Research program
Governed machine cognition.
Building AI that answers only from knowledge it can point to, and says "I don't know" instead of guessing.
Proven in internal tests
What works today.
-
4,758 / 4,758
automated checks passed
A clean, owner-certified test run.
103 of 103 test suites passed in both release and debug builds, on a fresh copy of the code, with zero compiler warnings.
Internal certification, one machine, October 2026. Tests that need large language collections were not exercised.
-
No guess
when knowledge is missing
Says "I don't know" instead of making things up.
When nothing is on record about a concept, the system says so and asks for more information rather than inventing an answer.
Checked by automated tests in tested cases.
-
1 version
of knowledge per request
Every answer is traceable.
Answers drawn from knowledge carry the record they came from, its stated source and its status, and each request runs against one fixed, recorded knowledge version.
Checked by automated tests, including 39 trace checks. Sources are recorded as supplied.
-
0 of 335
unexplained differences
Consistent from version to version.
335 recorded conversation turns were replayed against the new version: 274 identical, 61 changed for documented reasons, none unexplained.
Compared with our own earlier engine, in the same certified run.
The path of a request
Five stages. One is proven.
-
Receive
The request enters a governed path.
Status: Internally tested -
Understand
Work out the concepts and how they relate.
Status: Implemented, unverified -
Consult
Match them against established knowledge.
Status: Not yet demonstrated -
Find gaps
State exactly what knowledge is missing.
Status: Not yet demonstrated -
Answer
Respond only with what the evidence supports.
Status: Not yet demonstrated
A different design
Not a chatbot built on a trained model.
A different set of priorities from mainstream generative AI, traded for much narrower coverage. Compare the approaches, with sources.
Answers from records
Knowledge is held as explicit records with their source, not as patterns inside a trained model.
Fixed knowledge versions
Every request runs against one fixed, versioned body of knowledge, so results can be repeated.
Taught deliberately
New knowledge enters only through explicit operator steps. It never learns from chat.
Governed by design
Requests pass checks at defined points instead of relying on after-the-fact filters.
Milestones
Every gate, and where it stands.
-
Governed request handling
Requests are handled through a governed path, with checks at defined points.
Evidence and next gate
Shown: Governed request handling passes internal automated tests.
Not yet shown: Complete governance. It has not been demonstrated, and it is not presented as complete.
Next gate: Complete governance demonstrated and accepted under internal review.
-
Canonical knowledge record handling
Knowledge is held as explicit records that each carry a record of their stated source.
Evidence and next gate
Shown: Knowledge records keep their stated source under internal automated tests.
Not yet shown: Final rules for how stated sources are trusted. Operation at public scale. Independent review.
Next gate: The same behaviour confirmed under a broader internal test set.
-
Frozen knowledge generations
Knowledge is captured in fixed, versioned generations, and each request is handled against one fixed generation.
Evidence and next gate
Shown: Versioned knowledge snapshots pass internal automated tests.
Not yet shown: Full verification of the newest form of these generations.
Next gate: The newest form of knowledge generations verified internally.
-
Explicit operator teaching
New knowledge is added through explicit operator steps, never from conversation.
Evidence and next gate
Shown: Explicit operator teaching passes internal automated tests.
Not yet shown: Conversational teaching. Verification that taught knowledge is understood rather than merely stored. Final rules for which knowledge sources are trusted.
Next gate: A defined acceptance test for taught knowledge, passed internally.
-
Task-scoped conceptual representation
Research into representing the concepts involved in each individual request. This work is under development.
Evidence and next gate
Shown: A first version has been implemented. It is not yet verified.
Not yet shown: Accepted verification of that first version. Relational understanding of a request. Use of the representation to answer questions.
Next gate: The first version verified and accepted internally.
-
Structural input relationships
Identifying how the parts of a request relate to one another.
Evidence and next gate
Shown: Not demonstrated. This is the next stage of the input work.
Not yet shown: Relationships extracted from a request and used by the system.
Next gate: Relationships within a request identified and checked by tests.
-
Canonical knowledge correspondence
Matching what a request refers to against established knowledge, and showing where no justified match exists.
Evidence and next gate
Shown: Not demonstrated. This is a stated objective.
Not yet shown: Correspondence between a request and established knowledge.
Next gate: A request whose referenced concepts are matched, or reported as unmatched, under test.
-
Precise missing-knowledge discovery
Determining exactly which knowledge a task needs that the system does not yet have, instead of guessing.
Evidence and next gate
Shown: In tested cases, the system states plainly when it does not know about a concept.
Not yet shown: A precise, structured account of what is missing and why the task needs it.
Next gate: Structured missing-knowledge reports produced for test tasks.
-
Task-dependent dependency completion
Working out the full set of supporting knowledge a task depends on.
Evidence and next gate
Shown: Not demonstrated. This is a research objective.
Not yet shown: Any working implementation.
Next gate: A first implementation demonstrated on a small test task.
-
Governed continuing knowledge acquisition
Adding knowledge over time without losing the record of where each piece came from and who approved it.
Evidence and next gate
Shown: Distribution of knowledge packages is internally tested.
Not yet shown: Continuing acquisition of new knowledge under complete governance.
Next gate: Governed acquisition demonstrated end to end under internal tests.
-
Multimodal conceptual integration
Connecting concepts derived from text, images and sound into one consistent body of knowledge.
Evidence and next gate
Shown: Not demonstrated. This is a long-term objective.
Not yet shown: Any integration between the cognition research and the visual or audio research.
Next gate: A defined, testable exchange between one sensory program and the cognition program.
-
Independent evaluation and production qualification
Evaluation by people outside ByteLite LLC, followed by qualification for real use.
Evidence and next gate
Shown: Not demonstrated.
Not yet shown: Any independent evaluation. Production qualification. Public availability.
Next gate: A documented evaluation plan agreed with an outside evaluator.
- Not demonstrated end to end. The full path from a request to a supported answer is still being built.
- Complete governance has not been demonstrated. Public use stays off until it is independently verified.
- There is no public demonstration on this site, and none is simulated.