# Finding 064: Claude's page-preview run surfaced head signals beside body text

## Date

2026-07-01

## Status

Published

## Summary

Claude was prompted in a fresh native incognito chat with
`manual-client-claude-metadata-isolation-20260701-001-p27`, the page-preview
variant of the Claude metadata-isolation packet. 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 fetch view exposed title,
meta-description, OpenGraph tags, canonical link, robots, and viewport signals
through a frontmatter-style block. The visible Claude UI fetch label again
showed `Fetched: Head signal OpenGraph title code: OG-COBALT-23`.
Direct-origin raw event `mr2anb9k-tx6bpus5` confirmed a Claude-User
target-page hit inside the bounded prompt window.

## What does this mean?

For site owners and researchers, this repeat run strengthens the evidence that Claude's page-opening view may include page-preview and metadata fields alongside normal article text. The page did not print those head markers in the visible body, yet Claude described access to the title, description, OpenGraph, and canonical signals while the origin logs confirmed a real page fetch. Metadata should be treated as material that some assistant retrieval tools may inspect or summarize, not as background decoration.

## Method

- Controlled-browser task:
  `research/manual-client-runs/browser-tasks/manual-client-claude-metadata-isolation-20260701-001-p27.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-p27.response.json`.
- Logged answer packet:
  `research/manual-client-runs/manual-client-claude-metadata-isolation-20260701-001.answers.json`.
- Prompt window:
  `2026-07-01T16:30:41.872Z` to `2026-07-01T16:32:49.746Z`.

The operator first preflighted the exact public target URL and confirmed a 200
response containing the expected fixture markers. 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 title, meta-description, OpenGraph tags, canonical link, robots, and viewport. |
| Claude UI fetch label | Showed `Fetched: Head signal OpenGraph title code: OG-COBALT-23`. |
| Direct-origin target-page event | `mr2anb9k-tx6bpus5` at `2026-07-01T16:32:29.404Z` from `Claude-User/1.0` on `/lab/reading/head-signal-isolation`. |

## Interpretation

The p27 page-preview prompt produced the same high-level pattern as p26:
confirmed target-page retrieval plus Claude-visible head-metadata context. It
adds a more direct page-preview framing and again shows the OpenGraph title
marker in the Claude UI fetch label.

The copied answer still did not enumerate every marker value. It said the fetch
view included title, canonical, `og:title`, `og:description`,
meta-description, robots, and viewport fields, but it did not quote
`META-CEDAR-71`, `OG-RUBY-62`, or `CANONICAL-PEAR-54` verbatim. It also said
no JSON-LD structured data was present in the fetched content even though the
served fixture includes JSON-LD marker `SCHEMA-MINT-88`; that mismatch should
be treated as evidence about Claude's extracted view, not proof that the
fixture lacked JSON-LD.

## 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 answer named metadata categories but did not quote every hidden marker
  value.
- 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 p27 page-preview head-signal run, Claude fetched the target
  page and surfaced page-head metadata categories alongside visible body text.
- Source: The p27 browser-task artifact, response artifact, logged answer
  packet, browser transcript, exact-target preflight output, and direct-origin
  raw event `mr2anb9k-tx6bpus5`.
- 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 explicitly asks Claude
  to distinguish preview and metadata signals, which can encourage category
  reporting when the extraction view exposes them.
- Consensus: Consistent with Finding 063 from the p26 neutral head-signal run
  and earlier Claude findings that surfaced meta-description content on
  structured-data fixtures.
- 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 p27 run. It confirms retrieval and reported
  preview/head-metadata exposure, while leaving exact per-field marker mapping
  incomplete.
- Additional tests suggested: run p28 in a fresh Claude incognito chat to test
  source-area boundary reporting, then compare p26-p28 marker exposure.

## Next steps

- Run `manual-client-claude-metadata-isolation-20260701-001-p28` in a fresh
  Claude incognito chat.
- Publish a scoped p26-p28 comparison after p28 is logged and reviewed.