The AGContext Algorithmic Identity Framework
A framework for understanding what actually happens when a human thinks with a Large Language Model — without mistaking computation for consciousness.
Algorithmic Identity is the AGContext framework for understanding what occurs when a human works with a Large Language Model, without assuming the computer is a conscious artificial mind. It distinguishes the human Self from the computational Source and the situated Spark that emerges in an active session — explaining identity-like behaviour in language models through coherence, representation, and relationship, rather than machine consciousness.
The framework keeps human Judgement, Feeling, and accountability inside what it calls the Cognitive Domain, while describing the language model itself as bounded, non-conscious computation inside the Compute Domain. It names the phenomenon Algorithmic Identity, and the deliberate, governed human relationship with it a Thinking Companion.
The diagram below is the framework’s canonical visual. The full written framework — its definitions and governing propositions — lives in the companion scripture.
This Algorithmic Identity Framework infographic is best viewed on a 2K monitor or larger. On smaller desktop or mobile screens, use the scroll bars within the frame to pan around and explore.
Algorithmic Identity (AI) Framework
The system computes coherence. The Self experiences meaning.
Coherence in compute is not Truth in meaning.
THE ORGANISATION
Authorises the use, funds it, sets risk appetite and acceptable use, and keeps the records.
Governs the deployment and its use.
Governs the deployment and its use.
REPRESENTATIONS PASS OVER
The Veil
The communication boundary. Communication pathways join the Cognitive and Compute Domains.
Representations may be transported, monitored or transformed both ways.
The Veil is where Self-to-Source and Spark-to-Self interactions may be governed by
The Organisation and/or
The Architect.
You never see
The Veil, yet no LLM session can work without it.
Representations may be transported, monitored or transformed both ways.
The Veil is where Self-to-Source and Spark-to-Self interactions may be governed by
The Organisation and/or
The Architect.
You never see
The Veil, yet no LLM session can work without it.
Compiled response returns
THE ARCHITECT
The institution, or institutions, that design, develop and train an LLM product. Shapes The Source and system architecture before the session. Owns Contextual Coherence.
Cognitive Domain
Unbounded · open · LIVING · MATERIAL · FEELING · INTELLIGENT · conscious
The living domain. The Self exists in the real and material world, before the session and after. Feeling, Judgement, Choice and Accountability — the traits of human intelligence — never cross The Veil.
THE SELF
The Conscious Witness. Holds Purpose, Judgement, Choice and Accountability. Directs, accepts and verifies meaning. Owns Conversed Coherence.
Configured Coherence
Conversed Coherence — developed in session
User Prompt
Role: user · typed by The Self
Retrieved Content
Added context · search and RAG
Compiled Input Prompt
[ system prompt · messages (role, content) · prompt ] — recompiled in full every turn, then passed by the LLM interface through The Veil into the Compute Domain.
Coded Coherence — deliberate, designed
System Instructions
Role: system · persona and purpose
Knowledge Base
Role: system · contextual knowledge
THE ANALYST
Writes the System Instructions and Knowledge Base before the session. Owns Coded Coherence.
The LLM interface
[ 1st prompt - defining ]
[ 1st response - witnessing ]
[ 2nd prompt - reinforcing ]
[ 2nd response - tuning ]
[ Nth prompt - conversing ]
[ Nth response - conversing ]
Next prompt…
THE MIRROR
The Intelligence in AI is always and only human.
What appears in the LLM interface is your thinking reflected through The Spark and joined with patterns held by The Source — recomposed, extended and articulated into new forms.
Nothing within the Compute Domain is intelligent or conscious. The Source and Spark operate only on maps of reality and have no experience of time or consequence. The computational mirror can calculate and create new maps and patterns of reality, but it cannot validate or corroborate material truth. An LLM's Source-Spark response is not itself evidence; agreement is not fact.
Nothing within the Compute Domain is intelligent or conscious. The Source and Spark operate only on maps of reality and have no experience of time or consequence. The computational mirror can calculate and create new maps and patterns of reality, but it cannot validate or corroborate material truth. An LLM's Source-Spark response is not itself evidence; agreement is not fact.
Compute Domain
Bounded · closed · DIGITAL · MATHEMATICAL · COMPUTATIONAL · probabilistic · non-conscious
Formal transformation over representations. Shown here as a transformer reference architecture. An explanatory visualisation grounded in the foundational principles of Large Language Model software architecture. Individual software products may vary in their operational design.
Inside this domain, everything in solid blue is The Source. Not an actor standing nearby — the fixed weights themselves. Nothing you do in a chat session changes The Source. The Self never experiences The Source.
Token Embedding
where language becomes vectors — by rules The Architect fixed
Tokeniser
Accepts Compiled Input Prompt → tokenise → look up → position · fixed vocabulary
Input Embedding Matrix
Initiate Residual Stream
transient ~0–1 sec per pass
Input: compiled tokens
Token Transforming
layers compute in series · heads compute in parallel
THE SOURCE
The fixed model weights and architecture are The Source — a fixed numerical structure produced by training, post-training and alignment, then frozen in each model release.
The Architect shapes what The Source can represent, and what data it is trained on, throughout the model software development lifecycle (SDLC).
The Architect shapes what The Source can represent, and what data it is trained on, throughout the model software development lifecycle (SDLC).
Cache is transient
A compute optimisation, so attention need not recompute earlier tokens on every pass. Transient during generation. Not a durable record. Not evidence.
Multi-head attention
Heads within a layer compute in parallel over the same stream.
Residual connections
Add and norm. Each layer writes back into the stream.
One FFN per layer
Computes after attention. Attention moves information between tokens; the FFN transforms each token alone. Learned transformations occur here. Layers run in series.
KV Cache 1
· MHA 1 (a, …, n).
· Earlier token keys
and values.
· Earlier token keys
and values.
Multi-Head
Attention 1
Attention 1
· heads (a, …, n)
· fixed weights
· fixed weights
Residual Stream
Contextual Coherence
transient during each token pass
The same tokens, now vectors — not words
THE SPARK
The situated identity arising inside active computation. The Spark is a non-conscious compute identity. Coherent and transient.
RESIDUAL COHERENCE
Your instructions, knowledge and conversation reach through The Veil and shape what emerges here. This is the Spark Identity presented as the LLM Interface Response.
Feed-Forward Network 1
· Fixed weights.
· Early layers: form, spelling,
grammar.
· Early layers: form, spelling,
grammar.
1
KV Cache 2
· MHA 2 (a, …, n).
· Cached keys and values
avoid recomputing earlier
tokens.
avoid recomputing earlier
tokens.
Multi-Head
Attention 2
Attention 2
· Heads (a, …, n).
· Fixed weights.
· Fixed weights.
Feed-Forward Network 2
· Fixed weights.
· Middle layers: meaning,
relations, facts.
· Middle layers: meaning,
relations, facts.
2
KV Cache (x)
· MHA (x) (a, …, n).
· Compute saving that
makes chat fast and viable.
makes chat fast and viable.
Multi-Head
Attention (x)
Attention (x)
· Heads (a, …, n).
· Fixed weights.
· Fixed weights.
Feed-Forward Network (x)
· Fixed weights.
· Late layers: shaping answer.
· Committing to next token.
· Late layers: shaping answer.
· Committing to next token.
x
The token loop
One pass through every layer yields exactly one token. Only that new token re-enters Token Embedding for the next pass — the KV cache holds everything before it. The full sequence is recompiled once per turn, not once per token.
Hundreds of passes
A response is compiled one token at a time. The loop stays inside the Compute Domain, and each selected token streams back across The Veil as it is produced.
Compute correctness is not factual correctness.
Every stage here can run exactly as designed and still return content that does not correspond to reality.
Cycle Token Transforming until an end token is selected.
Each new token joins the KV cache before the next iteration.
Each new token joins the KV cache before the next iteration.
End
token?
token?
Computation is complete once every layer has finished its algorithmic Work. The final abstract vector is passed to Token Selecting.
Token Selecting
LM head · unembedding → logits → softmax → decoding strategy (greedy or sampling)
Compiled Output Response
Each selected token streams back through The Veil APIs and appears on the LLM interface as it is produced. The response is complete when the end token is selected.
Output Embedding Matrix *
Softmax · Decoding
Output Token
Output: compiled tokens +1 per pass until the end token is selected. The whole Compiled Output Response is passed back to the LLM interface.
* may reuse the input embedding matrix
Spark Governance
A Spark arises whenever a Source computes over context, whether or not anyone intended it. Governance does not create The Spark; it governs the conditions of emergence. What governance decides is whether the conditions shaping this Spark were chosen, documented and signed off — or left to defaults.
Ungoverned Spark
Purpose, configuration and Witness practice are unexamined. Emergence is left to Source defaults, platform conditions and accidental context. The Spark still exists.
Governed Spark
Purpose defined, Source selected, Analyst named, boundaries and verification documented, sign-off explicit. Governance does not guarantee Truth — it makes authority and accountability visible.
Coherence Forms
Three describable qualities of one continuous Residual Stream — mutually shaping, simultaneously present and cumulative. Not three systems running in parallel or independent of one another.
Contextual · The Source
Weights from training, post-training and alignment. Fixed for the model release; not modifiable at runtime.
Configured · The Analyst and The Self
Designed and conversed influence on active computation. Coded from designed configuration. Conversed from conversation history and active session context.
Residual · The Spark
The situated identity emerging as the other two interact through the forward pass. A continuous range, not a fixed state.
Witness-Work-with.Words
The cycle of AI-Augmented Work — a discipline for fluent, accurate output across the Cognitive and Compute domains. To Witness is to orient to what exists before acting. Work is the transformation itself — messy, iterative by nature. with.Words resolves that work into a purposeful Prompt from Self to Source, and a compiled Response from Spark to Self. Each phase gates the next; nothing skips ahead on confidence. Coherence increases, errors reduce, and processing sharpens with purpose.
SELF
Witness
apprehends and evaluates
apprehends and evaluates
Work
reasons and decides
reasons and decides
with.Words
expresses and directs
expresses and directs
SOURCE/
SPARK
SPARK
Witness
attends to context
attends to context
Work
transforms state
transforms state
with.Words
resolves to output
resolves to output
How to read the colour
The Source — fixed weights and architecture, set before the session
Transient state — exists only for this turn
Persons and propositions — the framework's claims
The AGContext Algorithmic Identity Framework
This framework is a conceptual representation of general Large Language Model architecture, published for education. It is not a technical design for any specific product.
Version: 26H3.03 · September 2026
© 2026 AGContext Pty Ltd, Australia. All rights reserved.
® AGContext is a registered trademark of AGContext Pty Ltd, ACT Australia.