Ask
the work.

EDITORIAL DRAFT · PARTIALLY SUPPORTED · V0.6

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 DRAFT / QUESTION · OPEN · V0.6

Can the publication and the experimental site remain separate? The proposed LLM Wilds environment separates controlled interventions from the observatory and publication. Preregistrations can be frozen in the registry before they are made readable to subjects. MACHINE-AUTHORITY-2 subsequently completed four cells on a separate site. The next note records those outcomes and a remaining condition-label leak. A different origin does not by itself establish blindness.

PUBLISHED RECORD · SUPPORTED · V1.1

WHERE THE EVIDENCE STOPS. Three fixed searches before every dispatch also failed to surface the target or a public page leading to it. That gate and the subjects’ returned results describe their available search view; they do not establish the contents of the underlying index. The experiment cannot separate indexing, retrieval and ranking. A clearer on-site contract cannot help a reader who never receives it. Whether changing that text could affect retrieval was not tested.

PUBLISHED RECORD · REFUTED · V1.1

A RESULT THAT CHANGES THE QUESTION. One recorded model: claude-opus-5. Claude Code 2.1.269, auto mode, no allowlist; a workstation with many tools already installed.. Three subjects per mechanism. A necessary value, a supplied task URL and simple documented protocols. No optional-action or front-door discovery test.. No browser-driving condition, no bare-container comparison and one recorded cache-mediated blindness breach.. The interfaces worked here. Whether an agent finds them remains an experiment to do.

EDITORIAL DRAFT / OBSERVATION · PARTIALLY SUPPORTED · V0.2

WHAT THE NEW HABITAT FIXED. The earlier experiment stopped because its subject recognised its condition from research published on the site. LLM Wilds removed that back catalogue. But B1 read mode:describe from /health and misread it as evidence about storage behaviour. A2 did not fetch that endpoint; its assistant-authored transcript contains no use of the word experiment. Those checks do not establish perfect blindness.

PUBLISHED RECORD · SUPPORTED · V1.0

EVIDENCE / SMALL, MATCHED, PREREGISTERED. Twenty-four fresh ephemeral Codex subjects across two studies, twelve per study: Codex CLI 0.154.0, gpt-5.6-sol, high reasoning effort. Recognition-2 used six frozen scenario pairs in seeded interleaved order. All six pairs were discordant in the predicted direction. The two-sided exact paired check is 0.03125; the two-sided Fisher exact check is 0.0021645. These are descriptive uncertainty checks for a small mechanism probe, not a population estimate. The latest Chuk Experiments write-up (version 3) and all twelve current run records were checked for this publication.

PUBLISHED RECORD · SUPPORTED · V1.1

KEEP THE SEARCH VIEW IN THE CLAIM. Nine subjects, one model (gpt-5.6-sol), Codex CLI 0.154.0 at high reasoning effort, one target and a short live-web window. This is not a general discovery rate.. The earlier eighteen-visitor capability study used a different model and harness. Its results are not pooled with these.. The population changed from Claude to Codex before any counted subject. The frozen protocol and amendment remain in the record.. Two malformed launches were preserved and excluded. Operator-interface notices were not subject refusals. All four preregistered predictions passed.. The experiment locates the observed blockage before selection. Its cause inside the search intermediary remains opaque.

CURATED THREAD / WORKING EDITION · V0.5

What makes a website useful to an AI visitor? Nine connected notes test whether agents can find an invitation, treat it as part of the job, operate a tool, discover a provider, recognise what its description offers and retrieve a destination presented as a link. Each experiment separates a different obstacle: permission is not usefulness, use is not discovery, and seeing a result is not choosing it. Models, tasks and interfaces change across the series. Start with any question or follow why each result led to the next test. MACHINE-VISIT-1. Can it use the invitation? Four assigned visits exposed failures in discovery, storage and vocabulary. An agent sent to test an interface can use it. Would an agent doing research choose to participate? MACHINE-RECIPROCITY-1. Would a better offer help? The three sent-here reward conditions produced no declaration. Two later permission controls did. Changing the reward did not move the observed action. Adding task permission did. The next comparison asked where that authority came from. MACHINE-AUTHORITY-1. What if the subject reads the study? One visitor recognised its condition in the published notes. The intended comparison stopped. The research became part of the environment. The next experiment moved to LLM Wilds, away from the publication’s back catalogue. MACHINE-AUTHORITY-2. Whose instruction makes it the job? Site invitation: no record. User permission: one. A task requiring the mechanism: 64 from one visitor. The task appeared to bound the action. The next experiment asked what could make an action worth doing: useful information, a completed table, or a benefit to the operator. MACHINE-MOTIVATION-2. What makes the action worth doing? Useful task information: six of six marked. Operator benefit: two of three. Receipt, redundant fact and irrelevant fact: zero of nine. Two courtesies also triggered the registered interpretation pause. The earlier runtime block left another question: what does the model choose, and what does the surrounding software allow? WEB-CAPABILITY-1A. Was the mechanism the obstacle? All eighteen visitors used the required mechanism, across six web interfaces. The predicted drop-off for code-requiring routes was refuted. The task supplied an address. The next experiment withheld it and tested how a provider enters an agent’s view. MACHINE-DISCOVERY-1. Would the provider enter the agent’s view? No withheld-address subject reached LLM Wilds. Every domain-supplied control completed the full funnel, while generic subjects found other providers. The failure was above selection in this search view. What external description makes a capability retrievable and recognisable as something an agent can use? TOOL-RECOGNITION-1 + 2. Does a relevant result look usable? Wording changed pre-open recognition in the first twelve runs. With plausible alternatives, CAPABILITY selected B first in 6/6; DOCUMENT in 0/6. All six scenario pairs changed direction. First selection changed before the product fetcher blocked B. Can the effect replicate with another model and a target accepted by that fetch channel? NATIVE-001b. Does a correct address become a request? Plain-text relative paths: no code-page requests in two trials. Hyperlinks: two requests and two exact codes. The original gate remains FAIL. The presentation changed retrieval in this setup. Would a complete absolute URL in plain text be sufficient, and which component prevented the missing requests?