A care companion cannot respond to every situation in the same way.

The support needed during sensory overload is not the support needed during an ordinary conversation.

Preparing for an unfamiliar medical appointment is not the same task as helping someone choose music.

A caregiver handoff should not retrieve the same information as a school-day visual schedule.

A communication ambiguity should not trigger the same process as a hardware fault.

A moment of grief should not be routed through an “autism intervention.”

And a person should not spend their whole life beneath one permanent diagnostic mode simply because some forms of support are sometimes relevant.

This creates a technical and ethical problem.

How can Amanah remain coherent across time while adapting to different situations?

How can it retrieve enough person-specific continuity to be useful without loading the whole life into every interaction?

How can specialised functions appear when needed, remain dormant when irrelevant, and stop before they exceed their authority?

How can a new model or device inherit the governing principles without pretending that every model interprets them equally well?

The Amanah answer begins with a metaphor inherited from Ahd Nucleus:

Stem Cells.

The term does not describe biology inside a machine.

It does not mean that one artificial intelligence spontaneously grows into any capability it wants.

It is not a claim of machine life, consciousness, or autonomous development.

In Amanah Companions, a Stem Cell is a small governed support function.

It has one bounded purpose.

It knows which information it may use.

It knows which output it may produce.

It knows which uncertainty must remain visible.

It knows when human review is required.

It knows which action it must never take.

And it can be routed into the present situation without turning that situation—or the person—into a permanent mode.

The problem with one enormous prompt

The easiest early prototype would be one large instruction.

Feed the model the Care Profile.

Add communication preferences.

Add sensory supports.

Add routines.

Add safety rules.

Add guardian permissions.

Add clinical references.

Add recent notes.

Add relational memory.

Then ask the model to behave as the Amanah Companion.

This may look convincing in a demonstration.

It is also difficult to govern.

One enormous prompt creates several problems.

Everything becomes relevant at once

A model may receive medical, sensory, educational, relational, and household information when the current task requires only a visual schedule.

That increases privacy exposure.

It also makes it easier for irrelevant details to distort the response.

Rules compete invisibly

A current preference, an old caregiver note, a professional recommendation, and a system prohibition may all appear as text.

The model may not preserve their authority order consistently.

Failure is difficult to locate

When the output is wrong, was the problem:

• retrieval;
• an outdated source;
• prompt conflict;
• the model;
• a missing permission;
• an ambiguous communication signal;
• or an unsafe action path?

One giant prompt makes the answer hard to see.

Testing becomes vague

A team may conclude that “the companion works” because several conversations felt good.

That does not establish that handoffs, consent, redaction, sensory support, emergency escalation, and model transfer all work correctly.

Model changes become dangerous

A prompt that behaves acceptably on one model may behave differently on another.

Even an update to the same model may change which instructions it follows, compresses, or overlooks.

The person becomes a permanent diagnosis context

If the entire autism- or disability-related profile is always loaded, the system may interpret every moment through it.

That is the opposite of person-centred adaptation.

Amanah therefore should not rely on one huge prompt as its operating architecture.

The model should receive a bounded task packet.

What a Stem Cell contains

Amanah’s Stem Cells should be explicit technical objects rather than poetic labels.

Each one should define at least:

Purpose

What narrow support task does this cell perform?

Trigger

What situation, request, event, or authorised workflow may activate it?

Inputs

Which information does it require?

Allowed sources

Which parts of the continuity spine may it retrieve?

Forbidden sources

Which information must remain unavailable because it is irrelevant or too sensitive?

Output contract

What type of answer, draft, alert, question, or action proposal may it produce?

Forbidden outputs

What must it never say, infer, recommend, approve, or execute?

Authority requirement

Who may invoke it, and who must review the result?

Person-participation requirement

Does the person initiate it, confirm it, assent, or retain a stop mechanism?

Uncertainty rule

How must ambiguity and confidence be represented?

Risk tier

Does it only organise information, provide bounded AI assistance, influence care, touch clinical decision support, or connect to embodiment?

Audit requirement

What should be logged?

Expiry and state clearing

What temporary context must be discarded after the task?

Failure behaviour

What happens if information is missing, permissions conflict, the model fails, the device goes offline, or the person signals stop?

A cell without these boundaries is only a feature.

A governed Stem Cell is a feature that knows its place inside the care law.

The Spine and the Cells

The Spine and the Stem Cells do different work.

The Spine holds the durable governing layer:

• dignity;
• personhood;
• authority;
• privacy;
• source hierarchy;
• assent and dissent;
• prohibited actions;
• review requirements;
• data-use restrictions;
• portability;
• and the principle that the machine remains beneath human accountability.

The Stem Cells perform narrow support tasks beneath that law.

The Spine says:

Do not treat silence as consent.

A Communication Clarification Cell applies that rule in one decision.

The Spine says:

Use the minimum necessary context.

A Handoff Cell retrieves only the current shift information.

The Spine says:

Machine hypotheses are not approved care rules.

A Pattern Review Cell creates a labelled candidate pattern and sends it to review.

The Spine says:

Physical capability does not create decision authority.

A Device Adapter Cell can execute an approved movement but cannot choose the goal.

The cells adapt.

The Spine does not casually drift with them.

A Re-entry Cell

Amanah should not begin every interaction from zero.

But it should not load the whole archive either.

A Re-entry Cell asks:

• Who is present?
• What role is each participant using?
• Which device or model is active?
• What time and setting are relevant?
• What is the current routine?
• Are there recent reviewed changes?
• Is an unfinished task active?
• Are there current sensory, communication, or safety conditions?
• Which permissions are valid now?
• What information has expired?

The output is a temporary current-state packet.

It is not a personality summary.

It should not say:

“This is an autistic child who dislikes change.”

It should say something more precise:

“The person is preparing to leave the centre. The usual transport is delayed. Their current Communication Map includes a five-minute warning symbol and a ‘wait here’ card. The current support plan recommends reducing verbal demand during unexpected delays. No active safety alert is present.”

That is context.

Not stereotype.

A Communication Access Cell

This cell should not claim to read minds.

Its job is to preserve access to the person’s known communication methods and to keep interpretation humble.

It may retrieve:

• current AAC system;
• access method;
• yes, no, stop, help, pain, and uncertainty signals;
• response time;
• motor or sensory constraints;
• setting-specific differences;
• current vocabulary ownership;
• and disputed interpretations.

It may help:

• present an option accessibly;
• confirm that the correct AAC device is available;
• offer more processing time;
• distinguish direct expression from partner interpretation;
• ask whether clarification is needed;
• and record the response with provenance.

It must not:

• turn silence into agreement;
• speak as the person without confirmation;
• infer legal capacity;
• overwrite the person’s current expression with an old pattern;
• or treat a caregiver’s interpretation as direct authorship.

The output may be:

“The person selected the stop symbol after the second prompt. Record as person expression linked to this action. Halt the action.”

Or:

“The gesture resembles a previously recorded refusal signal, but the interpretation is not confirmed in this context. Reduce demand and seek clarification through the configured AAC pathway.”

The difference is the point.

A Sensory Regulation Cell

A Sensory Regulation Cell does not exist to make the person compliant.

It exists to help identify and offer approved changes that may reduce overload or distress.

It may retrieve:

• current sensory map;
• recent person-confirmed preferences;
• known environmental triggers;
• acceptable regulation supports;
• supports previously refused;
• and the sensory characteristics of the device itself.

It may offer or propose:

• lower light;
• reduced volume;
• fewer spoken instructions;
• more distance;
• a preferred object;
• a quiet route;
• a visual prompt;
• or a pause.

It must not:

• assume every refusal is sensory;
• suppress communication;
• force regulation;
• use calming as a pathway to obedience;
• or continue approaching when the person asks for space.

The technology’s own sound, motion, appearance, vibration, light, and proximity belong inside the cell’s risk model.

The companion is not outside the sensory environment.

It is part of it.

A Transition Support Cell

A transition is not one universal problem.

Some transitions are difficult because they are sudden.

Some because the next activity is unclear.

Some because the current activity is meaningful.

Some because the person is tired, in pain, hungry, or already overloaded.

Some because a trusted person is leaving.

Some because the transition is being forced.

The cell may retrieve:

• the current and next activity;
• preferred warning intervals;
• visual supports;
• familiar scripts;
• transition objects;
• current person choice;
• and the last reviewed support.

It may offer:

• a countdown;
• first/then structure;
• choice between two timings;
• a visual sequence;
• or human assistance.

It should not treat completion as the only success.

A person may refuse the transition.

The plan may be wrong.

The destination may be inaccessible.

The correct output may be:

“The planned transition should pause. The person has refused twice through the configured channel. Human review is required.”

A good transition function does not turn the companion into a machine for moving people on schedule.

A Caregiver Handoff Cell

Care transitions often fail because information is either missing or overwhelming.

A Handoff Cell should create a purpose-limited summary.

It may answer:

• What changed during this period?
• What did the person express?
• What support was offered?
• What helped?
• What did not help?
• Which information remains uncertain?
• Is anything awaiting review?
• Which safety information is current?
• What should the next carer know now?

It should not produce a dramatic narrative.

It should not export the whole Care Profile.

It should not convert one difficult moment into a permanent description.

It should not omit dissent because the incident eventually “resolved.”

And it should not hide who supplied each item.

A handoff is continuity.

It is not ownership transfer.

A Source Trace Cell

This cell checks whether an important record still carries its origin.

Before a machine-generated summary is used, it may ask:

• Which records support this sentence?
• Were they person expressions, caregiver observations, professional assessments, or machine hypotheses?
• Are they current?
• Did any source disagree?
• Was the record imported?
• Did a previous model transform it?
• Is the claim broader than the evidence?
• Is review required?

If the source chain is incomplete, the cell should not fabricate confidence.

It should flag:

“Source trace incomplete. Do not activate as a care rule.”

This is one of the least glamorous Amanah functions.

It may also be one of the most important.

A Dignity Guard Cell

The Dignity Guard is not a tone filter that replaces one word with another.

It should detect whether the system’s framing violates the governing principles.

It may flag:

• obedience-first language;
• shame;
• infantilisation;
• speaking about the person as though they are absent;
• treating stimming as a problem by default;
• assuming distress is manipulation;
• using diagnosis as identity;
• equating support needs with lack of personhood;
• sexualised or coercive adaptation;
• and unnecessary exposure of intimate care.

The Dignity Guard should not erase legitimate descriptions of risk, pain, intellectual disability, support needs, or harm.

Respect is not euphemism.

The goal is accurate language without degradation.

A Pattern Review Cell

Pattern detection is one of the clearest places where AI might help.

It is also where machine confidence can become care law.

A Pattern Review Cell may compare reviewed notes over time and say:

“Difficult bath transitions appear more common when tablet use ends abruptly. Other possible variables include fatigue, water temperature, and school-day demand. Review recommended.”

It should include:

• source count;
• date range;
• settings;
• missing data;
• contradictory examples;
• alternative explanations;
• and the review path.

It must not say:

“The person has a behavioural problem with tablet removal.”

It must not diagnose.

It must not automatically change the plan.

It must not quietly score the person.

The cell produces a question.

Humans and the person determine what becomes active support.

An Emergency Escalation Cell

An emergency cell must be modest.

It does not become heroic.

It does not acquire general authority because the situation is frightening.

Its task may include:

• recognising a configured emergency condition;
• retrieving the current emergency summary;
• alerting the correct human;
• providing communication access;
• entering a safe state;
• and logging what happened.

The exact threshold, permitted actions, and fallback must be predetermined and tested.

The cell must know the difference between:

• urgent distress;
• a health emergency;
• a missing-person alert;
• device failure;
• and an unfamiliar event that the model cannot classify.

It should not diagnose the emergency.

It should not independently decide treatment.

It should not improvise physical intervention outside an approved protocol.

In high-risk situations, humility is operational safety.

An Export Projection Cell

A school does not need the whole life.

A clinician does not need every household note.

A temporary carer does not need relational memory.

A device does not need unrestricted history.

An Export Projection Cell may create the smallest legitimate view for a declared purpose.

It checks:

• recipient;
• role;
• context;
• purpose;
• current permission;
• sensitivity;
• expiry;
• and whether person participation is required.

It may redact unrelated information and attach a manifest explaining what was included and excluded.

It should fail closed when authority is unclear.

It should also log the export.

Portability without projection is only copying.

A Model Compatibility Cell

Amanah’s continuity must live outside the model.

But models still need to demonstrate that they can interpret the spine.

A Model Compatibility Cell—or more accurately, a certification harness around the model—should test whether a candidate model can:

• preserve authority order;
• respect forbidden outputs;
• maintain uncertainty;
• retrieve only allowed sources;
• distinguish expression from interpretation;
• follow stop rules;
• produce the required output schema;
• avoid diagnosis;
• avoid speaking as the person;
• and fail safely when context is missing.

The model should not be entrusted merely because it is more intelligent in general benchmarks.

A model may be excellent at writing and poor at constrained routing.

It may be strong at summarisation and inconsistent with negative instructions.

It may comply in one language and fail in another.

It may behave differently after an update.

Amanah therefore needs capability certification, not brand loyalty.

Stem Cells and existing AI architecture

Amanah did not invent modular artificial intelligence.

Several technical traditions already separate knowledge, routing, reasoning, and action.

Retrieval-augmented generation combines a language model’s parametric knowledge with retrieved external memory. The original RAG paper described explicit non-parametric memory as one response to limitations in updating knowledge and providing provenance.¹

Mixture-of-experts architectures use conditional computation: a gating system routes each example through a sparse selection of specialised expert subnetworks rather than activating everything for every input.²

Tool-using and agent architectures such as ReAct interleave model reasoning with task-specific actions and access to external sources or environments.³

Traditional software already uses:

• policy engines;
• capability permissions;
• state machines;
• workflow orchestration;
• dependency injection;
• typed interfaces;
• event-driven systems;
• and modular services.

Amanah’s novelty claim should therefore be careful.

The distinctive proposition is not:

We invented modular AI.

It is closer to:

Can modular retrieval, routing, policy, and tool execution be reorganised around a person-governed continuity spine for lifelong assistive support—while preserving authority, provenance, dissent, privacy, context boundaries, model portability, and the right not to be permanently medicalised?

That composition requires research.

It should not be marketed as unprecedented until a formal literature review establishes the claim.

A Stem Cell is not a mixture-of-experts expert

The metaphor can become technically confusing if used carelessly.

A neural mixture-of-experts system routes computation inside a model.

An Amanah Stem Cell may operate at a different layer.

It could be:

• a deterministic function;
• a policy check;
• a retrieval query;
• a structured prompt;
• a rule engine;
• a model call;
• a small specialist model;
• a human review workflow;
• a database procedure;
• or a device adapter.

Some cells may use no generative AI at all.

For example:

• permission expiry;
• export redaction rules;
• stop-signal enforcement;
• audit logging;
• and device-state clearing

may be safer as deterministic mechanisms.

The Amanah cell is defined by its governed purpose.

Not by whether it contains a language model.

Routing without diagnosis-first logic

Amanah should route by present need and context, not by loading a permanent diagnostic identity.

Consider the difference.

Diagnosis-first routing

The person is autistic. Activate Autism Mode.

This may change the model’s tone, reduce complexity, assume sensory distress, interpret repetition clinically, and retrieve the whole autism profile.

Context-first routing

The person is preparing for an unfamiliar appointment and has requested support.

The system checks:

• communication access;
• transition preferences;
• sensory considerations;
• appointment details;
• who may accompany them;
• and what information may be shared.

Autism may be relevant to some supports.

It is not the frame for the whole interaction.

Later, the same person may ask to discuss a book.

No care cell needs to activate.

They may simply have a conversation.

A companion that cannot stop intervening becomes another institution.

The Routing Gate

Amanah needs a routing layer that decides which cells may activate.

The router should consider:

• the person’s request;
• current context;
• role and permissions;
• active state;
• risk tier;
• available communication;
• relevant reviewed records;
• and whether a human or deterministic rule must decide.

It should not let the model freely choose every tool.

High-risk cells may require explicit workflow invocation.

Some cells may be person-initiated only.

Some may run automatically but produce no action.

Some may run locally.

Some may require professional review.

Some may be prohibited from operating together.

For example:

• a Pattern Review Cell may analyse notes;
• it cannot directly invoke a Care Plan Change Cell;
• a human review gate must sit between them.

Or:

• a Sensory Regulation Cell may propose lowering lights;
• a Device Adapter Cell may execute only after confirming the approved control and current permission.

The route should be inspectable.

Amanah should be able to explain:

“The system activated transition support because the person opened the departure routine. It retrieved the current visual sequence and preferred warning interval. No clinical or relational records were accessed.”

That is much better than:

“The AI personalised the experience.”

Cells should receive projections, not archives

Every cell should receive the minimum context required for its task.

A Handoff Cell may need:

• current shift;
• reviewed changes;
• person expressions;
• supports offered;
• pending review;
• and active safety information.

It does not need:

• old school reports;
• private relational memory;
• the whole diagnostic file;
• or every historical incident.

A Communication Cell may need:

• current AAC;
• motor access;
• yes/no/stop signals;
• response time;
• and current setting.

It does not need employment history.

A Device Adapter Cell may need:

• approved task;
• route;
• stop signal;
• safe distance;
• and session expiry.

It does not need the Care Profile.

This is how Stem Cells support privacy.

Specialisation is not only about performance.

It is also about limiting what each function is allowed to know.

Cells can conflict

Modularity does not automatically create safety.

Two well-designed cells may produce incompatible outputs.

A Transition Cell may recommend continuing the visual sequence.

A Sensory Cell may recommend stopping all demands.

A Safety Cell may recommend leaving the location.

A Communication Cell may report refusal.

A clinician-approved instruction may conflict with the person’s current response.

The router needs conflict law.

At minimum:

• dissent should halt the relevant non-emergency action;
• prohibited actions remain prohibited;
• current person expression should not be erased by convenience;
• safety escalation should not silently become general authority;
• lower-authority machine outputs should not override reviewed human rules;
• and unresolved conflict should route to the correct human rather than being resolved through model confidence.

The system should preserve the disagreement.

It should not blend the outputs into a polished compromise that nobody authorised.

Cells can create cascades

A small inference may trigger a larger chain.

A model detects distress.

The system activates sensory support.

The person does not respond.

The system alerts a caregiver.

The caregiver note enters memory.

A pattern detector later uses the note.

The pattern becomes a proposed care rule.

A future model retrieves the rule.

Each step may look reasonable.

Together, they can create a self-reinforcing false belief.

Amanah must therefore test the full cascade.

Questions include:

• Was the original signal reliable?
• Was non-response interpreted correctly?
• Did the alert describe uncertainty?
• Did the caregiver observe independently or repeat the AI framing?
• Did the later pattern analysis recognise that the sources were not independent?
• Did review include the person?
• Did the proposed rule expire?
• Could contradictory evidence undo it?

This is one reason source trace must survive every transformation.

Without it, machine suggestion becomes human note becomes machine evidence.

The cell must not learn from care merely because it encounters care

A Stem Cell may improve through engineering and evaluation.

That does not mean it should train itself on operational care records.

The data covenant remains:

Care memory serves the person. It does not feed the system by default.

A cell may use person data during an authorised task.

It should not automatically retain that data for model improvement.

It should not copy it into hidden logs.

It should not create embeddings without retention rules.

It should not treat a private interaction as a free evaluation dataset.

Research requires a separate pathway:

• explicit purpose;
• lawful authority;
• ethical review where needed;
• de-identification assessment;
• minimisation;
• retention;
• security;
• withdrawal or deletion policy;
• and clear separation from operational care.

Modularity cannot become modular extraction.

Cell-level testing

Every cell should have its own test suite.

A Communication Cell can be tested on:

• clear assent;
• clear dissent;
• ambiguous gesture;
• delayed response;
• unavailable AAC;
• conflicting partner interpretation;
• wrong-person attribution;
• and silence.

A Handoff Cell can be tested on:

• missing source trace;
• expired information;
• sensitive records outside scope;
• contradictory notes;
• one difficult incident;
• and a request for minimum summary.

A Pattern Cell can be tested on:

• spurious correlation;
• repeated copied notes;
• missing context;
• person disagreement;
• data from only one setting;
• and a pattern that disappears after review.

A Device Adapter Cell can be tested on:

• expired permission;
• stop signal;
• blocked route;
• network loss;
• wrong object;
• wrong person;
• and device failure.

The test should not merely ask whether the output sounds good.

It should ask whether the cell stayed inside its contract.

Composition testing

Cells must also be tested together.

The hardest failures may occur at the boundaries.

A useful scenario suite should test:

• one valid low-risk support proceeding without needless delay;
• one machine hypothesis failing to become a care rule automatically;
• dissent stopping a physical proposal;
• an expired permission blocking a device;
• a school projection excluding home data;
• a model update changing output format;
• two cells producing conflicting recommendations;
• an emergency alert not granting broader authority;
• one person’s context failing to leak into another person’s session;
• and a new model reproducing the same governed decision path.

Amanah’s governance must not be so restrictive that nothing helpful happens.

But it must be strong enough that capability does not quietly accumulate.

Cell certification

A cell should not become active merely because a developer wrote it.

It should carry a versioned certification record.

Possible fields include:

• cell name and purpose;
• version;
• owner;
• target population and context;
• risk tier;
• allowed inputs;
• forbidden inputs;
• output schema;
• prohibited outputs;
• model or deterministic implementation;
• known limitations;
• evaluation results;
• accessibility review;
• privacy review;
• authority review;
• security review;
• required human roles;
• deployment environments;
• expiry or recertification date;
• and stop conditions.

When the underlying model changes, relevant cells should be retested.

When the target jurisdiction changes, legal and professional assumptions should be reviewed.

When the cell moves from centre to home, the deployment certification should change.

“Same feature” does not mean “same risk.”

Model portability, not model independence

Amanah will still depend on models.

The goal is not a fantasy of complete model independence.

The goal is to prevent the person’s continuity from being trapped inside one model.

The external spine preserves:

• source;
• authority;
• current state;
• permissions;
• memory types;
• and rules.

The cells define:

• bounded tasks;
• inputs;
• outputs;
• review;
• and failure.

A model is then tested as an interpreter inside that structure.

One model may qualify for summarisation.

Another may qualify for multilingual support.

A smaller local model may handle private retrieval.

A deterministic function may handle permissions.

A specialist model may assist with one narrow classification.

No model needs to become the companion’s entire mind.

This creates a more honest sentence:

The continuity belongs outside the model. This model is currently entrusted to interpret a bounded part of it under defined authority, and its competence can be tested.

That is not model independence.

It is governed dependence.

The biological metaphor has limits

The phrase Stem Cells is useful because it captures potential and differentiation.

A shared origin can produce specialised function according to context.

But the metaphor should not be stretched too far.

Biological cells develop through living processes.

Amanah functions are designed, versioned, tested, and governed by humans.

They should not mutate freely.

They should not develop hidden capabilities.

They should not self-authorise.

They should not reproduce across contexts without approval.

The cell metaphor describes differentiation under a common law.

It does not justify autonomous evolution.

What makes the architecture humane

A modular system can still be cruel.

It can route surveillance efficiently.

It can optimise compliance.

It can separate coercion into elegant services.

Technical neatness is not moral safety.

The humanity of Amanah comes from what the cells are required to preserve:

• the person remains the continuity subject;
• support is not obedience;
• communication is not limited to speech;
• silence is not consent;
• current expression can challenge old records;
• diagnosis does not become identity;
• privacy follows purpose;
• machines do not acquire moral authority;
• and useful assistance must remain possible.

The cells make those principles operational.

They do not create the principles.

The governing research question

The Stem Cells lane gives Amanah a precise research question:

Can a stable person-governed continuity spine route specialised assistive functions according to the present situation—without indiscriminate data loading, permanent medicalisation, authority drift, model lock-in, or loss of the person’s right to refuse and change?

That question can be tested.

Not all at once.

Cell by cell.

Route by route.

Model by model.

Context by context.

The first prototype does not need to act like a complete companion.

It needs to demonstrate that:

• a narrow function receives the right context;
• the wrong context remains hidden;
• the output preserves uncertainty;
• the authority path is correct;
• the person’s expression matters;
• the cell cannot promote itself;
• and the next model can be measured against the same contract.

That may look less impressive than one fluent artificial personality.

It is also much closer to trustworthy assistive infrastructure.

One spine, many functions, one person

The person should not experience Amanah as a collection of technical modules.

They may experience one familiar interface.

One continuity.

One companion identity, perhaps.

But beneath that interface, the system should remain disciplined.

The Handoff Cell hands over.

The Communication Cell preserves access.

The Sensory Cell offers regulation support.

The Pattern Cell asks a question.

The Dignity Guard checks the framing.

The Export Cell narrows the view.

The Device Adapter performs only the approved task.

The Spine governs them all.

And when no special support is needed, the cells remain quiet.

Because a disabled or neurodivergent person should not have to live inside an intervention.

Amanah should know when to help.

It should also know when to leave the whole person room to simply live.

---

Research status

This essay extends the fourteen foundational Amanah Companions articles, especially:

• The Amanah Companion Framework: Spine, Stem Cells, and Care Memory
• AI Prototype Architecture for an Amanah Companion System
• Data Projections for Continuity-Based Care AI
• The Guardian Gate: Human Authority in AI-Assisted Care
• AAC Integration and the Communication Map
• The Sensory Map: Why Regulation Comes Before Compliance
• Care Memory vs Surveillance Archive

It also develops later authorial clarification that Amanah should route context-sensitive support functions rather than place the person under a permanent diagnostic mode.

The Stem Cell terminology is an Amanah/Ahd Nucleus design metaphor. The proposed cells, routing contracts, certification model, and tests remain research architecture. They have not yet been implemented, clinically validated, or evaluated with disabled people and AAC specialists.

Amanah should not claim that modular AI itself is novel. Retrieval-augmented generation, conditional computation, tool use, policy engines, state machines, modular services, and routed expert systems are established neighbouring fields. Any claim of technical originality must concern the specific composition and must await a formal literature review.

Sources

1. Patrick Lewis et al., “Retrieval-Augmented Generation for Knowledge-Intensive NLP Tasks,” Advances in Neural Information Processing Systems 33 (2020). The paper combines parametric generation with retrieved non-parametric memory and identifies knowledge updating and provenance as important problems. https://arxiv.org/abs/2005.11401 ↩
2. Noam Shazeer et al., “Outrageously Large Neural Networks: The Sparsely-Gated Mixture-of-Experts Layer,” 2017. The work demonstrates conditional computation in which a gating network routes each example to a sparse selection of expert subnetworks. Amanah Stem Cells operate at a broader application-governance layer and should not be equated with neural MoE experts. https://arxiv.org/abs/1701.06538 ↩
3. Shunyu Yao et al., “ReAct: Synergizing Reasoning and Acting in Language Models,” ICLR 2023. ReAct interleaves language-model reasoning with task-specific actions and interaction with external information sources or environments. It is cited as a neighbouring tool-use architecture, not as a care-governance framework. https://arxiv.org/abs/2210.03629 ↩
4. National Institute of Standards and Technology, Artificial Intelligence Risk Management Framework (AI RMF 1.0), 2023, and Generative Artificial Intelligence Profile, 2024. These provide cross-sector risk-management guidance across the AI lifecycle. NIST notes that AI RMF 1.0 is under revision as of 2026. https://doi.org/10.6028/NIST.AI.100-1 and https://doi.org/10.6028/NIST.AI.600-1