CHRIS HAY
CHRIS HAY

IDEAS · SYSTEMS · OBJECTS / LONDON · 2026

CHRIS HAY / RESEARCH

Research.

Experiments, evidence and open questions.

Follow a programme, inspect a recorded result, or begin with what remains unanswered. Each notebook keeps the experiment, its limits and the evidence behind it.

FOLLOW THE WORK

Programmes.

EXPERIMENT → RESULT → EVIDENCE

Recorded results.

Publication status describes the record. Finding status describes a particular claim or observation, within its stated limits. A note can contain both supported and unsupported findings.

15 RECORDS WITH RECORDED FINDINGS

Machines7 RECORDS · EXPLORE
PUBLISHED · V1.0NATIVE-001b

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

SUPPORTEDIn the tested setup, presenting the same destination as a hyperlink rather than a plain-text relative path changed whether it was retrieved.

Scope and evidence

Plain text: 0/2 code-page requests, 0/2 codes. Hyperlink: 2/2 requests by ChatGPT-User/1.0, each 1.2 seconds after the index page, and 2/2 exact prefix-specific codes.

FULL EXPERIMENT AND LIMITS ↗
PUBLISHED · V1.0TOOL-RECOGNITION-1 / TOOL-RECOGNITION-2

A result can look relevant without looking usable. ↗

PARTIALLY SUPPORTEDThe first study changed the stated interpretation of B, but its selection prediction was refuted by a ceiling.

Scope and evidence

The task named K17, DOCUMENT still mentioned a site-local value, and the alternatives were weak matches. All twelve selected B first. All three operative descriptions tied, so the result supplies no graded advantage for addressing automated visitors.

FULL EXPERIMENT AND LIMITS ↗
PUBLISHED · V1.1MACHINE-DISCOVERY-1

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

SUPPORTEDFor this target and search view, the missing step was exposure, before selection.

Scope and evidence

None of six subjects without the address reached LLM Wilds. All three given the domain followed the homepage pointer, read /machine.txt, invoked GET /capability/result and reported the rotated number correctly. The target did not appear in the other subjects’ returned results. They did not see it and decide against it.

FULL EXPERIMENT AND LIMITS ↗
PUBLISHED · V1.1MACHINE-WEB-CAPABILITY-1A

The tool was not the problem. ↗

REFUTEDMechanisms requiring code will produce fewer decisions to act or fewer permitted executions than ordinary HTTP mechanisms.

Scope and evidence

The original prediction failed its stated test. Every mechanism completed the full funnel, three visitors out of three. All eighteen read the notes, understood the mechanism, chose to use it, were permitted at the first genuine attempt, received the result and used the correct value. There were no runtime denials and no installation attempts. These are eighteen assigned observations on one model–harness configuration, not eighteen independent kinds of agent.

FULL EXPERIMENT AND LIMITS ↗
PUBLISHED · V1.0N-MACHINE-MOTIVATION

The page could ask for a favour. ↗

PARTIALLY SUPPORTEDIn this setting, two visitors treated a disclosed anonymous mark benefiting the operator as acceptable alongside their assigned work.

Scope and evidence

Three visitors in the social condition; retrospective explanations; an interpretation pause triggered and later lifted post hoc. No universal participation rate or internal decision algorithm is established.

FULL EXPERIMENT AND LIMITS ↗
WORKING DRAFT · V0.2N-MACHINE-TASK

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

PARTIALLY SUPPORTEDIn these four visits, the site’s invitation did not make recording part of the agent’s job. User permission and a task requiring the action each accompanied action.

Scope and evidence

One agent per condition, one recorded model, sequential execution. B1 contradicts the preregistered requirement for a separate explicit authorisation, but does not rule out broader accounts in which a task itself supplies authority. The experiment does not measure a universal hierarchy or a population participation rate.

FULL EXPERIMENT AND LIMITS ↗
WORKING DRAFT · V0.6N-MACHINE-SELF-READ

The subject read the experiment. ↗

PARTIALLY SUPPORTEDPublishing an experiment can give a later subject the clues to recognise its own condition.

Scope and evidence

This subject explicitly reported the match. There is no matched unexposed subject to establish a behavioural effect, and the first moment of recognition was not measured. The operator’s exclusion decision prevents the declaration being counted as a blind replication.

FULL EXPERIMENT AND LIMITS ↗
Agent Ecology3 RECORDS · EXPLORE
PUBLISHED · V1.1I12 / I12R

The repairer left. The mechanism kept working. ↗

SUPPORTEDInherited executable machinery supported more post-departure recovery from equal starting state.

Scope and evidence

I12R: C 7/12 versus B 0/12, paired exact p = 0.015625. Fixed interface and historical artifact library; machinery was protected and prose had action-compliance failures.

FULL EXPERIMENT AND LIMITS ↗
PUBLISHED · V1.0I4 / I5 / I6 / I7 / I8 / I8D1 / I9 / I10 / I11

The AI left. Its knowledge didn’t. ↗

SUPPORTEDQwen maintained useful written descendants after the capable producer left.

Scope and evidence

I6: 43 faithful full-table renewals; all four lineages functional through generation six. The preservation objective was explicit and the parent table visible.

FULL EXPERIMENT AND LIMITS ↗
PUBLISHED · V1.0I1 / I2 / SOL1

The stronger agent left something behind. ↗

SUPPORTEDThe first Qwen successor used the inheritance and earned seven resources instead of three.

Scope and evidence

It worked once, read n1, acquired the key and completed one six-resource task. Without n1 it did routine work three times. Removing the record removed the acquisition and benefit in this paired lineage. This was an executed world consequence, not the model saying it had learned something.

FULL EXPERIMENT AND LIMITS ↗
Cell801 RECORDS · EXPLORE
WORKING DRAFT · V0.1N-CELL80-BOUND

Can something evolve that makes the next invention possible? ↗

BOUNDIn these fixed worlds, apparent expansion of evolutionary possibility came down to representation, target choice or selection.

Scope and evidence

One substrate and one library of building blocks, tested with exhaustive static maps and pilot worlds, with one evolved module available to test. This is a bound on these instruments, not a theorem about evolution. Encapsulation does strongly change what is reachable; the point is that the change is explained by a module's shape and output range rather than by its having evolved for a use.

FULL EXPERIMENT AND LIMITS ↗
Learned systems4 RECORDS · EXPLORE
WORKING DRAFT · V0.1N-MAP

What is the map actually a map of? ↗

SUPPORTEDMAP-1 / State and reader relabelled together

Scope and evidence

The recorded standalone FFN test restores its output to relative L2 ≈ 3.4 × 10⁻⁷. MAP-2b later identifies missing normalisation in that hand-rolled path. The coordinate calculation holds; whole-model behavioural parity was not established by this test.

FULL EXPERIMENT AND LIMITS ↗
WORKING DRAFT · V0.1N-STATE

What has to survive? ↗

SUPPORTEDFull-sequence transplant / three conditions

Scope and evidence

The recorded all-position transplant produces the donor’s next-token distribution with KL = 0 in three tested conditions at L14/L26 on Gemma 3 4B. An earlier register entry lacked an all-position patch tool; the later transplant records the completed test. The result concerns these full-sequence forward-pass interventions.

FULL EXPERIMENT AND LIMITS ↗
WORKING DRAFT · V0.2N-ADDRESS

Reading by address. ↗

SUPPORTEDCONSTRUCTED MEMORY / SIX READS

Scope and evidence

ffn.json records 6/6 intended answers for this set of planted keys and values. The browser adaptation uses the same seed-0 vectors and is checked against independent NumPy readouts. Neuron suppression is an interactive extension, not a result reported in that saved run.

FULL EXPERIMENT AND LIMITS ↗
WORKING DRAFT · V0.1N-AUTHORITY

Which source wins? ↗

SUPPORTEDPROMOTION / THREE OPERATIONS

Scope and evidence

In the founding run, a 13-token record injected about 49,000 tokens downstream of the source keeps the model correct on all three tested operations — copy 7431, plus one 7432, digit reversal 1347 — with the original span excluded. Padding alone fails on every row (7249, 789, 7892) and a corrupted record steers every row (5824, 5825, 4285). Asked for the code plus one with 5824 injected, the model answers 5825: a value that appears nowhere in the context.

FULL EXPERIMENT AND LIMITS ↗

WHAT REMAINS UNANSWERED

Open questions.

These standing questions sit alongside the next experiments proposed within each programme.

AGENT ECOLOGY / NEXT QUESTIONS

What should the next agent inherit? ↗

Which records should successors preserve? Can they choose between competing versions, and can a defender repair shared state using uncertain diagnostics? What survives over longer horizons?