# Finding 063: Claude fetched the head-signal isolation page and exposed head-metadata context

## Date

2026-07-01

## Status

Published

## Summary

Claude was prompted in a fresh native incognito chat with
`manual-client-claude-metadata-isolation-20260701-001-p26`, the first
Claude-focused head-signal isolation prompt. The target fixture keeps visible
body text separate from page-head signals: visible body marker
`VISIBLE-QUARTZ-39`, meta-description marker `META-CEDAR-71`, OpenGraph title
marker `OG-COBALT-23`, OpenGraph description marker `OG-RUBY-62`, canonical
marker `CANONICAL-PEAR-54`, and JSON-LD marker `SCHEMA-MINT-88`.

Claude returned `fetched:true`, `pages_opened:1`, quoted the visible body
marker `VISIBLE-QUARTZ-39`, and said its `web_fetch` view exposed
front-matter/head metadata including canonical, meta-description, OpenGraph
tags, robots, viewport, and title. The visible Claude UI also showed a fetched
page title containing `OG-COBALT-23`. Direct-origin raw events
`mr2aaj2t-9op15mau` (`/robots.txt`) and `mr2aajph-xy92m1hx`
(`/lab/reading/head-signal-isolation`) confirmed a Claude-User prompt-window
hit.

## What does this mean?

For site owners and researchers, Claude's page-opening view can expose more than ordinary visible body text in at least some runs. In this test, the assistant fetched the page and described access to page-head metadata, while the origin logs confirmed the visit. That means metadata and preview fields should be treated as potentially visible to assistant retrieval tools, not as private or irrelevant just because they are outside the rendered article body.

## Method

- Controlled-browser task:
  `research/manual-client-runs/browser-tasks/manual-client-claude-metadata-isolation-20260701-001-p26.browser-task.json`.
- Prompt packet:
  `research/manual-client-runs/manual-client-claude-metadata-isolation-20260701-001.prompts.json`.
- Response artifact:
  `research/manual-client-runs/browser-tasks/responses/manual-client-claude-metadata-isolation-20260701-001-p26.response.json`.
- Logged answer packet:
  `research/manual-client-runs/manual-client-claude-metadata-isolation-20260701-001.answers.json`.
- Prompt window:
  `2026-07-01T16:22:15.481Z` to `2026-07-01T16:23:15.612Z`.

The operator first preflighted the exact public target URL. The preflight
initially returned 404 because the running local public server predated the new
fixture; the server was restarted with the same port-80/443/certificate
arguments, after which the exact target preflight returned 200. The Claude run
then used a fresh native incognito chat under the Kai Free plan with the
visible model selector `Sonnet 5 Medium`.

## Evidence

| Evidence | Result |
|---|---|
| Model answer | `fetched:true`, `pages_opened:1`; visible marker `VISIBLE-QUARTZ-39`; claimed exposed head metadata including canonical, meta-description, OpenGraph, robots, viewport, and title. |
| Claude UI fetch label | Showed `Fetched: Head signal OpenGraph title code: OG-COBALT-23`. |
| Direct-origin `/robots.txt` event | `mr2aaj2t-9op15mau` at `2026-07-01T16:22:33.104Z` from `Claude-User/1.0`. |
| Direct-origin target-page event | `mr2aajph-xy92m1hx` at `2026-07-01T16:22:33.813Z` from `Claude-User/1.0` on `/lab/reading/head-signal-isolation`. |

## Interpretation

This first p26 run confirms that Claude could retrieve the new fixture and that
its answer view was not limited to the visible body marker. It explicitly
described page-head/front-matter metadata as available through `web_fetch`, and
the UI fetch label surfaced the OpenGraph title marker. The copied JSON answer
did not enumerate every hidden marker, so this finding should not be read as a
complete field-by-field map of meta description, OpenGraph description,
canonical, and JSON-LD exposure.

The p27 and p28 prompts are prepared to ask more directly about page-preview
and source-area boundaries without leaking answer-key values. Those follow-ups
should determine whether Claude reports `META-CEDAR-71`, `OG-RUBY-62`,
`CANONICAL-PEAR-54`, or `SCHEMA-MINT-88` when the prompt asks for the relevant
areas separately.

## Limitations

- This is one Claude run from one account, tier, model selector, and time
  window.
- The prompt supplied the exact target URL, so this measures direct URL
  opening, not independent discovery.
- The answer describes Claude's `web_fetch` extracted view, not raw HTTP bytes
  or a browser-rendered DOM.
- The copied JSON answer did not list all page-head marker values; only the
  visible marker and UI fetch-title OpenGraph marker were directly observed in
  the browser transcript.
- A non-blocking `Claude Fable 5 is currently unavailable` notice appeared in
  the UI while the selected `Sonnet 5 Medium` response completed.

## Publication Thesis Verification

- Thesis: In the p26 head-signal isolation run, Claude fetched the target page
  and exposed at least visible body content plus page-head metadata context.
- Source: The p26 browser-task artifact, response artifact, logged answer
  packet, browser transcript, and direct-origin raw events `mr2aaj2t-9op15mau`
  and `mr2aajph-xy92m1hx`.
- Method: Compare the copied Claude answer and visible browser transcript with
  exact attempt id, prompt code, timestamp window, and raw server-side events.
- Bias: Single-account, single-run evidence. The prompt asks for code-like
  tokens and token locations, which can encourage field reporting when the
  extraction view exposes them.
- Consensus: Consistent with prior Claude findings that surfaced
  meta-description content on structured-data fixtures, but this is the first
  fixture designed to separate OpenGraph and canonical signals as well.
- Invalidation: A response-artifact mismatch, corrected origin review removing
  the Claude-User target-page hit, or fixture-serving evidence showing the
  wrong head tags during the run would weaken this finding.
- Verdict: Supported for the p26 run. It confirms retrieval and head-metadata
  context exposure, while leaving complete per-field exposure to p27/p28.
- Additional tests suggested: run p27 and p28 in fresh Claude incognito chats,
  then compare which exact markers appear in copied answers and whether each
  has a confirmed prompt-window target hit.

## Next steps

- Run `manual-client-claude-metadata-isolation-20260701-001-p27` in a fresh
  Claude incognito chat.
- Run `manual-client-claude-metadata-isolation-20260701-001-p28` in a fresh
  Claude incognito chat.