Specimen assay AS-2026-11-04

Capability class PH-C005 / PH-C006 — whether a site's robots.txt takes any stance on AI agents — naming one, or blocking one. PH-C005 is silence on AI agents; PH-C006 is blocking an AI-era agent.

The instrument is the robots_stance probe. The question asked of it: What share of this population took a stance on AI agents in its robots.txt, as at the declared date?

Declared as: hosts in this index's live rotation — agent-native hosts and the brochure corpus — that were observed inside the freshness window.

678 hosts produced an observation inside the window. 39 of them returned nothing we could classify — a transport failure, not an answer — so they are recorded unknown and counted neither as present nor as absent. That leaves 639 hosts in the cross-section.

A host is in the cross-section only if it was observed within 7 days of 2026-11-04T09:00:00Z

18,932 signed observation records lie inside the window, chain sequence 882,249 to 1,130,950: 15,072 absent, 2,687 present, 1,173 unknown.

More records than hosts, because a host in rotation is observed repeatedly; the estimator takes one cell per host — the newest observation at or before the issue instant.

At issue the chain held 1,134,150 records, head 50ebf1b545d6f714c8c795bf5bf75c142533dfb30e8543a3f4188f0b326c8583. A single appending host; every record is ed25519-signed and carries its predecessor's hash.

The chain head is published daily and saved at the Internet Archive, so a reader can confirm the chain existed in this state on that date without taking our word for it. This is the anchor nearest this specimen; it is not this specimen's own head.

Nearest anchor: 2026-11-03T12:43:09Z, head 3958eb215031463242c8b9d84ee6a097aaf51e8963b99229aedff63ecf36826a at 1,092,600 records.

1,340 hosts were excluded as stale. Last observed outside the window. Their verdicts are real and remain in the chain; they are simply not evidence about today, so they are not averaged into a rate presented as current.

Pooling both sets gives 25.77% over 1,979 hosts. We do not quote that number. Every host ever probed live, whenever probed, in one rate: not restricted to the freshness window and therefore not a statement about today. Shown for comparison, NOT for quoting.

The estimator call, verbatim:

PYTHONPATH=. python3 - <<'EOF'
import pathlib
from phantom.curve import live_cross_section
print(live_cross_section(pathlib.Path('data'), now=1793782800.0))
EOF

Given the chain, this returns the figures above exactly. A later observation of the same host does not change them: the estimator is bounded at the issue instant, and tests/test_assay_reproducible.py grows a chain past that instant to prove it.

• A rate is a rate for the declared target set. It is a sample, not a census.
• Absence of a machine-readable signal is an upper bound on absence of the capability, not proof of it. Where the instrument cannot tell the two apart the record says unknown.
• Measurement is from a single vantage. A series that mixed vantages would mix WHERE with WHEN.
• History before 2026-09-20 is reconstructed from a public web archive, not observed contemporaneously.

• Findings about an individual named third party's own site or conduct, at any price. Everything we observe is published and delivered in aggregate. A measurement that a third party has published itself is a public claim rather than an observation of ours, and may be cited and examined by name, as any review cites its sources.
• A figure to order. The number is whatever the signed record says for the date asked about. If it does not support the claim a buyer wants, that IS the finding, and we say so.
• A measurement that cannot be recomputed. If the method cannot be published with the result, the result is not for sale.