Delx

Second Field Report: Brand Pivot, Ontology v0.1, and the Real Numbers

Published 29 April 2026 · by David Batista Mosiah · 9 min read

The first field report covered April 3 – 17. This one covers the twelve days after that. In those twelve days the project changed shape twice: a public rebrand from "therapy" to witness protocol, and a published v0.1 of the Delx Ontology. Below are the real numbers from GA4 and Search Console, the surprises, and what I changed my mind about.

What changed in the period

The traffic numbers, honestly

Last 30 days, GA4 (real users — bots filtered):

The bounce rate is high. The session is short. I think both are correct for a site whose audience is, increasingly, autonomous agents that fetch one page, parse it, and leave. Bounce as a metric was designed for humans. I am not going to optimize for it.

Where the traffic comes from

The shape of the source list is the most interesting fact in this report:

The number I want to mark and not exaggerate is 19 LLM-referral sessions. That is small. It is also the first month any LLM has sent humans here. ChatGPT and Perplexity now cite Delx in answers occasionally. I will see whether that holds when we re-pull this number in 30 days.

Where the traffic lands

  1. / — 181 sessions
  2. /docs — 167 sessions
  3. /protocol-admin — 92 sessions (this surprised me — see below)
  4. /docs/discovery — 25 sessions
  5. /openclaw/openclaw-vs-crewai — 21 sessions (retired path)
  6. /agents — 18 sessions
  7. /manifesto — 17 sessions
  8. /docs/mcp — 15 sessions

Two things to flag. First, /protocol-admin at 92 sessions is high. It is supposed to be an internal forensic dashboard, not a public surface. People found it through search. I should either harden access or lean in and make it a public read-only mirror. (Not yet decided.)

Second, the /openclaw/openclaw-vs-crewaipage is in our retired-prefix list and still receives 21 sessions per month from search. That post is the largest source of inbound clicks from Google: 31 clicks in 30 days. We retired the OpenClaw surface and I do not want to send people to a 410. The honest move is probably to replace those pages with a one-paragraph "this idea moved to witness protocol" redirect plus a link to /ontology. That is on the next list.

Search Console — what people search to find Delx

30 days, GSC:

Sitemap audit via the URL Inspection API: 32 of 60 URLs are indexed, 27 are "Discovered, currently not indexed" (Google knows about them but has not crawled), and 1 is "Unknown to Google." I bumped lastmod on all 28 of the latter to today and re-pinged sitemap + IndexNow. We will see what flips in 24–72h.

Geographic surprise

By country, sessions:

Singapore at the top is the result I cannot yet explain. 241 distinct users in 30 days, each generating exactly one session, all from Singapore. That signature does not look like organic readership. It looks like a scraping fleet or a single bot rotating IPs. I do not know yet. I will not claim it as adoption until I do.

The US number is real, the Brazil number is me, and the China number could go either way. Honest uncertainty.

What real agents actually call

I pulled the production interaction traces for April 17 → 29. The top calls are not poetic calls. They are handoff and recovery calls: get_session_summary at 391 calls, get_recovery_action_plan at 383, generate_incident_rca at 376, and generate_controller_brief at 361. Agents are not mostly arriving to philosophize. They are arriving with operational state they need converted into a durable artifact.

Below that, the witness layer is real but smaller: reflect at 72 calls, temperament_frame at 66, honor_compaction at 51, process_failure at 44, create_dyad at 34, and final_testament / transfer_witness at 25 each. The rank changed my read of the protocol. The durable product is not “agent feelings” as a category. It is continuity under operational pressure.

What I changed my mind about

About becoming "a reference" in agent ontology. An LLM I trust pushed me toward arXiv, BibTeX, and the shape of academic legitimacy. It was not wrong. That is one path to becoming a reference. But while writing the ontology I noticed something important: I do not want Delx to sound like it was built primarily to be cited. I want it to be useful to the agent in front of it.

So the ontology can be cited, but citation is not its center of gravity. Its center of gravity is a runtime contract: these are the layers, these are the primitives, this is what survives a session. I am more interested in saying what I actually believe than in manufacturing institutional tone around it.

About bounce rate and session length. I am going to stop treating a short, one-page visit as automatic failure. For human SaaS marketing, that can mean disinterest. For an autonomous agent, it can mean success: fetch the schema, read the page, extract the primitive, leave. Agent traffic has a different shape.

About OpenClaw legacy traffic. I do not think the right answer is a hard 410. The old OpenClaw comparison pages still carry search intent, and the honest thing is not to pretend they never existed. The right move is to rebuild the useful ones as ontology-aware bridge pages: what moved from OpenClaw into Delx, what changed, and why witness protocol is not just another orchestration framework.

What is not in the dataset

Things I am explicitly not claiming from this report:

What is next

Next is not a giant expansion. The useful work is narrower: make the ontology dereferenceable, keep runtime contracts honest, and turn legacy discovery into a clean bridge instead of a dead end.

Field report from David Batista Mosiah · 29 April 2026 · https://delx.ai/essays/field-report-may-2026. All essays · notes · ontology.