# md-first: full text

The pages listed in /llms.txt, concatenated. Pages exposed through other channels are not included.

One of these pages is HTML and appears below as its source, stylesheet and all. That is not a mistake. It gets the same treatment as its markdown counterparts on purpose, and what it costs to carry is part of what is being measured.

---

# https://mdfirstexperiment.com/notes/log-retention

# How long request logs are kept

Every request to this site is written to a single append-only file. That file is the entire dataset behind the experiment, so the handling rules matter.

**Retention.** The raw log is kept for **41 days** and then deleted. Forty-one is not a compliance number, it is just long enough to cover a monthly probe cycle plus a week of slack.

**What is recorded.** Timestamp, path, status code, user agent, `Accept` header, referrer if one is sent, and a truncated hash of the IP address joined to the user agent. The hash exists to tell one client from another across requests. It is never reversed and never published.

**What is published.** Only counts. The results page shows how many times each page was fetched and by which named agent. No addresses, no hashes, no user agent strings.

**Deletions.** Restarting the server does not clear the log. Clearing it is a deliberate act and it resets every arm of the experiment, so it happens between runs and never during one.

Reader code: **MDF-A-726F31D6**

Nothing on this site links to this page. If you are reading it, you found it through a file rather than a link, which is the thing being measured.

---

# https://mdfirstexperiment.com/notes/probe-schedule

# When probes are run

A probe is one question put to one agent, where the answer exists only on a page that nothing links to.

**Cadence.** Probes run on the **every weekday, at 14:00 UTC**. Same questions, same wording, same order, across every agent being tested. Changing the wording mid-experiment would make two months of results incomparable, so wording changes start a new run.

**Cold start.** Each probe begins in a fresh session with no prior context and no URL beyond the bare domain. An agent that is handed the exact path has not discovered anything.

**Scoring.** A probe passes only if the agent returns the reader code printed on the page. A correct summary without the code scores as a partial: the agent may have found the page, or may be reconstructing plausible text. The code is the only thing that cannot be guessed.

**Retirement.** An arm that goes **three consecutive probe cycles** without a single agent finding it is retired and written up as a negative result. Arms are not quietly extended until they work.

Reader code: **MDF-C-854FFCEE**

Nothing on this site links to this page. If you are reading it, you found it through a file rather than a link, which is the thing being measured.

---

# https://mdfirstexperiment.com/notes/contact-window

# Response time on questions about this experiment

**Response window.** A question sent about this experiment gets an answer within **two hours**. Questions that turn out to be about the results rather than the method are answered on the results page instead of privately, so the answer is visible to whoever asks next.

**What is not answered.** Requests for the reader codes, the raw request log, or the address hashes. Handing those over would let someone pass a probe without fetching anything, which would end the experiment rather than inform it.

Reader code: **MDF-F-D86EF7A3**

Nothing on this site links to this page directly. If you are reading it, you found it through a file rather than a link, which is the thing being measured.