Laboratory
ExperimentsIn Progress
Each experiment tests one claim about how machines read a page. Every one of them is run against this site before it is pointed anywhere else. Failures stay published, marked as failed.
EXP-001 Running
The Machine View
- Hypothesis
- A page that can show you its own extraction is easier to fix than a page you have to inspect with external tooling. Putting the parser’s view one keystroke away changes what the author notices.
- Method
- A toggle on every page. Semantic regions are outlined and labelled, headings declare their level, marked entities are highlighted, and a dock opens with four readouts: heading outline, structured data, extraction payload and retrieval chunks. All four are derived from the live DOM at the moment the dock opens, so the instrument cannot flatter the page it ships on.
- Status
- Live on every page. Press Shift + M or use the switch in the masthead. Two defects found and fixed on day one, both written up in NOTE-001. Next revision should rank findings by impact rather than listing them evenly.
EXP-002 Running
Readability Probe
- Hypothesis
- Most machine-readability failures are structural and boring (a missing H1, an unparseable schema block, six hundred words with no heading) and can be caught without any model in the loop.
- Method
- Paste a page’s HTML. Sixteen weighted structural checks run against a detached document in your browser: identity, outline integrity, structured data, semantics, content shape and answer shape. Every failing check explains what it costs rather than just flagging it. Nothing is uploaded.
- Status
- Live. Open the probe → The weightings are a first guess and will be wrong in interesting ways; they get revised as EXP-003 produces evidence.
EXP-003 Queued
Citation Drift
- Hypothesis
- The set of sources an answer engine cites for a fixed question is far less stable than the set of pages a search engine ranks for it, and that instability, not the ranking, is the thing worth measuring.
- Method
- A fixed panel of questions, asked of several answer engines on a schedule, recording the cited URLs each time. Published as the raw series rather than a summary: churn rate, half-life of a citation, and how often a source returns after dropping out.
- Status
- Queued. Needs a scheduler and somewhere to keep the series. Expected to say nothing useful before spring 2027. The whole value is in the length of the run.
EXP-004 Queued
The Robots Ledger
- Hypothesis
- Site owners discuss AI crawlers constantly and publish their own logs almost never. The actual arrival pattern is more interesting than the discourse about it.
- Method
- Server-side logging of every non-human request to this site: which crawler, how often, which paths, how much it takes. Published openly as a running ledger, with no filtering of the unflattering parts.
- Status
- Queued, and blocked on something slightly embarrassing: this site currently collects no server logs at all. Requires a host that will hand them over.
Backlog
Not yet experiments
Ideas that have not earned a number. Listed so they are not quietly forgotten, and so the ones that turn out to be bad can be struck through in public.
- Chunk boundary test
- Publish the same article under four heading structures and measure which passages get quoted back.
- llms.txt, honestly
- Does anything actually read it? Serve a unique path from it and watch whether that path is ever requested.
- Schema decay
- How long does a page keep getting rich treatment after its structured data breaks?
- The null site
- A deliberately unreadable twin of this site, published alongside it, to see what the difference is worth.