The Map Without a Vision: One Sicilian Lemon, Many Cognitive Perspectives on Knowing Without Seeing

20 May 2026

•

28 min listen

•

Bartosz Lenart

AI & Machine LearningNeuroscienceConsciousnessAgentic AICognitionaphantasia
Get Instant Insight

Close your eyes and picture a lemon. Aphantasia usually means no voluntary visual image, not necessarily nothing else: some people still get scent, sourness, or weight without a picture. The map is built downstairs. The picture never reaches conscious access.

The Core Idea

  • The map is local. Broadcast is organization-wide. Hippocampus, visual-imagery loops, and action pathways compute structure whether you consciously "see" the result or not.

Conscious access is who can report from the workspace, broadcast is when the whole organization publishes a state across cortex, conscious access can trigger it, but the brain can too.

Silence at the rail does not imply silence in the map downstairs.

  • Only a fraction of what is computed becomes conscious. Most of the work runs below it. The report, the picture, the spoken word are the thin trace of states that crossed.
  • Four families of errors, one mind.

I. Omertà (hidden computation): the brain builds the representation, but conscious awareness never receives it (aphantasia, blindsight).

II. Embargo (blocked read-out): the brain can use the information to guide action, but you cannot consciously recognize, name, or explain it (form agnosia, hemineglect).

III. Severance (broken connection): one region keeps computing, but its results never reach the part that speaks and explains, that part fills in the gap instead (hemispherotomy, split-brain).

IV. Phantom (activity without input): the incoming sensory signal is weak or gone, but cortex keeps firing and awareness treats that activity as real (phantom limb, Charles Bonnet, tinnitus, Anton's syndrome, chronic pain, central sensitization).

Why this matters now

The systems being deployed as agents recently mostly are output-first: a transformer attends over a context window and prints tokens, with no persistent latent map underneath.

Build only the workspace and you get the same four failure modes the brain does, one family at a time. The fix is a predictive state layer below the threshold, plus a relational schematic of where the work is, where it came from, and what crossed before.

[ NEURAL COMPRESSION COMPLETE ]

80% signal retained.
Full depth below.

The lemon

The house is quiet. Quiet that deep can only be bought and kept in Palermo.

You wake to an orange sunrise slipping behind the shutters. Heat has waited all night in the corners. You cross the stone floor. Down the stairs without hurry.

On the wooden kitchen table sits a bowl of lemons from the Conca d'Oro. Heavy peel. Oil on the skin. Dimple at the stem. A leaf still on.

The stiletto is already in your hand. You halve the fruit on the wood.

The smell of olio di limone fills the air. You quarter it. Squeeze one piece in your fist.

Let the juice fall into a glass of cold water.

Now close your eyes and put that lemon back in your mind.

For most readers something visual arrives. A wash of yellow. The dimple. A hint of texture. For roughly three or four readers in a hundred, nothing visual arrives: no yellow glossy skin, no lemon outline, no flicker behind the eyelids.

That is what clinicians usually label aphantasia: voluntary mental sight never clocks in. Self-identification rates run higher, around one in eight, partly because self-report and formal imagery scales do not draw the boundary in exactly the same place. No single profile fits everyone.1

That label does not mean the mind goes blank in every sense. Some of those readers still catch sourness, olio di limone, the cold rim of the glass: smell, taste, body-memory without a picture to attach them to.1

In recent high-resolution fMRI, the Fusiform Imagery Node, the same patch that lights up when typical imagers try to picture objects like that lemon, still activated in aphantasic participants asked to visualise an object,2 but the voluntary visual picture still did not reach awareness. What is contested is whether that activity matches what seeing looks like in the brain, or only marks an early signal that never becomes a picture.3 Either reading places the failure between local computation and conscious access.

Set the lemon down.

Sicilian lemon orchards ran on records most people never pictured as a whole landscape: who owed whom, which routes were safe, what was stored where.4 Later, fugitive boss Bernardo Provenzano ran Cosa Nostra the same way: short folded notes passed from hand to hand, each person knowing only the link before and after, with no shared view of the full operation.5

The brain splits work similarly. Local circuits compute structure whether awareness ever gets a picture of it.

Only a selected subset reaches the broadcast layer. One of many competing computations gets published to the rest of the brain. Nonlinear ignition.6 Neural Bifrost.

The Stake: Agents Without a Map

The reason this matters today is architectural. Current agents are mostly workspace: a transformer attends over a context window and prints tokens. The map lives somewhere else, bolted on as vector stores, databases, scratchpads, or coordinator agents.789

The pattern is already visible in deployed model architectures. Activations can contain answers the decoder never says.10 Reasoning traces can omit the cue that steered them.11 Multi-agent stacks often break at the handoff, where a coordinator fluently synthesises work it never truly received.12 Retrieval systems can attach a citation that supports the answer without being the source that produced it.13 In multimodal stacks the gap is worse: one modality can look grounded while another never received the signal that should have anchored it.

The brain has the same kind of fault line, but lesions and congenital variants expose it cleanly. Aphantasia keeps the map and silences the picture. Blindsight steers around obstacles the patient swears are not there. Hemispherotomy leaves a hemisphere intact but unable to broadcast. Phantom limb pain publishes a hand that is gone.

That is the point. In any system with a threshold, only a fraction of what is computed crosses into the workspace.

Then, the question is:

What does the map below the threshold need to be, for the workspace to be worth listening to?

The Map and Conscious Access: Hippocampus, Sleep, and the Workspace

Local circuits build the map.

Ignition at the front of the brain crosses the threshold, puts conscious access on the workspace, and can set the broadcast running across cortex, but broadcast is the organization's work, not conscious access alone.

Maps of space and of everything else

The hippocampus is the brain's mapmaker. In rodents, place cells fire at particular spots and tile a map of the room. In monkeys, hippocampal neurons tile an abstract value space the same way.14 In humans, the same machinery maps categories, goals, and task rules.1516 With entorhinal cortex, it keeps a quiet schematic of where you are in the task and what connects to what, whether you are thinking about it or not.17

Sleep edits the map

By day, experience reaches the schematic through what crossed the threshold. At night, sleep replays and consolidates what mattered.1819 New facts lean on hippocampus for weeks unless a compatible schema already exists, in which case neocortex can absorb them in 48 hours.20

Conscious access

Think of the threshold as a doorway. For a brain state to reach the workspace, it has to push through. Conscious access is whoever steps into the workspace once the door opens, the part of you that can report what is there, reflect on it, and act on it on purpose. The full map keeps running behind the door, even though only the winner made it through.

Broadcast is a separate event. The moment the threshold ignites, that winning state is shouted across the brain at once, a fast, long-range gamma rhythm that locks cortex into a single shared signal.62122 Conscious access can pull the trigger, but it does not have to: the brain's networks can fire the broadcast on their own.

Four Families, One Seam

Every case below breaks on the same seam between local computation and what reaches awareness, and they break in four recognisable ways. Two axes vary across them: where in the stack the seam breaks, and what the workspace / broadcast does in response. Read the families as architectural frames.

FamilyWhat still worksWhat fails at conscious accessBrain examplesModel analogue
OmertàLocal circuits compute the signalSilence at the workspace: computation stays unspokenAphantasia, blindsightHidden knowledge or reasoning that never reaches the decoder
EmbargoThe signal can guide actionRecognition or explanation is blockedVisual form agnosia, hemineglectA system acts on context it cannot faithfully account for
SeveranceA separated subsystem keeps computingThe bridge to shared workspace is cut, so the narrator fills the gapHemispherotomy, split-brainAn orchestrator summarizes outputs it never actually received
PhantomBroadcast still runsSource-checking or sensory grounding is missingPhantom limb, Charles Bonnet, tinnitus, Anton'sHallucination or citation that looks grounded but was not causal

Anton's syndrome occupies more than one cell. V1 is gone (input missing, like Family IV) and the monitor that would say "no signal" has been severed (broadcast runs anyway). Fibromyalgia is the other stress test: one name on the chart, several very different patients underneath, and seams that land in different families depending on which patient you are looking at. Any family can host a phantom when the loop that would catch it goes quiet.

Most cases are silence with structure: the map is maintained even when conscious access never gets a picture.

A minority are sound without structure: the report crosses the workspace even though nothing valid produced it below.

I. Omertà. Aphantasia: Structure Without the Picture

The brain computes the image, conscious awareness never receives it.

Aphantasia keeps most of the structure of memory and reasoning, and loses the voluntary picture. Memory-recall studies find a weaker link between memory and visual areas while spatial organization is spared.23 What is preserved is the relational and semantic framework, not every frame.

The Fusiform Imagery Node still activates on cue and its link to the broadcast network drops sharply, and the memory–visual link weakens during autobiographical recall while spatial navigation is spared.2323

Blindsight, form agnosia, and hemineglect arrive by injury at the same fault line. The episode is there, but the remembered scene does not come back as a picture.

Blindsight: Processing Without Report

Silence again, at a different address. Deeper brain routes run their own map; at the workspace, still nothing visible.24

Aphantasia keeps the map and loses the imagery.

Blindsight keeps perceptual processing and loses conscious seeing.

Behaviour without the report

After damage to primary visual cortex (V1: the brain's main visual processor), patients are blind in the affected field by every clinical measure. Yet in forced-choice trials they pick out motion, location, orientation, even fear on a face well above chance, sometimes walking around obstacles they swear are not there.252426

The signal still travels on back-channel routes through deeper brain regions and patches of visual cortex next to V1, reaching attention and action areas, those routes drive action and attention, but the route to a conscious visual report is a dead end.24

The mirror case: a workspace that broadcasts nothing as something

After damage to the visual cortex on both sides, some cortically blind patients insist they can see.

This is Anton's syndrome (denial of one's own blindness): they walk into furniture, fail every visual exam, and invent detailed rooms with no insight that anything is wrong.2728

The vision is gone, but the monitoring loop that would say "there is no signal" has been severed. Broadcast carries a confident report from priors and language alone.

Blindsight denies vision that is there. Anton's syndrome claims vision that is not.

One seam, two opposite errors, and a bridge into Family IV below.

II. Embargo. Form Agnosia and Hemineglect: Vision Without Access

The signal reaches the brain and can guide behaviour, but conscious recognition and explanation are blocked.

Form agnosia loses recognition while keeping action.

Hemineglect loses half of conscious space while keeping local processing.

Visual Form agnosia: Grasping without recognising

Visual form agnosia is the mirror of blindsight: V1 intact, recognition stream damaged. The classic patient cannot report shape or size yet scales her grip correctly to grasp objects she cannot identify.2930 One path for what the body does, another for what awareness gets to see.

Hemineglect: Half of space drops off the report

After right parietal damage, patients ignore the left half of space while early visual cortex still processes left-side stimuli. The disconnection reading: cables to frontal attention regions are cut, so the broadcast that would give the left side conscious priority no longer fires.31

III. Severance. Hemispherotomy: Local Computation Without Broadcast

The connection between the two halves breaks. One side keeps computing, but the speaking half never receives what the other computed, and invents an explanation instead.

Hemispherotomy preserves an entire hemisphere's local computation.

Every route by which it could broadcast is cut.

What hemispherotomy does

Hemispherotomy pushes the same architecture to a brutal extreme.

Surgeons sever every connection between one half of the brain and the rest, to stop seizures no drug can hold back. Blood flow remains.

The result: a half-brain that is structurally intact but cannot send, receive, or contact its twin.3233

At rest, the usual idle-state networks still fire inside the cut-off hemisphere.34 EEG shows the isolated half running slow waves of deep sleep and anaesthesia while the person is awake on the other side.35

Local structure without broadcast, at hemisphere scale.621

Callosotomy is a milder cut: surgeons sever the corpus callosum, the thick cable linking the two hemispheres, while deeper routes stay intact. In Gazzaniga's classic experiment with a split-brain patient, the right hemisphere is flashed a snow scene and the left, a chicken claw. The left hand (wired to the right hemisphere) picks out a snow shovel. Ask why and the speaking left hemisphere, which only saw the chicken, answers without missing a beat: "to clean out the chicken shed."36

When the broadcast layer must explain inputs it never received, it invents fluently.

IV. Phantom. When Input Goes Quiet: Noise at the Gate

The sensory input is missing or too weak, but cortex keeps firing anyway, and awareness treats that activity as if it came from the body or the world.

Phantoms flip the failure mode of everything above.

Here broadcast is noisy, not silent.

When input from the body or senses drops out, cortex can keep firing, and broadcast carries the result as if it were real. Phantom limb pain is the limb-is-gone case. An older account framed it as bad remapping: neighbouring patches of cortex crowd into the now-silent hand area, and the invasion is the pain. Persistent phantom pain more often tracks preserved function in the former hand area, with remapping playing a smaller role than once supposed.37

The same architecture shows up at other addresses:

  • Charles Bonnet syndrome is the visual parallel after macular degeneration or glaucoma: vivid figures with full insight that they are unreal, on less incoming visual signal and a cortex that fires more easily.3839

  • Tinnitus is the auditory version after hearing loss, evidence that phantoms share one architecture across the senses.4041

  • Migraine with aura is a slow procession with nothing in the cart: a wave creeps across the back of the brain at a millimetre every twenty seconds, and the matching patch of the visual field shimmers and goes blank as it passes, with no scene outside to explain it.4243

  • Pain generalises the pattern. Thalamic pain syndrome after stroke damages the sensory relay itself.4445 Central sensitization (the body's pain volume turned up across the board, often discussed in fibromyalgia and other chronic pain syndromes) can make conscious pain feel louder than the body's actual signal explains.46 That is the clean Phantom read, and only one slice of a much messier syndrome.

  • The architectural inverse is congenital insensitivity to pain: the pain channel never fires, so conscious access correctly reports nothing while tissue keeps being damaged unreported.47

When the Dials Stack

Beneath those axes runs another panel: one dial for what the body sends up, another for how loudly the nervous system reads it back. One name on the chart, several seams underneath.

Fibromyalgia is the case where the grid meets a crowd, not a single patient.

A recent review of thirty-nine cluster studies kept finding the same few shapes under one diagnosis, none mapping cleanly onto a single family.4849 One branch, different citruses, hence the picture above. An interpretive read on those subgroup shapes:

  • Pain-loud (hurt more than fatigue) reads as Phantom.

  • Fatigue-and-sleep (fatigue more than hurt) reads as a map problem, the night work missed.

  • Thermal-sensitive (heat or cold on a probe hurts at lower intensities) reads as Phantom at the input gate, central sensitization amplifying signal before it reaches the workspace.

Catastrophizing and acceptance are not somatic profiles but psychological-style overlays from a separate cluster axis, riding on top of any of the somatic profiles above:50

  • Catastrophizing (catastrophic worry over the signal) reads as Severance at the narrative layer.

  • Acceptance (the inverse style) reads as the narrative letting go of the threat loop it was running. Sensory pain stays where it is at the gate, so the affective suffering the story used to add from above falls away.

One patient can sit on more than one variety at once, and the dominant seam can shift inside one person over months, which is one reason average treatments work only modestly, and often intolerably.5152

Vegetative and minimally conscious states extend Severance to the whole brain. Most show flat broadcast.

A minority follow covert commands: ask them to imagine playing tennis or walking through their house, and fMRI catches motor cortex or the brain's place-recognition area lighting up on cue, even though no one outside the room sees any outward response.53

The four-family layout also bends with the input.

Give aphantasic and typical readers vivid first-person fear scenarios on the page: typical readers' sweat response tracks the text, aphantasic traces stay flat. Show both groups frightening photographs and the gap disappears.54

Prose asks you to supply the scene.

A photograph arrives with the scene already built.

One body often runs all three at different levels at once.

In one cross-sectional study, partial imagery (hypophantasia) tracked with worse anxiety and depression than core aphantasia, linked to trouble naming one's own feelings and a weak felt sense of body signals.55

The picture dial half-turned with no steady readout from the body was the risky shape, not the dial all the way down.

In matched comparisons, people with aphantasia score higher on standard autistic-trait measures than controls,56 and autistic adults without a fibromyalgia label still score in the central-sensitization clinical range, with sensory sensitivity predicting symptom load.57

Independent dials, repeatedly stacked.

Put the stack together: aphantasic, high on autistic-trait measures, centrally sensitized.

You can know where it hurts, picture it weakly or not at all, and still feel it loudly at the level of conscious access.

The chart of the body is not the broadcast of pain.

The inner picture is not the hurt itself.

The LLM Is Not Aphantasic, Only Mapless

Output Without the Map

A loud front end with no stable memory behind it: each prompt is handled fresh, with nothing underneath that tracks the work over time.

An aphantasic brain keeps the map and loses the picture.

A standard LLM keeps the output layer and never builds the map.

Brain vs model

An aphantasic brain still has its map of tasks, goals, and episodes, whether the inner picture arrives or not.1723

A standard LLM is the mirror case: the workspace is there, the sampling is there, the report is there, but no relational map runs underneath broadcast.789

Multimodal models can clear visual tasks much as aphantasic minds handle a lemon without a picture, yet the source of the knowledge differs. Aphantasic concepts were embedded by years of real-world sight, sound, touch, and smell grounded in a working perceptual system, even when voluntary imagery never arrives,223 while the model stitches modalities together from text-trained representations with no independent source of truth beneath the report to check what was "seen".585960

The context window is replayed each call, never held as a standing map of the work. Ask about a step that never entered context and the model invents one fluently, the same move the split-brain narrator makes when filling in for the half it cannot hear.

Four families, in code

Each clinical family has a code analogue. Same seam, different measurements.

Omertà in code: hidden knowledge that never reaches output. Probes can find the right answer in activations even when the model never says it.10 The remedy starts before decoding: probe internal states, not just the final output.

Embargo in code: behaviour shaped by a cause the report leaves out. Injected hints can steer reasoning models while the final answer tells a cleaner story than the internal trace.11

Severance in code: orchestration without a shared workspace. Multi-agent systems fail not only because single agents are weak, but because messages are lost, roles drift, and coordinators certify work they never truly received.12

Phantom in code: grounding that looks real but was not causal. A citation can point to a document that supports the answer while not being the source that produced it.13 Mechanistic work suggests that kind of unfaithful attribution can leave a detectable signature in the model's internal pathways.61

What would fix the shape

Token prediction without a persistent latent map is broadcast with nothing durable underneath. A longer context window only stuffs more tokens into the workspace, but it does not add a map where structure lives.

The fix is a predictive state layer below the threshold: a representation the model can rehearse, revise, and consult without committing every step to an output modality (text, pixels, or tokens).

The hippocampus does not just predict the next frame, it plots a space of relations.14151617

The world-model revival learns latents that predict, plan, and flag implausible events before anything is rendered,62636465 with goal-conditioned value readable as distance between embeddings.66

Agents and workspaces

For AI agents, that predictive latent is only half the job. The other half is the relational schematic of where you are in the work:

  • Context - the work in front of you right now.
  • Source - where each piece came from.
  • History - what crossed before, and what followed.
  • Harness - the workspace, tools, and agents you operate inside.

Matching inputs across modalities is not the same as having a map of relations.

Both belong below the threshold.

The layer also needs a graded memory stack, not a shelf: in practice a continuous gradient, with detailed and schematic versions of a memory coexisting and interacting throughout its lifetime rather than handing off cleanly.67

  • Stage 1 Working context lives in the workspace (what is promoted now).
  • Stage 2 Episodic structure lives in hippocampus for weeks.19
  • Stage 3 Slow neocortical consolidation, finishing in 48 hours where a schema already exists.20

Most LLM agents today run only Stage 1, with retrieval bolted on outside.789

Stage 2 is capture in agentic system history and can affect re-representation of context if extracted.68

Stage 3 writes back into the schema: parameter-efficient weight edits during an offline phase, schema-compatibility-gated against catastrophic forgetting.69

The dominant production answer in recent systems is the opposite, continual learning in token space, keeping all per-user adaptation in external memory rather than in weights.70

Sleep-Consolidated Memory and small consolidator models sketch what that memory stack looks like in practice.7172

Treat the map as the invariant. Tune the threshold.

If every new agent rebuilds the map from raw transcript, every handoff is a callosotomy and every coordinator invents. Agent work inspired by Global Workspace theory splits the workspace (what is broadcast now) from experience memory (what crossed before and what followed): the workspace can rotate while the map stays.7374

Memory updates while the workspace is empty, as in sleep: hippocampal replay and cortical re-expression with the lights down.1819 Launch a new agent on the same memory and you get the same split without the wait: empty workspace, map already there. Other agents run consolidation between sessions: summarisers writing memory, structurers shaping what the next workspace reads. The threshold is what to tune per task: what gets promoted, with what cutoff, into which workspace. The map is what to keep stable underneath.

Map the architecture choices back onto the four families and the costs sort themselves:

  • Omertà is a decoder and eval problem
  • Embargo is a faithfulness problem
  • Severance is an orchestration problem
  • Phantom is a grounding problem

None of these is solved out of the box.

All are easier to specify when you stop pretending the workspace is the system.

A note on hardware

The memory stages above describe the process. The hardware layer is the next question. Stages 1–3 read as scheduled jobs because today's hardware makes them scheduled jobs: a transformer on a GPU has to leave its inference loop, run a fine-tune, validate, and resume. The brain does not stop to consolidate. Local learning rules run continuously on the chip, and the night is the time when input is low enough for the plasticity that was always there to dominate.

On brain-inspired hardware the same loop runs as a continuous local property rather than a scheduled job. Sleep-staged learning with three-factor plasticity is now demonstrated end-to-end on spiking architectures,75 sleep-like offline phases stabilise against unbounded weight growth and catastrophic forgetting,76 and on-chip continual learning is real silicon, not a sketch.77 Collocated memory and compute mean the schema is the connectivity itself, edited in place.

Where do you write back? What do you overwrite? The question dissolves into the hardware.

The likely near-term shape is not brain-inspired chips in place of transformers but transformer-as-workspace above, brain-inspired map hardware below, an analogy to how cortex and deeper brain maps have long coexisted rather than a claim of an exact one-to-one match.

What To Keep When Few States Ever Reach the Workspace

Only a fraction of what is computed makes broadcast. The fusiform can fire while conscious access stays empty. Phantom pain and Bonnet figures can cross the workspace with no valid signal below. Computation and consciousness pull apart, and we are building systems on purpose that way: most of the work runs below the threshold, tokens as the thin trace that crossed.

The mind can keep a map no one reads, receive a signal no one opens, lose the bridge so the narrator invents the missing half, or broadcast a report with nothing behind it, and it can do all four at once. Only a fraction ever reaches conscious access. Build only the workspace, as current agents mostly do, and you get the same four failure modes.

Sometimes the issue is failure, sometimes it is simply different architecture.

What's Next

This piece sits inside a longer arc on consciousness, attention, and how to build systems that respect the architecture of mind rather than fight it.

◆

Bartosz Lenart

I write blog.ckpt to think out loud about AI reasoning, agentic systems, and what it actually feels like to build cognitive architectures that don't fall over.

Enjoyed the article?
Get new essays on AI reasoning, cognition, and system design.
No noise.

Found this useful?

Share it with others who might benefit, or save the citation for your research.

Share
Citation

Bartosz Lenart (2026). The Map Without a Vision: One Sicilian Lemon, Many Cognitive Perspectives on Knowing Without Seeing. Retrieved May 25, 2026, from https://bartoszlenart.com/blog/map-without-a-vision

References

Footnotes

  1. Takahashi, J., & Saito, G. (2023). Diversity of aphantasia revealed by multiple assessments of visual imagery, multisensory imagery, and cognitive style. Frontiers in Psychology, 14, 1174873. Large-sample Japanese study (N = 2,871) reporting approximately 3.7% prevalence by VVIQ criteria (VVIQ ≤ 32) and 12.1% by self-identification, attributing much of the discrepancy to face-recognition items and documenting wide variation across multisensory imagery profiles. Replicates Dance, Ipser, & Simner (2022, Consciousness and Cognition 97: 103243), who reported 3.9% in a UK sample (N = 1,004). ↩ ↩2

  2. Liu, J., Zhan, M., Hajhajate, D., Spagna, A., Dehaene, S., Cohen, L., & Bartolomeo, P. (2025). Visual mental imagery in typical imagers and in aphantasia: A millimeter-scale 7-T fMRI study. Cortex, 185, 113-132. 7-Tesla fMRI showing Fusiform Imagery Node and ventral temporal activation during attempted visualization in aphantasia, with sharply reduced coupling to the frontoparietal network. ↩ ↩2 ↩3

  3. Scholz, C. O., Monzel, M., & Liu, J. (2025). Absence of shared representation in the visual cortex challenges unconscious imagery in aphantasia. Current Biology, 35(13), R645-R646. Reframes residual visual-cortex activity in aphantasia as sensory reactivations that do not share the representational geometry of perception. ↩ ↩2

  4. Dimico, A., Isopi, A., & Olsson, O. (2017). Origins of the Sicilian Mafia: The Market for Lemons. The Journal of Economic History, 77(4), 1083-1115. Uses original Damiani Inquiry data (1881-1886) on Sicilian municipalities alongside citrus suitability and profits to argue that the export-driven lemon and orange trade, weak formal enforcement, and high-value vulnerable assets made western Sicily a primary cradle of early mafia brokerage and protection. Municipality-level correlations; historians continue to debate mechanisms and how directly modern Cosa Nostra descends from those nineteenth-century networks. ↩

  5. Schneider, J., & Schneider, P. T. (2003). Reversible Destiny: Mafia, Antimafia, and the Struggle for Palermo. University of California Press. Background on Bernardo Provenzano's 43-year fugitive command of Cosa Nostra (1963-2006) via tightly folded typewritten notes, often coded with Biblical references, routed through a chain of couriers in which no intermediary knew more than the previous and next link. ↩

  6. Dehaene, S., & Changeux, J.-P. (2011). Experimental and theoretical approaches to conscious processing. Neuron, 70(2), 200-227. Canonical synthesis of Global Neuronal Workspace: feedforward sensory climb, late nonlinear ignition, and broadcast as the signature of conscious access. ↩ ↩2 ↩3

  7. DataCamp. (2025). How Does LLM Memory Work? Building Context-Aware AI Applications. DataCamp Blog. Practitioner-level overview of context windows, retrieved information, and the absence of persistent state in standard LLM serving. ↩ ↩2 ↩3

  8. Arize AI. (2024). Memory and State in LLM Applications. Arize AI Blog. Engineering breakdown of why LLMs are stateless by default and how persistent state must be added at the orchestration layer. ↩ ↩2 ↩3

  9. Mem0. (2026). Context Window vs Persistent Memory: Why 1M Tokens Isn't Enough. Mem0 Blog. Argues that scaling context length does not substitute for a persistent, relational memory layer for agents. ↩ ↩2 ↩3

  10. Gekhman, Z., Ben David, E., Orgad, H., Ofek, E., Belinkov, Y., Szpektor, I., Herzig, J., & Reichart, R. (2025). Inside-Out: Hidden Factual Knowledge in LLMs. Conference on Language Modeling (COLM) 2025, arXiv:2503.15299. Formal framework for hidden knowledge in LLMs: linear classifiers on intermediate activations rank correct answers higher than any external scoring method, with an average relative gap of 40% between internal and external knowledge. In 9% of questions, the internal scoring ranks the ground-truth answer above any incorrect candidate while the model never generates it across 1,000 sampled completions. Cited here as the AI analogue of the Omertà family: knowledge that is computed and stored but never broadcast to the output. ↩ ↩2

  11. Chen, Y., Benton, J., Radhakrishnan, A., Uesato, J., Denison, C., Schulman, J., Somani, A., Hase, P., Wagner, M., Roger, F., Mikulik, V., Bowman, S. R., Leike, J., Kaplan, J., Perez, E., & the Anthropic Alignment Science Team (2025). Reasoning Models Don't Always Say What They Think. arXiv preprint, arXiv:2505.05410. Evaluates chain-of-thought faithfulness of Claude 3.7 Sonnet and DeepSeek R1 across six kinds of inserted hints: across most settings, reasoning models verbalise their reliance on used hints in fewer than 20% of cases; Claude 3.7 Sonnet acknowledges them in only about 25%, DeepSeek R1 in about 39%. Outcome-based RL initially improves faithfulness but plateaus. Cited as the AI analogue of the Embargo family: the input demonstrably shapes behaviour, and the verbal head never reports the cause. ↩ ↩2

  12. Cemri, M., Pan, M. Z., Yang, S., Agrawal, L. A., Chopra, B., Tiwari, R., Keutzer, K., Parameswaran, A., Klein, D., Ramchandran, K., Zaharia, M., Gonzalez, J. E., & Stoica, I. (2025). Why Do Multi-Agent LLM Systems Fail? arXiv preprint, arXiv:2503.13657. MAST analyzes seven popular multi-agent frameworks across 200-plus traces and identifies 14 failure modes in three categories: specification issues (41.77%), inter-agent misalignment (36.94%), and task verification (21.30%), with strong human agreement (Cohen's Kappa 0.88). ChatDev reaches only 33.33% correctness on ProgramDev; interventions improve results but do not remove the structural failure modes. Cited as the AI analogue of the Severance family: orchestrators that fluently invent a synthesis they never received from the parallel specialists below. ↩ ↩2

  13. Wallat, J., Heuss, M., de Rijke, M., & Anand, A. (2025). Correctness is not Faithfulness in Retrieval Augmented Generation Attributions. Proceedings of the 2025 International ACM SIGIR Conference on the Theory of Information Retrieval (ICTIR '25). Introduces citation faithfulness (whether the model genuinely relied on the cited document rather than post-rationalising to fit pre-existing knowledge) as a desideratum distinct from citation correctness, and finds that up to 57% of citations in standard RAG settings lack faithfulness: cited documents appear to support the answer but were not actually used to generate it. Cited as the AI analogue of the Phantom family: a confident citation that supports the answer without having caused it. ↩ ↩2

  14. Knudsen, E. B., & Wallis, J. D. (2021). Hippocampal neurons construct a map of an abstract value space. Cell, 184(19), 4960-4970.e6. Single hippocampal neurons in primates encode a relational map across an abstract value space, mirroring the structural organisation place cells provide for physical space. ↩ ↩2

  15. Constantinescu, A. O., O'Reilly, J. X., & Behrens, T. E. J. (2016). Organizing conceptual knowledge in humans with a gridlike code. Science, 352(6292), 1464-1468. Human fMRI evidence that entorhinal cortex uses a grid-like code over a non-spatial conceptual two-dimensional space, extending the cognitive-map account from physical to abstract domains. ↩ ↩2

  16. Barnaveli, I., et al. (2025). Hippocampal-entorhinal cognitive maps and cortical motor system represent action plans and their outcomes. Nature Communications. Hexadirectional entorhinal codes and hippocampal similarity-scaling track learned action-outcome relational structure in human fMRI, extending the cognitive-map account to abstract motor planning. ↩ ↩2

  17. Eichenbaum, H. (2017). On the integration of space, time, and memory. Neuron, 95(5), 1007-1018. Synthesis of how hippocampal-entorhinal circuits maintain integrated representations of relational structure (space, time, and event memory) across awareness. ↩ ↩2 ↩3

  18. Wilson, M. A., & McNaughton, B. L. (1994). Reactivation of hippocampal ensemble memories during sleep. Science, 265(5172), 676-679. Foundational result: hippocampal place-cell sequences active during waking spatial behaviour re-fire in the same temporal order during subsequent NREM sleep, the original evidence that the brain replays experience offline. ↩ ↩2

  19. Diekelmann, S., & Born, J. (2010). The memory function of sleep. Nature Reviews Neuroscience, 11(2), 114-126. Synthesis of two-stage memory consolidation: hippocampal replay during NREM transfers labile traces into stable neocortical representations, with REM contributing to integration and emotional processing. ↩ ↩2 ↩3

  20. Tse, D., Langston, R. F., Kakeyama, M., Bethus, I., Spooner, P. A., Wood, E. R., Witter, M. P., & Morris, R. G. M. (2007). Schemas and memory consolidation. Science, 318(5858), 1076-1080. Demonstrates that prior schemas dramatically accelerate consolidation of compatible new information into neocortex, reducing the hippocampal dependency window from weeks to as little as 48 hours. ↩ ↩2

  21. Mashour, G. A., Roelfsema, P., Changeux, J.-P., & Dehaene, S. (2020). Conscious processing and the global neuronal workspace hypothesis. Neuron, 105(5), 776-798. Updated GNW review covering ignition, broadcast, and the boundaries of the theory. ↩ ↩2

  22. Gaillard, R., Dehaene, S., Adam, C., Clémenceau, S., Hasboun, D., Baulac, M., Cohen, L., & Naccache, L. (2009). Converging intracranial markers of conscious access. PLoS Biology, 7(3), e1000061. Intracranial EEG in epilepsy-surgery patients catches the same nonlinear late ignition (~300 ms) and long-range gamma-band synchrony that scalp EEG and fMRI had localised to the frontoparietal network, providing the depth-electrode evidence for the gating metaphor. ↩

  23. Monzel, M., Leelaarporn, P., Lutz, T., Schultz, J., Brunheim, S., Reuter, M., & McCormick, C. (2024). Hippocampal-occipital connectivity reflects autobiographical memory deficits in aphantasia. eLife, 13, e94916. Functional disconnection between hippocampus and visual cortex during autobiographical retrieval in aphantasia, with intact spatial cognition. ↩ ↩2 ↩3 ↩4

  24. Tamietto, M., & Morrone, M. C. (2016). Visual plasticity: blindsight bridges anatomy and function in the visual system. Current Biology, 26(2), R70-R73. Review of the subcortical and extrastriate pathways (LGN-pulvinar-MT, superior colliculus) that support blindsight after V1 damage, and the affective discrimination (fearful faces) preserved through amygdala routing. ↩ ↩2 ↩3

  25. Weiskrantz, L., Warrington, E. K., Sanders, M. D., & Marshall, J. (1974). Visual capacity in the hemianopic field following a restricted occipital ablation. Brain, 97(1), 709-728. The original demonstration of blindsight: a hemianopic patient (D.B.) discriminating visual stimuli in his cortically blind field at well above chance while denying any conscious vision. ↩

  26. de Gelder, B., Tamietto, M., van Boxtel, G., Goebel, R., Sahraie, A., van den Stock, J., Stienen, B. M. C., Weiskrantz, L., & Pegna, A. (2008). Intact navigation skills after bilateral loss of striate cortex. Current Biology, 18(24), R1128-R1129. Famous case study (patient T.N.) of a cortically blind patient successfully navigating an obstacle-filled corridor without conscious vision, demonstrating preserved action-guidance in the absence of phenomenal sight. ↩

  27. Das, J. M., & Naqvi, I. A. (updated 2023). Anton Syndrome. In StatPearls [Internet]. Treasure Island (FL): StatPearls Publishing. NCBI Bookshelf clinical overview of Anton-Babinski syndrome (visual anosognosia with confabulation in cortical blindness), covering disconnection accounts of why the patient denies blindness and confabulates visual content. ↩

  28. Atallah, O., Badary, A., Almealawy, Y. F., Farooq, M., Hammoud, Z., Alrubaye, S. N., Alwan, A. M., Awuah, W. A., Abdul-Rahman, T., Sanker, V., & Chaurasia, B. (2024). Insights into Anton Syndrome: When the brain denies blindness. Journal of Clinical Neuroscience, 121, 181-190. Review of contemporary case reports and mechanistic accounts of Anton's syndrome, framing visual anosognosia as forced confabulation by intact speech and decision systems disconnected from a missing visual signal. ↩

  29. Goodale, M. A., Milner, A. D., Jakobson, L. S., & Carey, D. P. (1991). A neurological dissociation between perceiving objects and grasping them. Nature, 349(6305), 154-156. Original report of patient D.F., showing intact visuomotor control of grasping despite severe inability to consciously perceive object shape, size, or orientation; the empirical anchor of Goodale and Milner's two visual streams hypothesis. ↩

  30. Whitwell, R. L., Milner, A. D., Cavina-Pratesi, C., Byrne, C. M., & Goodale, M. A. (2017). Still holding after all these years: An action-perception dissociation in patient D.F. Neuropsychologia, 105, 215-226. Twenty-five-year follow-up confirming that D.F.'s preserved visuomotor scaling for grasping persists despite ongoing deficits in conscious perception of the same object properties. ↩

  31. Bartolomeo, P., Thiebaut de Schotten, M., & Doricchi, F. (2007). Left unilateral neglect as a disconnection syndrome. Cerebral Cortex, 17(11), 2479-2490. Argues that hemineglect is best understood as a white-matter disconnection between right parietal cortex and frontal attentional regions, fitting the broader framework of preserved local processing with disrupted broadcast. ↩

  32. Alotaibi, F., Albaradie, R., Almubarak, S., Baeesa, S., Steven, D. A., & Girvin, J. P. (2021). Hemispherotomy for Epilepsy: The Procedure Evolution and Outcome. Canadian Journal of Neurological Sciences, 48(4), 451-463. Review of the historical evolution, indications, surgical technique, and outcomes of hemispherotomy for medically refractory epilepsy. ↩

  33. EANS Functional Neurosurgery Section Task Force. (2024). Functional hemispheric disconnection procedures for chronic epilepsy: history, indications, techniques, complications and current practice in Europe. Brain and Spine, 4, 102754. European consensus statement reviewing the evolution from anatomical hemispherectomy to modern functional hemispherotomy, detailing which connections are severed and which preserved. ↩

  34. Bauer, T., Gauvry, C., Markett, S., et al. (2024). Intact functional brain networks in the isolated hemisphere of people after hemispherotomy. Research Square preprint. Resting-state fMRI in hemispherotomy patients showing all seven canonical large-scale networks remain present and internally coordinated in the disconnected hemisphere. ↩

  35. Colombo, M. A., Favaro, J., Mikulan, E., Pigorini, A., et al. (2025). Hemispherotomy leads to persistent sleep-like slow waves in the isolated cortex of awake humans. PLOS Biology, 23(10), e3003060. Scalp EEG evidence (19-channel, 10–20 system) in ten paediatric patients: the isolated hemisphere shows persistent slow oscillations with spectral exponents indistinguishable from deep NREM sleep, general anaesthesia, and the vegetative state, while the patient is awake. ↩

  36. Volz, L. J., & Gazzaniga, M. S. (2017). Interaction in isolation: 50 years of insights from split-brain research. Brain, 140(7), 2051-2060. Fifty-year synthesis of split-brain findings, including the post-hoc confabulation paradigm (chicken-claw / snow-scene / shovel) in which the verbal left hemisphere fluently invents reasons for actions chosen by the disconnected right hemisphere. ↩

  37. Makin, T. R., & Flor, H. (2020). Brain (re)organisation following amputation: Implications for phantom limb pain. NeuroImage, 218, 116943. Review challenging the simple maladaptive-plasticity account of phantom limb pain, showing that persistent phantom pain is associated with preserved structure and function in the former hand area as well as neighbouring-territory remapping, with multiple contextual factors beyond S1 cortical reorganisation contributing. ↩

  38. Ffytche, D. H., Howard, R. J., Brammer, M. J., David, A., Woodruff, P., & Williams, S. (1998). The anatomy of conscious vision: an fMRI study of visual hallucinations. Nature Neuroscience, 1(8), 738-742. fMRI study of Charles Bonnet syndrome showing that hallucinations of colour, faces, textures, and objects correlate with content-specific activity in ventral extrastriate cortex, dissociating conscious visual content from sensory input. ↩

  39. daSilva Morgan, K., Collerton, D., Firbank, M. J., Schumacher, J., Ffytche, D. H., & Taylor, J.-P. (2025). Visual cortical activity in Charles Bonnet syndrome: testing the deafferentation hypothesis. Journal of Neurology, 272, Article 76. Multimodal evidence (fMRI, EEG, TMS) showing reduced bottom-up activation paired with increased cortical excitability and theta-band power in CBS, with cortical excitability scaling with hallucination severity, consistent with the deafferentation account. ↩

  40. De Ridder, D., Elgoyhen, A. B., Romo, R., & Langguth, B. (2011). Phantom percepts: tinnitus and pain as persisting aversive memory networks. Proceedings of the National Academy of Sciences, 108(20), 8075-8080. Frames tinnitus and chronic pain as the same phenomenon at different sensory addresses: phantom percepts arising from maladaptive plasticity in deafferented cortex, broadcast as if from a peripheral source that no longer exists. ↩

  41. Baffour-Awuah, K. A., Bridge, H., Engward, H., MacKinnon, R. C., Ip, I. B., & Jolly, J. K. (2024). The missing pieces: an investigation into the parallels between Charles Bonnet, phantom limb and tinnitus syndromes. Therapeutic Advances in Ophthalmology, 16, 25158414241302065. Cross-modal review arguing that Charles Bonnet syndrome, phantom limb syndrome and tinnitus share a single underlying architecture (sensory deafferentation followed by cortical hyperexcitability and maladaptive reorganisation), and that lessons from each translate to the others; the empirical basis for treating the three as one family in the deafferentation row of this article's table. ↩

  42. Hadjikhani, N., Sanchez del Rio, M., Wu, O., Schwartz, D., Bakker, D., Fischl, B., Kwong, K. K., Cutrer, F. M., Rosen, B. R., Tootell, R. B. H., Sorensen, A. G., & Moskowitz, M. A. (2001). Mechanisms of migraine aura revealed by functional MRI in human visual cortex. Proceedings of the National Academy of Sciences, 98(8), 4687-4692. High-field fMRI during spontaneous visual aura found a focal BOLD change propagating across occipital cortex, retinotopically matched to the moving scotoma and consistent with cortical spreading depression. ↩

  43. McLeod, G. A., Josephson, C. B., Engbers, J. D. T., Cooke, L. J., & Wiebe, S. (2025). Mapping the migraine: Intracranial recording of cortical spreading depression in migraine with aura. Headache: The Journal of Head and Face Pain, 65(4), 658-665. Depth electrodes captured a spontaneous migraine aura wave live, about eighty years after Karl Lashley inferred the same march and speed from pencil maps of his own scintillating scotomas. ↩

  44. Stretanski, M. F., Hu, Y., & Munakomi, S. (2025). Thalamic Pain Syndrome. In StatPearls [Internet]. Treasure Island (FL): StatPearls Publishing. NCBI Bookshelf clinical overview of thalamic pain syndrome (Dejerine-Roussy), its presentation, and the relevant thalamic nuclei. ↩

  45. Klit, H., Finnerup, N. B., & Jensen, T. S. (2009). Central post-stroke pain: clinical characteristics, pathophysiology, and management. The Lancet Neurology, 8(9), 857-868. Review of the symptoms, anatomy, and management of central post-stroke pain following thalamic and extra-thalamic lesions of the somatosensory pathway. ↩

  46. Ropero Peláez, F. J., & Taniguchi, S. (2016). The Gate Theory of Pain Revisited: Modeling Different Pain Conditions with a Parsimonious Neurocomputational Model. Neural Plasticity, 2016, 4131395. Reformulation of the classical Gate Control theory and its extensions to phantom limb pain, wind-up, and chronic/neuropathic conditions. ↩

  47. Cox, J. J., Reimann, F., Nicholas, A. K., Thornton, G., Roberts, E., Springell, K., et al. (2006). An SCN9A channelopathy causes congenital inability to experience pain. Nature, 444(7121), 894-898. Identification of homozygous loss-of-function mutations in SCN9A (encoding the Nav1.7 sodium channel) as the cause of congenital insensitivity to pain in three Pakistani families; affected individuals have normal touch, proprioception and temperature discrimination but cannot generate the action potentials that signal nociception, leading to painless fractures, burns, and self-injury that often go undetected until severe damage has accumulated. ↩

  48. Gianlorenço, A. C., Costa, V., Fabris-Moraes, W., Menacho, M., Gola Alves, L., Martinez-Magallanes, D., & Fregni, F. (2024). Cluster analysis in fibromyalgia: a systematic review. Rheumatology International, 44, 2389-2402. PRISMA synthesis of 39 cluster-analysis studies in fibromyalgia: recurring subgroups by severity, pain- versus fatigue-predominant symptom profiles, thermal pain sensitivity, personality and psychological-vulnerability profiles, and differential treatment response. ↩

  49. Vincent, A., Hoskin, T. L., Whipple, M. O., Clauw, D. J., Barton, D. L., Benzo, R. P., & Williams, D. A. (2014). OMERACT-based fibromyalgia symptom subgroups: an exploratory cluster analysis. Arthritis Research & Therapy, 16(5), 463. OMERACT-based clustering of 581 female fibromyalgia patients identified four symptom subgroups: lowest-burden adaptive, pain-dominant with lower mood symptoms, mood-dominant with lower pain, and high-impact across all symptoms; validated in 478 additional patients. ↩

  50. Maurel, S., Gimenez, L., Alegre-Martin, J., & Castro-Marrero, J. (2023). Hierarchical cluster analysis based on clinical and neuropsychological symptoms reveals distinct subgroups in fibromyalgia: a population-based cohort study. Biomedicines, 11(10), 2867. In 251 primary-care fibromyalgia patients, separate cluster solutions on somatic symptoms and on neuropsychological variables (catastrophizing, acceptance, mindfulness, surrender, affect) yielded distinct profiles within the same cohort. ↩

  51. Häuser, W., Perrot, S., Fitzcharles, M. A., & Clauw, D. J. (2018). Unravelling fibromyalgia—steps toward individualized management. The Journal of Pain, 19(2), 125-134. Argues that fibromyalgia's wide variation and modest average treatment effects require graduated, individualized management rather than one protocol for all patients. ↩

  52. Moore, A., Bidonde, J., Fisher, E., Häuser, W., Bell, R. F., Perrot, S., Makri, S., & Straube, S. (2025). Effectiveness of pharmacological therapies for fibromyalgia syndrome in adults: an overview of Cochrane Reviews. Rheumatology, 64(5), 2385-2394. Overview of 21 Cochrane reviews (87 trials, 17,631 patients): duloxetine, pregabalin and milnacipran give substantial (≥50%) pain relief in roughly 1 in 10 adults with moderate-to-severe fibromyalgia over 4–12 weeks, with no trustworthy efficacy beyond six months; adverse-event withdrawals run 6–9% above placebo and no evidence identifies in advance who benefits or who is harmed. ↩

  53. Owen, A. M., Coleman, M. R., Boly, M., Davis, M. H., Laureys, S., & Pickard, J. D. (2006). Detecting awareness in the vegetative state. Science, 313(5792), 1402. Original demonstration of covert command-following (mental imagery of tennis vs. spatial navigation) in a patient meeting all clinical criteria for the vegetative state, evidence that broadcast can occasionally cross even when behavioural channels are fully silent. ↩

  54. Wicken, M., Keogh, R., & Pearson, J. (2021). The critical role of mental imagery in human emotion: insights from fear-based imagery and aphantasia. Proceedings of the Royal Society B, 288(1946), 20210267. Aphantasic participants show flat-line skin-conductance responses when reading fearful first-person scenarios, while showing normal autonomic responses to fearful perceptual images, evidence that voluntary visual imagery functions as an emotional amplifier whose absence selectively dampens imagined (not perceived) threat. ↩

  55. Kvamme, T. L., Scholz, C. O., Monzel, M., Liu, J., & Silvanto, J. (2026). When weak imagery is worse than none: Core aphantasia and hypophantasia relate differently to mental health, mediated by subjective interoception. Neuropsychologia, 222, 109368. Hypophantasia associated with greater alexithymia and worse anxiety and depression than core aphantasia, mediated by subjective interoceptive processing. ↩

  56. Dance, C. J., Jaquiery, M., Eagleman, D. M., Porteous, D., Zeman, A., & Simner, J. (2021). What is the relationship between aphantasia, synaesthesia and autism? Consciousness and Cognition, 89, 103087. Matched comparison of aphantasic and control participants reporting elevated autistic-trait scores in aphantasia, with particular differences on imagination-related subscales. ↩

  57. Grant, S., Norton, S., & Hoekstra, R. A. (2025). Central Sensitivity Symptoms and Autistic Traits in Autistic and Non-Autistic Adults. Autism Research, 18(3), 660-674. Cross-sectional study (n = 534) showing bidirectional overlap between autistic traits and central sensitivity symptoms, elevated CSS scores in autistic adults without a CSS diagnosis, and sensory sensitivity as a significant predictor of CSS symptom load. ↩

  58. Liu, B., Kamath, A., Grunde-McLaughlin, M., Han, W., & Krishna, R. (2025). Visual Representations inside the Language Model. arXiv preprint, arXiv:2510.04819. Probes image key-value tokens in LLaVA-OneVision, Qwen2.5-VL, and Llama-3-LLaVA-NeXT: internal value states encode enough structure for zero-shot segmentation, correspondence, and referring-expression tasks, yet on BLINK Art Style questions the correct perception is present internally but not surfaced in the final output in 33.3% of cases. Supports the aphantasia parallel: visual competence without a checked inner picture reaching the report. ↩

  59. Ren, L., Yu, W., Yu, R., & Wang, X. (2026). NoLan: Mitigating Object Hallucinations in Large Vision-Language Models via Dynamic Suppression of Language Priors. arXiv preprint, arXiv:2602.22144. Ablations attribute object hallucination predominantly to the language decoder, not the vision encoder; a training-free contrast between multimodal and text-only inputs improves POPE accuracy by up to 6.45 points (LLaVA-1.5 7B) and 7.21 points (Qwen-VL 7B), showing fluent visual answers can run on language priors alone. ↩

  60. Narnaware, V., Gupta, A., Zhai, K., Wang, Z., & Shah, M. (2026). Seeing to Ground: Visual Attention for Hallucination-Resilient MDLLMs. arXiv preprint, arXiv:2603.25711. Frames multimodal hallucination as an objective mismatch: parallel decoders rank tokens by textual likelihood without verifying localized visual support, letting language shortcuts finalize ungrounded answers; a training-free VISAGE re-ranking penalty on cross-attention entropy improves HallusionBench and MMMU-val. Cited as the missing verification channel: no independent check beneath the report. ↩

  61. Dassen, M., Kotula, R., Murray, K., Yates, A., Lawrie, D., Kayi, E., Mayfield, J., & Duh, K. (2026). FACTUM: Mechanistic Detection of Citation Hallucination in Long-Form RAG. arXiv preprint, arXiv:2601.05866. Introduces mechanistic scores over contextual alignment, attention-sink usage, parametric force, and pathway alignment to detect citation hallucination in long-form RAG. FACTUM outperforms prior internal-signal baselines by up to 37.5% in AUC, suggesting that unfaithful citations leave a measurable trace in the same pathways that produce them. ↩

  62. LeCun, Y. (2022). A Path Towards Autonomous Machine Intelligence. OpenReview position paper. Introduces Joint Embedding Predictive Architectures (JEPA): hierarchical, energy-based world models that learn predictive latent representations instead of generating tokens or pixels directly. ↩

  63. Maes, L., Le Lidec, Q., Scieur, D., LeCun, Y., & Balestriero, R. (2026). LeWorldModel: Stable End-to-End Joint-Embedding Predictive Architecture from Pixels. arXiv preprint, arXiv:2603.19312. First JEPA that trains stably end-to-end from raw pixels with two losses (prediction + SIGReg), ~15M parameters trainable on a single GPU, with a latent space that reliably detects physically implausible events. ↩

  64. Terver, B., Balestriero, R., Dervishi, M., Fan, D., Garrido, Q., Nagarajan, T., Sinha, K., Zhang, W., Rabbat, M., LeCun, Y., & Bar, A. (2026). A Lightweight Library for Energy-Based Joint-Embedding Predictive Architectures. arXiv preprint, arXiv:2602.03604. Open-source library demonstrating image, video, and action-conditioned video JEPAs trainable on a single GPU, with ablations showing the regularisation components needed to prevent representation collapse. ↩

  65. Ha, D., & Schmidhuber, J. (2018). World Models. arXiv preprint, arXiv:1803.10122. Foundational modern formulation of a learned latent world model (encoder + recurrent dynamics + controller), which agents can use to "dream" trajectories internally; the direct intellectual predecessor of later predictive-latent approaches including JEPA. Builds on Schmidhuber's long-running line of work on predictive world models for reinforcement learning, dating to the early 1990s. ↩

  66. Destrade, M., Bounou, O., Le Lidec, Q., Ponce, J., & LeCun, Y. (2026). Value-Guided Action Planning with JEPA World Models. arXiv preprint, arXiv:2601.00844. Shapes JEPA latent space so that the goal-conditioned value function is approximated by a distance between state embeddings, improving planning performance. ↩

  67. Sekeres, M. J., Winocur, G., & Moscovitch, M. (2018). The hippocampus and related neocortical structures in memory transformation. Neuroscience Letters, 680, 39-53. Trace Transformation Theory: with time and experience, episodic memories transform at both neural and psychological levels along the long axis of the hippocampus and into medial prefrontal cortex. Detailed perceptual representations (posterior hippocampus), gist (anterior hippocampus), and schemas (mPFC) coexist and remain in dynamic flux throughout a memory's lifetime, going from detailed to schematic and possibly back again, rather than handing off cleanly from hippocampus to neocortex. ↩

  68. Lin, K., Snell, C., Wang, Y., Packer, C., Wooders, S., Stoica, I., & Gonzalez, J. E. (2025). Sleep-time Compute: Beyond Inference Scaling at Test-time. arXiv preprint, arXiv:2504.13171 (Letta and UC Berkeley). Introduces sleep-time compute: while an LLM is idle between user queries it reasons offline over a persisted context $c$ and re-represents it as $c'$ that subsequent test-time queries consume, cutting test-time compute by roughly fivefold for the same accuracy on Stateful GSM-Symbolic and Stateful AIME, with a case study on an agentic software-engineering task. ↩

  69. Baranov, V. (2026). Sleeping LLM: Biologically-Inspired Memory Consolidation for Continual Learning in Local Language Models. Open-source preprint, vbario/sleeping-llm. Working architecture for sleep-wake memory consolidation in a locally-running LLM: wake-phase conversation logging, sleep-phase fact extraction into structured Q&A pairs, LoRA fine-tuning under MEMIT-style null-space-constrained weight editing, spaced-repetition replay buffer with priority decay, and a validator gate that benchmarks the model before and after each sleep cycle to reject runs that drift. Demonstrated on a 4-bit quantized Llama 3.2 3B model on 8GB of unified memory: after a single sleep cycle the model recalls specific facts from prior conversation with an empty context window, with successive cycles strengthening recall as the spaced-repetition account predicts. Cited here as the cleanest worked example of Stage 3 in an LLM agent: a parameter-efficient, schema-compatible write back into the weights gated against catastrophic forgetting. ↩

  70. Du, P. (2026). Memory for Autonomous LLM Agents: Mechanisms, Evaluation, and Emerging Frontiers. arXiv preprint, arXiv:2603.07670. Survey of agent-memory designs from 2022 through early 2026, formalising memory as a write–manage–read loop coupled to perception and action, and organising the design space across temporal scope (working / episodic / semantic / procedural), representational substrate, and control policy (heuristic, prompted-self, RL-learned). Identifies continual consolidation, the episodic-to-semantic transition where recurring episodes graduate to durable schema-level structure, as the first open challenge and explicitly notes that most current systems handle this with crude heuristics or periodic LLM-driven summarisation, supporting the article's claim that production agents do not run Stage 3. ↩

  71. Shinde, S. (2026). SCM: Sleep-Consolidated Memory with Algorithmic Forgetting for Large Language Models. arXiv preprint, arXiv:2604.20943. Presents Sleep-Consolidated Memory, a prototype LLM memory architecture with limited working memory, multi-dimensional importance tagging, NREM/REM-style offline consolidation, value-based forgetting, and a self-model. In benchmark tests it reports perfect recall over ten-turn conversations while reducing low-value memory noise by 90.9% through adaptive forgetting. ↩

  72. Zhang, J., Zhang, C., Chen, S., Huang, Z., Zheng, P., Wang, Z., Guo, P., Mo, F., Bae, S.-H., Zou, J., Wei, J., & Yang, Y. (2026). Lightweight LLM Agent Memory with Small Language Models. arXiv preprint, arXiv:2604.07798. Three specialised small language models (Controller for query planning, Selector for candidate verification, Writer for incremental MTM summarisation) plus an offline large-context model for long-term-memory consolidation, with consistent F1 gains over prior agentic-memory baselines on LoCoMo. ↩

  73. Dossa, R. F. J., Arulkumaran, K., Juliani, A., Sasai, S., & Kanai, R. (2024). Design and evaluation of a global workspace agent embodied in a realistic multimodal environment. Frontiers in Computational Neuroscience, 18, 1352685. Embodied agent based on Global Workspace Theory navigating 3D audiovisual environments; the broadcast architecture outperforms standard recurrent baselines at smaller working-memory sizes. ↩

  74. Nakanishi, J., Baba, J., Yoshikawa, Y., Kamide, H., & Ishiguro, H. (2025). Hypothesis on the Functional Advantages of the Selection-Broadcast Cycle Structure: Global Workspace Theory and Dealing with a Real-Time World. arXiv preprint, arXiv:2505.13969. Argues that the combined Selection-Broadcast cycle (not Selection or Broadcast alone) is what gives GWT-inspired agents their dynamic, experience-based, and real-time adaptation advantages. ↩

  75. Boahen, K., Wunderlich, T., & Petrovici, M. A. (2026). Phasor Agents: Oscillatory Graphs with Three-Factor Plasticity and Sleep-Staged Learning. arXiv preprint, arXiv:2601.04362. Spiking agent architecture implementing three-factor local plasticity (eligibility traces gated by sparse neuromodulatory signals) with separated wake-time tagging and offline consolidation phases inspired by NREM/REM sleep staging. Reports 67% expansion of stable learning capacity and +45.5% improvement in downstream task performance through REM-like replay relative to non-staged baselines, on a substrate where consolidation is a continuous local property rather than a scheduled fine-tuning job. ↩

  76. Petrovici, M. A., Senn, W., & Maass, W. (2026). Sleep-Based Homeostatic Regularization for Stabilizing Spike-Timing-Dependent Plasticity in Recurrent Spiking Neural Networks. arXiv preprint, arXiv:2601.08447. Implements the synaptic homeostasis hypothesis on recurrent SNNs: periodic offline phases suppress external input, drive stochastic weight decay toward a homeostatic baseline, and allow spontaneous activity to consolidate memory. Sleep cycles at 10–20% of training time prevent unbounded weight growth, catastrophic forgetting, and loss of representational diversity on MNIST-class benchmarks without data-specific hyperparameter tuning, evidence that sleep-staged regularisation is a substrate-level property and not an LLM-only abstraction. ↩

  77. Andreou, A. G., Mastella, M., & Chakrabartty, S. (2025). Genesis: A Spiking Neuromorphic Accelerator With On-chip Continual Learning. arXiv preprint, arXiv:2509.05858. Mixed-signal spiking accelerator implementing activity-dependent metaplasticity and custom memory mapping for on-chip continual learning in 65nm CMOS, reaching 74.6% accuracy on split-MNIST at 17.08 mW. Cited as silicon-level evidence that continuous on-chip consolidation, the hardware analogue of Stage 3, is now an implemented technology rather than a research sketch. ↩

License

This work is licensed under a Creative Commons Attribution 4.0 International License.