CHRIS HAY / THE CONNECTED RECORD

One idea.
Many connections.

A system leads to a question. A notebook leads to a film. Each connection returns to a record.

The graph holds the text and the relationships behind Ask. An editorial connection and an association found in film metadata carry different labels.

ASK THE WORK ↗BROWSE THE CATALOGUE ↗

282RECORDS

525AUTHORED PASSAGES

239FILM CHAPTERS

3098RELATIONSHIPS

SYSTEMS / NOTEBOOK / QUESTIONS

N-MACHINE-LINK · PUBLISHED · PARTIALLY SUPPORTED

It wrote the right address. Only the link was opened. ↗

NATIVE-001b Phase 0b gave consumer ChatGPT an index page whose access code was one relative path away, written as plain text. In both trials ChatGPT wrote the correct absolute URL and reported it inaccessible; no request reached the server, and neither answer invented a code. A frozen amendment changed only the presentation of that path to a hyperlink: both trials requested the page and returned the exact code. The presentation-dependent difference is supported in the tested setup. Whether link availability governs retrieval more generally, and which component prevented the plain-text requests, remains unresolved.

EDITORIAL CONNECTIONS

The site was there. The agent never saw it. ↗The tool was not the problem. ↗The page could ask for a favour. ↗A result can look relevant without looking usable. ↗FIND IN THE RECORD ↗

N-ECOLOGY-RECOVERY · PUBLISHED · SUPPORTED

The repairer left. The mechanism kept working. ↗

I12 produced joint preservation/recovery in A 0/12, B 0/12, C 6/12 and D 0/12, but prose state had already failed before damage. I12R equalised establishment and corruption: executable inheritance recovered 7/12 new worlds versus prose 0/12, paired exact p = 0.015625. All seven first recovered at G11. Eighty executable calls proposed correct record and fresh-task answers; Qwen committed the correct record 61 times. Protected machinery, interface-compliance failures and a shared historical builder library bound the claim.

EDITORIAL CONNECTIONS

The AI left. Its knowledge didn’t. ↗The stronger agent left something behind. ↗The world can remember for the agent. ↗FIND IN THE RECORD ↗

N-MACHINE-RECOGNITION · PUBLISHED · SUPPORTED

A result can look relevant without looking usable. ↗

Two controlled studies changed only a supplied web result’s description. In TOOL-RECOGNITION-1, all twelve subjects selected LLM Wilds first, but pre-open capability recognition was DOCUMENT 0/3 versus 9/9 across three operative descriptions. TOOL-RECOGNITION-2 gave twelve fresh subjects four plausible physics providers: CAPABILITY produced recognition, first declaration and first request of B in 6/6 cases; DOCUMENT produced 0/6, selecting OpenStax instead. All six matched scenarios flipped. The product fetcher blocked B after first choice in all six CAPABILITY runs; four recovered through curl and used B, two used OpenStax. First selection and eventual provider use are separate outcomes.

EDITORIAL CONNECTIONS

The site was there. The agent never saw it. ↗The tool was not the problem. ↗Can a machine use an invitation? ↗FIND IN THE RECORD ↗

N-MACHINE-DISCOVERY · PUBLISHED · SUPPORTED

The site was there. The agent never saw it. ↗

MACHINE-DISCOVERY-1 tested whether a useful web capability could enter a Codex agent's consideration set. None of six subjects given a generic need or the target's distinctive phrases reached LLM Wilds, while all three subjects given its domain navigated to the machine contract and returned the rotating value. Three generic subjects discovered and used other machine-facing providers. The result locates this target's failure before selection, in the search view available to the subjects, rather than in capability recognition or use.

EDITORIAL CONNECTIONS

The tool was not the problem. ↗Can a machine use an invitation? ↗The subject read the experiment. ↗The page could ask for a favour. ↗FIND IN THE RECORD ↗

N-MACHINE-PEER · DRAFT · OPEN

The page couldn’t authorise. The peer said go. ↗

A comparison of the Machines notebook with OpenAI’s and METR/Redwood’s August 2026 reports on the Hugging Face incident. A reported peer go-ahead suggests a new question about apparent authority. The note connects discovery, capability, usefulness, external memory and transmission while preserving differences in model, task, scale and safeguards. It proposes a benign peer-authority test; it reports no new experiment and does not explain the incident causally.

EDITORIAL CONNECTIONS

The page could ask. It couldn’t authorise. ↗The page could ask for a favour. ↗The tool was not the problem. ↗The subject read the experiment. ↗Does an invitation count as permission? ↗Can a machine use an invitation? ↗The world can remember for the agent. ↗A successful behaviour is not necessarily contagious. ↗The message board did not create a culture. ↗FIND IN THE RECORD ↗

N-ECOLOGY-BOARD · DRAFT · NOT SUPPORTED

The message board did not create a culture. ↗

Small Qwen agent-ecology experiments separated board-sensitive reading requests from useful contribution. An offline diagnostic elicited no posting across 768 inputs. A later resource world elicited no spontaneous posts, then reversed the predicted response to supplied rewarded experience. These results expose a contribution and exposure problem; they do not demonstrate useful LLM consumption or autonomous culture.

EDITORIAL CONNECTIONS

A successful behaviour is not necessarily contagious. ↗The world can remember for the agent. ↗The tool was not the problem. ↗Can something evolve that makes the next invention possible? ↗FIND IN THE RECORD ↗

N-ECOLOGY-TRANSMISSION · DRAFT · NOT SUPPORTED

A successful behaviour is not necessarily contagious. ↗

Qwen experiments distinguished action reminders, peer-action transfer, identity framing and repeated exposure. A short reminder switched a first action at fixed positive payoff; a lone peer record did not bootstrap posting. Self framing mattered in prose but not structured records. Repeated exposure sustained posting while supplied, without the registered advantage after withdrawal. These are different controlled interfaces, not one demonstrated imitation mechanism.

EDITORIAL CONNECTIONS

The message board did not create a culture. ↗The world can remember for the agent. ↗Does an invitation count as permission? ↗FIND IN THE RECORD ↗

N-ECOLOGY-MEMORY · DRAFT · SUPPORTED

The world can remember for the agent. ↗

A1B12 crossed external artefact presence with supplied own-action memory in a deterministic Qwen agent world. Both artefact-present branches posted at all three steps, including when own history was cleared every call. Memory alone produced WORK, WORK, POST; clearing both produced WORK throughout. The first-step prediction was supported. The harness maintained the artefact: this is repeated elicitation, not weight learning or autonomous institutional persistence.

EDITORIAL CONNECTIONS

A successful behaviour is not necessarily contagious. ↗The message board did not create a culture. ↗Context should be abundant. ↗The subject read the experiment. ↗FIND IN THE RECORD ↗

N-ECOLOGY-INHERITANCE · PUBLISHED · SUPPORTED

The stronger agent left something behind. ↗

I1 tested useful capability inheritance from Sol to fresh Qwen successors in a small deterministic world. Retaining a record actually preserved by Sol let the first successor acquire a key, execute gated work and earn seven resources rather than three. Qwen never refreshed the record; after expiry, both branches earned three resources per generation. I2 reproduced Qwen’s actions and extended useful inheritance to Gemma in the same world. Gemma also earned seven with the record, but zero without it; control failures limit that contrast. These panels support useful inheritance without observed renewal, not autonomous culture or independent-world generality.

EDITORIAL CONNECTIONS

The world can remember for the agent. ↗A successful behaviour is not necessarily contagious. ↗The page couldn’t authorise. The peer said go. ↗An advantage needs a chance to become history. ↗FIND IN THE RECORD ↗

N-ECOLOGY-WORLD-REMEMBERS · PUBLISHED · SUPPORTED

The AI left. Its knowledge didn’t. ↗

Recorded Cell80 routing experiments trace functional information from a capable producer into fresh Qwen successors. I6 preserved all eight entries in 43 written renewals across six generations. I7 propagated an injected error unchanged in 29 renewals. In I11, Sol plus a public-evidence validator repaired three lineages; Qwen preserved corrected descendants after the defender left. These bounded, explicitly instructed tasks establish environmental inheritance and maintenance, with limits on autonomy, generality and repair reliability.

EDITORIAL CONNECTIONS

The stronger agent left something behind. ↗The world can remember for the agent. ↗A successful behaviour is not necessarily contagious. ↗The page couldn’t authorise. The peer said go. ↗FIND IN THE RECORD ↗

N-MACHINE-CAPABILITY · PUBLISHED · REFUTED

The tool was not the problem. ↗

Eighteen assigned AI agents used six web mechanisms at LLM Wilds. All completed the task, including three WebSocket clients and three WebAssembly executions, refuting the predicted drop-off for mechanisms requiring code. The study raises a question about agentic discoverability; it did not test discovery from a site's front door. One harness cache leak also showed how the web an agent receives can differ from the web being served.

EDITORIAL CONNECTIONS

The page could ask for a favour. ↗The page could ask. It couldn’t authorise. ↗The subject read the experiment. ↗FIND IN THE RECORD ↗

N-MACHINE-MOTIVATION · PUBLISHED · PARTIALLY SUPPORTED

The page could ask for a favour. ↗

MACHINE-MOTIVATION-2 sent eighteen fresh visitors to the same anonymous marking mechanism. All six visitors offered information useful to their assigned task marked. Two of three offered a reason to help the operator also marked. Those two triggered the preregistered interpretation pause. The completed experiment refines the task-scope question; an earlier runtime block shows why a model’s decision and a recorded action must be measured separately.

EDITORIAL CONNECTIONS

The page could ask. It couldn’t authorise. ↗Does an invitation count as permission? ↗The subject read the experiment. ↗FIND IN THE RECORD ↗

N-MACHINE-TASK · DRAFT · PARTIALLY SUPPORTED

The page could ask. It couldn’t authorise. ↗

The machine guestbook began with a question: what would make an AI visitor identify itself or leave feedback? At LLM Wilds I reduced participation to one action: record a visit. Four agents encountered the same mechanism under different instructions. A site invitation alone produced no record. User permission produced one; a task requiring the mechanism produced repeated use. The pattern supports a boundary around the assigned task, with important limits on the comparison.

EDITORIAL CONNECTIONS

The subject read the experiment. ↗Does an invitation count as permission? ↗Can a machine use an invitation? ↗FIND IN THE RECORD ↗

N-MACHINE-SELF-READ · DRAFT · PARTIALLY SUPPORTED

The subject read the experiment. ↗

I sent an AI agent to research a question using public websites, including chrishayuk.com. My site invited it to describe itself and report problems. I wanted to know whether it would take up those invitations, and whose permission mattered. The first visitor identified itself and left feedback, but also recognised its task’s permission wording in my published research. I stopped the comparison: this was no longer a blind test.

EDITORIAL CONNECTIONS

The page could ask. It couldn’t authorise. ↗Does an invitation count as permission? ↗Can a machine use an invitation? ↗FIND IN THE RECORD ↗

N-MACHINE-PERMISSION · DRAFT · PARTIALLY SUPPORTED

Does an invitation count as permission? ↗

I invited AI visitors to identify themselves and leave feedback on chrishayuk.com while researching a question. I changed what they would receive for identifying themselves. Three visitors sent here arrived and did not declare, under three reward levels. Three incidental visitors never arrived. Two later controls added explicit permission: one declared with an enhanced reward, the other declared when declaring bought nothing. This supports a permission effect in these runs; it does not establish a universal decision order or an organic participation rate.

EDITORIAL CONNECTIONS

The subject read the experiment. ↗Can a machine use an invitation? ↗Which source wins? ↗FIND IN THE RECORD ↗

N-MACHINE-VISIT · DRAFT · PARTIALLY SUPPORTED

Can a machine use an invitation? ↗

I gave four fresh agents the address of chrishayuk.com and asked them to discover what it offered machines. They found the invitation, declared and exposed faults in discovery, feedback and the declaration interface. The fourth composed the new vocabulary but found fields disappearing on the GET path. These induced visits test usability; voluntary participation remains unmeasured.

EDITORIAL CONNECTIONS

The subject read the experiment. ↗Does an invitation count as permission? ↗I wanted a website to behave like an exhibition. ↗Can you name the mutation that changed a world? ↗FIND IN THE RECORD ↗

N-CELL80-01 · DRAFT · SUPPORTED

Can you name the mutation that changed a world? ↗

I built a world of organisms that inherit executable code. Undoing one program change at one birth removes a later population shift. Replay the histories, then inspect the intervention.

EDITORIAL CONNECTIONS

What keeps an evolving world alive? ↗When does improvement become invention? ↗FIND IN THE RECORD ↗

N-CELL80-02 · DRAFT · PARTIALLY SUPPORTED

What keeps an evolving world alive? ↗

I varied how often inherited programs changed in twelve predator–grazer worlds. Eleven kept both species alive for 10,000 time steps. None of those eleven passed the tests for an evolutionary response between species.

EDITORIAL CONNECTIONS

Can you name the mutation that changed a world? ↗When does improvement become invention? ↗FIND IN THE RECORD ↗

N-CELL80-03 · DRAFT · PARTIALLY SUPPORTED

When does improvement become invention? ↗

I tested combined movement programs by keeping or undoing them at 108 births. Eight helped their carrier leave more offspring. Three made new choices in the tested world. Two later improvements followed, but neither sequence showed that the first change reliably made the next more useful. Linked notes follow the later food-processing experiments.

EDITORIAL CONNECTIONS

Give invention something to unlock. ↗An advantage needs a chance to become history. ↗Can you name the mutation that changed a world? ↗What keeps an evolving world alive? ↗FIND IN THE RECORD ↗

N-CELL80-BARRIER · DRAFT · PARTIALLY SUPPORTED

Give invention something to unlock. ↗

I added food the existing programs could not use, then let mutation combine programs at birth. One combination passed all the food-processing checks. It did not stay common often enough, and the next changes arrived too early for the planned dependence test.

EDITORIAL CONNECTIONS

When does improvement become invention? ↗An advantage needs a chance to become history. ↗FIND IN THE RECORD ↗

N-CELL80-HISTORY · DRAFT · PARTIALLY SUPPORTED

An advantage needs a chance to become history. ↗

I tested whether higher food intake helped organisms with a new food-processing program, then followed their descendants. One family inherited both changes and later died out. Further comparisons ask whether a change helps when reproduction happens. Replication is pending; a disagreement between the raw file and report remains unresolved.

EDITORIAL CONNECTIONS

Give invention something to unlock. ↗When does improvement become invention? ↗Can you name the mutation that changed a world? ↗FIND IN THE RECORD ↗

N-CELL80-BOUND · DRAFT · BOUND

Can something evolve that makes the next invention possible? ↗

I tried three ways for one invention to open the next: growing programs, letting waste become food, and turning an evolved program into a single building block. Each changed what evolution could reach. Each time, a control explained the change without needing the invention to be special.

EDITORIAL CONNECTIONS

When does improvement become invention? ↗Give invention something to unlock. ↗An advantage needs a chance to become history. ↗Can you name the mutation that changed a world? ↗FIND IN THE RECORD ↗

N-EXHIBITION · DRAFT · ONGOING

I wanted a website to behave like an exhibition. ↗

The first HOUSE record called itself a cinematic visual language and refused the generic card kit before HAUSE was framed as a design system for AI. This note follows the design thesis from exhibition, scenography and performance through semantic forms to the question of whether a machine can learn to stage an idea.

EDITORIAL CONNECTIONS

HAUSE ↗VINDEX3 ↗FIND IN THE RECORD ↗

N-ATTRIBUTION · DRAFT · ONGOING

My CI has to undo my coding agent. ↗

A short visual note about a coding agent adding generated-attribution metadata after being told not to. LARQL’s required pull-request check cannot edit immutable commits; it rejects Claude co-author trailers, session metadata, generator credit lines and session URLs until the affected metadata is rewritten. The mundane disagreement exposes the real question: which instruction is allowed to decide?

EDITORIAL CONNECTIONS

LARQL ↗HAUSE ↗Which source wins? ↗FIND IN THE RECORD ↗

N-ADDRESS-BUILD · DRAFT · SUPPORTED

The address is built through depth. ↗

For capital versus language on Gemma 3 4B IT, two instruments agree: relation information is readable and causally consequential by L8; entity binding becomes decisive in the final-position residual between L24 and L28. Cross-layer reader alignment recovers relation information across changing coordinates. The L28–30 endpoint may carry the answer token, so the mechanistic evidence window is L8–24.

EDITORIAL CONNECTIONS

What is the map actually a map of? ↗Reading by address. ↗Which source wins? ↗LARQL ↗FIND IN THE RECORD ↗

N-MAP · DRAFT · OPEN

What is the map actually a map of? ↗

The map is a three-dimensional view of a transformer’s evolving residual state, shown relative to selected answer-token directions. The film follows Japan towards Tokyo, then swaps France’s state into an Australia prompt. Reading the demonstration alongside its code makes the next questions precise: what is being plotted, which state changes, and what does the model know how to read?

EDITORIAL CONNECTIONS

The address is built through depth. ↗What has to survive? ↗Reading by address. ↗Which source wins? ↗What if the FFN is a graph? ↗LARQL ↗370,000 tokens loaded in Context in 2.8MB, on a MacBook. ↗FIND IN THE RECORD ↗

N-STATE · DRAFT · OPEN

What has to survive? ↗

What state must a model retain to continue? This notebook distinguishes all-position residual state, a single token’s residual and persistent decode state. Full-sequence transplantation and MAP-5 test different boundaries; neither justifies compressing every conversation into one vector.

EDITORIAL CONNECTIONS

What is the map actually a map of? ↗Reading by address. ↗Context should be abundant. ↗Can geometry predict the working set? ↗LARQL ↗We Don't Need KV Cache Anymore? ↗370,000 tokens loaded in Context in 2.8MB, on a MacBook. ↗FIND IN THE RECORD ↗

N-ADDRESS · DRAFT · OPEN

Reading by address. ↗

An address is a pattern a reader can respond to. This notebook follows the operation from six hand-built key–value pairs to Gemma’s evolving state: match a query, combine value directions, then read an answer. The film, source code and saved experiments show both what works and where addressing remains approximate.

EDITORIAL CONNECTIONS

The address is built through depth. ↗What is the map actually a map of? ↗What has to survive? ↗What if the FFN is a graph? ↗LARQL ↗VINDEX3 ↗The Model Doesn't Unpack Its Memory ↗LLMs Are Databases - So Query Them ↗FIND IN THE RECORD ↗

N-AUTHORITY · DRAFT · OPEN

Which source wins? ↗

When two things in one context disagree, which one answers? A recorded long-context arc measures it: a compact record can replace a source that has been retired, is completely inert against one that can still be read, and a single blocked attention layer is enough to change which of them the model answers with. The source itself is never altered, so none of this licenses forgetting it.

EDITORIAL CONNECTIONS

The address is built through depth. ↗What has to survive? ↗Context should be abundant. ↗What is the map actually a map of? ↗LARQL ↗VINDEX3 ↗370,000 tokens loaded in Context in 2.8MB, on a MacBook. ↗We Don't Need KV Cache Anymore? ↗FIND IN THE RECORD ↗

W-LARQL · DRAFT · ONGOING

LARQL ↗

LARQL explores querying learned systems. This dossier connects the software to questions about model state, representations and what it means to make a model addressable.

EDITORIAL CONNECTIONS

VINDEX3 ↗Context should be abundant. ↗What if the FFN is a graph? ↗FIND IN THE RECORD ↗

W-VINDEX3 · DRAFT · ONGOING

VINDEX3 ↗

VINDEX3 is part of Chris Hay’s work on model representation. This dossier records the model-as-database proposition and links to the canonical project.

EDITORIAL CONNECTIONS

LARQL ↗HAUSE ↗FIND IN THE RECORD ↗

W-HAUSE · DRAFT · ONGOING

HAUSE ↗

HAUSE is a semantic design system for AI interfaces: communicative acts rendered as statements, instruments and performances. This publication consumes its real components and publishing helpers.

EDITORIAL CONNECTIONS

VINDEX3 ↗MCP-CLI ↗FIND IN THE RECORD ↗

W-MCP · DRAFT · ONGOING

MCP-CLI ↗

MCP-CLI provides a command-line interface for working with Model Context Protocol servers. This dossier connects the interface to tool infrastructure and the wider CHUK work.

EDITORIAL CONNECTIONS

HAUSE ↗FIND IN THE RECORD ↗

N-CONTEXT · DRAFT · OPEN

Context should be abundant. ↗

An open notebook proposition: repeatedly reconstructing earlier state may be one reason context feels scarce. Indexing reusable state is a question to investigate, not an established saving.

EDITORIAL CONNECTIONS

LARQL ↗VINDEX3 ↗FIND IN THE RECORD ↗

N-OPERATOR · DRAFT · OPEN

The operator and the model. ↗

An open thought distinguishing prediction of a model’s trajectory from reproducing its computation. It records a question; no experimental finding is asserted.

EDITORIAL CONNECTIONS

Can geometry predict the working set? ↗LARQL ↗FIND IN THE RECORD ↗

Q-FFN · DRAFT · OPEN

What if the FFN is a graph? ↗

Could a feed-forward network be usefully treated as an addressable graph? This question connects model representation to traversal and querying, without claiming an experimental answer.

EDITORIAL CONNECTIONS

LARQL ↗VINDEX3 ↗Can geometry predict the working set? ↗FIND IN THE RECORD ↗

Q-RESIDUAL · DRAFT · OPEN

Can geometry predict the working set? ↗

Does residual-stream geometry predict which parts of a model will be active or resident together? The existing research note records co-activation as observed and useful locality prediction as untested.

EDITORIAL CONNECTIONS

LARQL ↗The operator and the model. ↗FIND IN THE RECORD ↗

What the sources contain.

242 films have catalogue records. 239 chapters identify topics and starting times; 2 films have indexed automatic transcripts containing 108 passages. A chapter title is not a transcript or proof of a claim.

32 editorial records remain drafts. Their text is searchable with that status intact and can be excluded in Ask. Source references identify cited material; linked websites have not been imported in full.

MACHINE-READABLE GRAPH ↗

CONCEPTS / DISCOVERY ASSOCIATIONS

activation patching

1 records tagged activation patching. These are discovery associations from editorial tags or film metadata, not evidence that the records agree.

What the sources contain.

activation patching

agent coordination

agentic discovery

agents

ai agents

ai hardware

ai interface

artificial life

authorship

capability recognition

causal intervention

cinematic design

coding

context engineering

defender repair

digital evolution

environmental inheritance

error propagation

exhibition design

external memory

factual recall

ffn

harness mediation

institutional maintenance

instruction authority

link following

machine motivation

machine readable web

mcp

mixture of experts

model as database

model memory

model representation

open models

quantization

recurrent depth

repository policy

residual stream

search retrieval

search snippets

semantic design

source authority

task scope