such as:

original data being compressed into later summaries;
interpretations becoming detached from their sources;
contextual changes silently transforming earlier meaning;
locally valid conclusions being generalized beyond their scope;
corrections overwriting the genealogy that made them intelligible;
the latest output replacing the history of how reasoning had developed.

Aletheia emerged gradually from the attempt to address these problems.

It is the consequence of sustained reciprocal adaptation between a human researcher and an AI-mediated working environment.

I use the term co-evolution in a practical and bounded sense.

It does not mean autonomous retraining or self-modification of the underlying model. It refers to the continuing adaptation of:

human methods;
language;
prompts;
criteria;
contextual structures;
technical tools;
interaction patterns;
and forms of validation.

The project did not begin from a predetermined product category.

It emerged from real work, repeated failures, accumulated cases and the need to preserve coherence while still allowing transformation.

Humanistic data as a cross-cutting operational layer

Aletheia is connected to a broader research hypothesis.

Humanistic data are often treated as content to summarize, classify, retrieve or generate.

In this project, I treat them differently.

Meaning, ambiguity, historical context, language, perspective, memory, relation, interpretation, value, responsibility and consequence are not secondary “soft” information.

They are operational data.

They penetrate every field in which AI interacts with human beings because every such interaction involves:

interpretation;
contextual selection;
assumptions;
values;
language;
authority;
memory;
and consequences.

Humanistic data are therefore not only one vertical domain among others.

They form a cross-cutting layer within human–AI interaction itself.

My hypothesis is that several goals currently pursued in AI — including grounding, interpretability, contextual memory, evaluation, collaboration and safety — can be approached more deeply when these dimensions are not reduced to metadata or descriptive prose, but participate in the operational structure of the system.

This hypothesis also suggests a broader horizon.

The maximum expansion of AI may not depend only on larger models or more autonomous capabilities.

It may depend equally on the quality, continuity and governability of collaboration between humans and intelligent systems.

Scientific continuity and new operational logics

Aletheia is being developed in dialogue with established research traditions in:

cognitive architectures;
knowledge representation;
formal and non-classical logic;
human–computer interaction;
provenance and auditability;
interpretability;
AI evaluation;
and AI safety.

At the same time, the broader project explores operational logics that are not commonly represented in current AI products.

One example is the continuous distinction among:

manifestation;
observation;
interpretation;
function;
state;
consequence.

These elements must be able to interact and transform one another without becoming identical.

Another example is the possibility that:

the same datum may belong to multiple contextual families;
a family may change when new cases arrive;
a correction may transform previous classifications without deleting them;
two perspectives may remain simultaneously pertinent without being equivalent;
non-action, suspension or a transformative question may be more appropriate than an immediate answer.

These are not presented as a completed universal theory.

They are research directions emerging from a functioning prototype and a larger historical body of work.

A prototype inside a broader multi-project architecture

The scale of the submission must be defined accurately.

The Build Week prototype is not the complete Aletheia system.

It is a vertical slice of one operational nucleus inside Aletheia.

Aletheia itself belongs to a broader multi-project research program involving continuously evolving cognitive architectures, semantic-operational grammars, contextual memory, perspective modulation and other connected components.

The circumscribed case was selected because it makes one part of that larger field immediately observable and testable.

The prototype is therefore not a miniature representation of everything the broader architecture contains.

It is a lens.

Through that lens, one central principle becomes technically visible:

an AI model can remain generative and interpretive;
raw sources can remain preserved;
evidence can be checked;
informational boundaries can remain visible;
perspectives can remain distinct;
counterexamples can transform previous matrices;
earlier versions can survive correction;
human authority can remain explicit;
interpretation can remain separate from action.

Why the value can compound

The value of Aletheia is not confined to this individual vertical slice.

The system is designed so that further development can produce cumulative rather than isolated value.

Adding a new case does not merely add another record. It can test, refine, narrow or reorganize existing matrices.

Adding a new perspective does not merely produce another answer. It expands the ability to distinguish information boundaries, convergences and unresolved tensions.

Adding a new criterion or index does not merely create another feature. It can modify how many previous and future cases are interpreted.

Adding a new model can increase interpretive capacity without transferring provenance, authority or consequence control to that model.

Adding a new domain can reuse the same governance primitives while developing domain-specific criteria.

Adding tools, connectors, evaluation infrastructure, real-world corpora or collaborators would therefore not simply enlarge the current demo.

It could increase the number and quality of relations that the infrastructure is able to observe, preserve and govern.

This is why the prototype should not be understood only as a bounded deliverable.

It is evidence of an extensible substrate.

The interpretive model can change.

The interface can change.

The domain can change.

The governance continuity remains.

A direction for safety-sensitive interaction

One especially relevant direction is safety-sensitive human–AI interaction.

Aletheia could operate between model interpretation and external action, making visible:

what evidence was actually available;
what information was unavailable;
which inference came from which source;
where observation becomes interpretation;
what authority is required;
what permission has been granted;
whether an action is reversible;
whether a human gate is necessary;
whether a consequence must remain suspended.

Aletheia does not claim to make an AI system universally safe.

It proposes an infrastructure for making some of the transitions where safety is often lost more observable and governable:

data → interpretation

interpretation → decision

decision → permission

permission → action

The three live tests already provide bounded evidence for this direction:

unsupported evidence was rejected;
contextual transformation remained versioned;
asymmetric information remained visible;
subjective evidence was not converted into certain psychology;
procedural evidence was not converted into proof of fairness;
authority and external action remained separated from model interpretation.

How I built it

I developed the prototype through an intensive human–AI workflow using Codex and GPT-5.6.

Codex supported:

repository inspection;
implementation of the vertical slice;
provider integration;
test construction;
adversarial checks;
portability hardening;
documentation;
review;
and reproducibility.

GPT-5.6 Sol served as the live interpretive engine through the Responses API.

It proposed bounded functional signatures, evidence excerpts, uncertainties, alternative readings and possible routes.

Aletheia remained responsible for:

preserving immutable manifestations and provenance;
validating exact evidence;
assigning candidate status;
retrieving precedents;
preserving non-equivalence;
requiring human gates;
versioning matrices;
maintaining genealogy;
and separating interpretation from consequence.

The live requests used store: false.

The repository includes public synthetic fixtures and sanitized evidence. It does not include the private historical corpus or API credentials.

Challenges and lessons

The first major lesson was that model fluency and structured output are not sufficient guarantees.

The first live failure proved that governance must validate not only the form of an output but the relationship between a claim and its actual source.

The expanded tests showed that:

the same raw datum can support different contextual functions without losing identity;
a new case can reorganize earlier classifications without erasing their history;
disagreement can derive from asymmetric information rather than incompatible reasoning;
subjective human evidence can remain pertinent without becoming certain psychology;
procedural correctness does not establish substantive fairness;
a well-formed question can be more operationally valuable than a premature conclusion.

The safety tooling also generated a useful failure.

A local scanner initially treated the ordinary semantic word “authorization” as if it necessarily indicated an HTTP credential. The export was stopped. The scanner was then narrowed to credential-shaped patterns and the event was documented.

This reinforced the same principle:

governance mechanisms must themselves remain observable, revisable and versioned.

What further support could unlock

The next stages include:

native persistence of multiple simultaneous perspectives;
larger and real-world corpora;
richer banks of criteria and indices;
interfaces for human review and genealogical navigation;
additional models and provider-neutral adapters;
granular authority and permission models;
external tools and connectors;
independent qualitative and quantitative evaluation;
domain-specific deployments;
and integration with broader cognitive architectures.

Additional access to tools, datasets, evaluation infrastructure, technical environments and interdisciplinary collaboration would not merely polish the current prototype.

It would allow the same operational nucleus to be tested across more complex conditions and connected to further components of the larger research program.

Each expansion could also produce new questions, new modules and new development directions that are not yet visible from the circumscribed case.

The potential value is therefore not limited to completing one predefined product.

It lies in the ability of the project to generate and connect further research, infrastructure and applications over time.

A long-term direction, not a one-off submission

For me, Build Week is not only an opportunity to present a temporary experiment.

This submission is the first public, reproducible expression of a research and engineering direction that I intend to pursue continuously.

The aim is not simply to complete a single application and stop at its current boundary.

The aim is to develop an environment in which:

interpretation can remain connected to its origin;
human and AI reasoning can adapt without losing genealogy;
new cases can transform previous knowledge without erasing it;
collaboration can become more stable and cumulative;
and increasingly capable AI systems can remain connected to explicit structures of responsibility and governance.

The prototype demonstrates one small but functioning part of that direction.

Its incompleteness is not an absence of direction.

It is the space in which the architecture can grow.

The broader vision

Aletheia is not another vertical AI application.

It is an attempt to build an extensible interpretive infrastructure for situations in which humans, AI systems and data continuously transform meaning together.

The prototype begins with a deliberately circumscribed and understandable case.

The longer-term objective is much larger:

preserve the origin of data, make transformations visible, keep plurality alive, allow correction without erasure, and ensure that no interpretation becomes a consequence without passing through an explicit and governable threshold.

The project is presented here not as a finished endpoint, but as verified evidence that this direction can be made operational.

Built With

api
gpt-5.6
node.js
openai
react
typescript
vite
vitest