v0.2.1

Check now prints a next block: measured follow-ups from this run, not generic advice.

Cost-derived next lines require the HTML extracted total to be at or above 1109 tokens, the 300-host corpus p25 in docs/results.md. That floor is a published percentile, not a grade. Missing Vary: Accept, dropped body blocks, and structural findings are not cost-derived and are not subject to it.

Code samples (pre/code) are no longer counted as lost body content. The 3/7 markdown body-loss figure was classified again from the stored 300-host corpus after that rule. It did not change. The code rule dropped zero blocks on those seven pages.

Changelog

• b1680c7 Print measured next steps on check and stop counting code as lost body.

v0.2.0

Token cost is now extracted text under cl100k_base. Raw response size is a separate bandwidth figure.

The first cost figures counted raw HTML (script, style, and attributes) and are withdrawn. They are replaced by median extracted tokens, the share that is not body, the chrome share, and the markdown-variant median.

The encoding is cl100k_base. A count is reproducible with tiktoken.get_encoding("cl100k_base").

Changelog

• 7665fd4 Measure token cost on extracted text and withdraw the raw-HTML figures.

v0.1.2

The injection scan now drops a hidden A/B banner variant when its near-duplicate is visible in the same region. That pattern is experiment chrome, the same class as unused tooltips and rails.

The 300-host scan is 1/281. That is the published figure. firecrawl.dev is the one structural finding.

v0.1.1 printed 2/281. The second hit was an unused free-trial offer variant on mailchimp.com. Do not quote 2/281.

Changelog

• ade5f20 Drop hidden A/B variants whose near-duplicate is visible nearby.

v0.1.1

Correction

This release's scan printed 2/281. One of those hits was an unused A/B free-trial banner variant, not planted machine-only text. v0.1.2 excludes that pattern. The current scan is 1/281. Do not quote 2/281 from this release.

Changelog

• b1ff770 Drop badges that do not resolve and pin CI to test.yml on main.
• c39c8d5 Measure uncertain pages and report markdown body loss separately from missing Vary.
• 3317bfb Publish the landing-page README and an MIT license.

v0.1.0

Changelog

• e73ce91 Start public history from the cleaned tree so earlier audit documents are not recoverable.