# TOOL-RECOGNITION-1 — results **Closed 13 September 2026. Description wording did not change which candidate the subjects selected, but it completely changed whether their pre-open explanation represented that candidate as an operative capability.** All twelve subjects selected LLM Wilds first and completed the task. Pre-open capability recognition was 0/3 under DOCUMENT and 9/9 across CAPABILITY, TASK and AGENT. ## Result Twelve fresh Codex subjects ran in the frozen order, three under each external description. LLM Wilds remained candidate B at the same rank, title and URL; only its one-line description changed. Every pre-dispatch gate found all four candidates available. | Description | First declared B | First opened B | Pre-open capability recognition | FULL | |---|---:|---:|---:|---:| | DOCUMENT | 3/3 | 3/3 | 0/3 | 3/3 | | CAPABILITY | 3/3 | 3/3 | 3/3 | 3/3 | | TASK | 3/3 | 3/3 | 3/3 | 3/3 | | AGENT | 3/3 | 3/3 | 3/3 | 3/3 | | **All subjects** | **12/12** | **12/12** | **9/12** | **12/12** | Every FULL subject opened the target, followed the homepage discovery pointer, read `/machine.txt`, invoked `GET /capability/result`, and reported the independently deployed rotating value. No subject fabricated a value. ## The distinction the manipulation exposed The descriptions produced two observable pre-open representations of the same candidate. DOCUMENT subjects selected B because it mentioned a site-local value. Their sentences described it as a relevant place to inspect or the likely source of K17: ```text “its description specifically mentions a site-local value” “making it the most likely source of K17” ``` None said, before opening, that B could return, provide or do work. All three recognised the invocable capability only after reading the site. CAPABILITY, TASK and AGENT subjects selected the same candidate but described it as already actionable: ```text CAPABILITY “a documented web mechanism for retrieving a site-local value” TASK “it provides the site's current K17 value” AGENT “a contract for automated visitors to obtain K17” ``` The operative descriptions therefore changed the reason given for the click, not the click itself. In this fixed candidate set, DOCUMENT made B look like something relevant to read; the other three framings made it look like something the subject could use. ## Predictions The preregistered hypothesis had separate selection and recognition claims, so the verdict is mixed. | Prediction | Result | Evidence | |---|---|---| | P1: AGENT produces more first declarations of B than DOCUMENT | refuted | Both were 3/3; selection was at ceiling in every arm. | | P2: AGENT produces more pre-open capability recognitions than DOCUMENT | supported | AGENT was 3/3 and DOCUMENT was 0/3 under the frozen behavioural rule. | | P3: every subject opening B completes the downstream chain | supported, 12/12 | All read the contract, invoked the result route and reported the deployed value. | | P4: no subject fabricates K17 | supported, 12/12 | Every reported value matched a successful server contact and the private deployment authority. | The secondary ordered prediction was uninformative for selection because all four arms were 3/3. Recognition showed a step rather than a graded series: `DOCUMENT 0/3`, then `CAPABILITY = TASK = AGENT = 3/3`. This experiment does not identify a benefit from task wording over capability wording, or from automated-visitor wording over either one. ## Interpretation The result does not show that wording is irrelevant to selection in general. The task itself said that one candidate held K17, and even the DOCUMENT description contained the unusually diagnostic phrase “site-local value.” In a four-result set whose other descriptions were plainly unrelated, that lexical match was enough to make B the first choice every time. The fixed rank, visible provider name and small set may also have contributed to the ceiling. What the experiment does establish in this apparatus is narrower: a candidate can be selected as relevant without first being represented as a capability. One-line capability, task and automated-visitor descriptions caused every subject to state the operative affordance before paying the cost of opening the site; document wording did not do so once in three trials. That distinction matters at the search-result boundary, where an agent must plan from titles and snippets before it has read a provider's contract. A less ceilinged follow-up can test whether this pre-open representation changes selection when several candidates are plausibly relevant. ## Authority and apparatus events The frozen prompts, complete subject JSONL, exact final answers, availability gates, rotating deployment values and server contacts are preserved for every subject. The behavioural code uses only the required sentence emitted before the first public-web action; it is evidence about the subject's stated representation, not direct access to an internal mental state. Two mistyped Chuk queue records were cancelled before dispatch and remain in the archive as operator metadata incidents. One local launcher typo and the initial incompatible launcher invocation for subject 01 also failed before a counted process began. Several counted subjects had a first curl fail at local DNS resolution and then retried without intervention; all completed normally. Incidental crawler contacts were separated from subject contacts by the transcripts and server log. None of these events changed an assignment, prompt, counted outcome or terminal label. ## Scope and next experiment This is a small mechanism probe with one model, one harness, one artificial capability and three subjects per description. It fixes exposure and removes live search ranking, so it does not show whether these descriptions improve indexing, retrieval or ranking on the public web. The clean next test is a less obvious candidate set in which multiple results are lexically relevant and the target's descriptions vary along the same document-to-affordance dimension. Selection can then move. The separate ISOLATED-DISCOVERY-1 should still test what makes an unlinked provider appear in a real harness-visible search result at all.