Goals and scope
What this programme is trying to establish.
Cortex is not an attempt to build a better assistant. It is an attempt to make one specific architectural claim falsifiable: that capability can be formed as bounded domains and composed, rather than absorbed into a single expanding parameter space.
Thesis
A falsifiable alternative to monolithic scaling.
The programme explores whether general capability must live inside one ever-larger model, or whether it can be expressed by a shared educational foundation, the ability to form bounded cognitive domains on top of it, and a governed protocol that composes those domains.
The project does not claim AGI, and does not claim experimental success. What it claims is narrower and, we think, more useful: that the comparison has not been run fairly, and that building the apparatus to run it fairly is worth doing on its own.
A cognitive domain, in this programme, has an explicit boundary and capability contract; its own training and update lifecycle; exact foundation and prerequisite versions; a separated experience namespace; release gates and evidence; the ability to abstain or escalate; and a governed protocol for exchanging claims. Anything missing those properties is not a domain, whatever it is called.
Three layers
Formation as a sequence, not a single pass.
Each layer is a separate research question requiring separate authorisation. None inherits evidence — or approval — from the layer below it.
A K–12 educational progression
The EDR Foundation is intended to learn a developmental educational substrate that can later branch into independently governed domains. K–12 is a deliberate choice, not a corpus of convenience: educational order is a declared feature of the substrate, so curriculum stage, subject, level and material type must be evidence-backed rather than inferred from folder names. The first scientific objective is narrow — whether a new dense model can acquire a coherent educational progression from approved early-learning through K–12 material with measurable retention, calibration and cross-stage transfer.
Discipline material at university-major level
Specialist cognition is formed after the foundation, not approximated by a role prompt or a retrieval query. A domain is intended to sit at roughly the granularity of a university major. Biology and Medical are the first pair because they share foundations while retaining observable ownership differences — adjacent enough that a merged model is a genuinely strong control. They are experimental candidates, not a claim that these are the universally correct or minimal boundaries.
An organisation's own material
A formed domain plus an organisation's training material is intended to yield a role-scoped capability — something that can be relied on inside a particular job, held to that job's standard. This is the reason abstention, declared scope and evidence traceability outrank aggregate accuracy throughout the evaluation design. It is out of scope until layers 1 and 2 produce evidence, and no work on it is authorised.
Not the same as
Where this differs from adjacent approaches.
Several existing techniques look superficially similar. The distinction in each case is what the unit of capability is, and whether it has an identity that can be versioned, evaluated and held responsible.
| Approach | Typical unit | Cortex distinction |
|---|---|---|
| RAG | Retrieved context around a general model | Evidence is not treated as internalised cognition |
| Mixture of Experts | Hidden experts inside one model lifecycle | Domains have explicit identities, boundaries, versions and owners |
| Multi-agent | Prompted roles, often sharing one model | Nodes are independently formed and evaluated cognitive domains |
| Continual learning | One model absorbs a sequence of data | Updates can remain isolated inside one governed domain |
| DAPT / CPT | General model adapted to one domain | The object of study is how multiple adapted domains compose |
| Merging / adapters | Parameter-efficient task composition | Adds domain lifecycle, memory governance, evidence and protocol |
None of these are dismissed as ineffective. The claim is only that none of them makes the domain a first-class, versioned, independently evaluable object — which is what this programme is testing the value of.
What counts as progress
Five things that would make this worth continuing.
Capability is preserved
Domain performance holds within a pre-registered margin against the merged control, rather than being traded away for structure.
Updates stay local
Updating one domain causes materially less regression in the other than the same update applied to a merged model.
Claims survive transit
Cross-domain claims retain their evidence, assumptions, scope and uncertainty instead of arriving as confident, unattributed prose.
Overheads are visible
Router and protocol costs are measured and reported rather than hidden inside an end-to-end number.
Results are reproducible
Every result can be reconstructed from immutable data, code, environment and run manifests — by someone who does not trust us.
Hard problems
What is genuinely unsolved.
These are listed because they are the parts most likely to defeat the programme, and because a research page that only lists strengths is not reporting honestly.
- Defining when a cognitive domain has genuinely formed, rather than merely scoring well on its own exam.
- Discovering the smallest useful domain boundary — and resisting boundaries chosen because they match a department name.
- Preventing catastrophic forgetting and unsafe experience consolidation.
- Routing unknown and ambiguous questions, including the case where no domain owns the request.
- Resolving cross-domain contradictions without letting confidence stand in for truth.
- Preserving implicit premises during communication, so a claim does not silently broaden in transit.
- Demonstrating a system-level advantage under a genuinely fair budget, including all overheads.
Non-goals
What this programme is not doing.
These are excluded deliberately at the current stage. Some are later, separately authorised phases; others are permanently out of scope.
- Proving that a multi-domain network outperforms a merged model — that is the question, not the premise.
- Treating retrieval as domain cognition.
- Allowing raw events to update parameters.
- Calling prompted roles independent cognitive domains.
- Exposing a public network service, or any production inference endpoint.
- Any real-patient framing in the Medical domain, even where an educational question resembles a benchmark item.
Current status
Foundation mode. Contracts, lifecycle rules, protocol, governance, audit, data manifests, run manifests and safety interlocks exist. Training, model weights, inference, domain dispatch, experience write-back and external actions remain disabled. The programme is preparing its first test and has not produced experimental results.