Web accessibility

A dated record of what actually happens when Dinoer’s own discovery step runs for real: what it finds, what it misses, and on what grounds. Counted from a real campaign, not estimated — published so a future run can compare rather than guess, and so anyone can contradict it with their own measurement.

2026-08-14 — Reputation-research campaign, SearXNG discovery vs. one unprepared search

Context: a real reputation-research subject (self-consenting — no third-party privacy question), three SearXNG-driven query targets, campagne.py, corpus locked by opencode.jsonc. 28 pages retained at the light tier; 0 of 1 escalation to the heavy tier succeeded.

The comparison that mattered: a single, unprepared query to a generalist, authenticated search assistant on the same subject surfaced three pages — a professional-network profile, a project page, a stock-photo credit — absent from all 28 collected pages, verified directly against collecte.jsonl, not assumed from the automatic report.

Root cause, identified rather than guessed: the three SearXNG queries used were aimed at company directories — the wrong kind of search for pages of that kind, a discovery-formulation gap, not a ranking or truncation defect downstream.

The second, independent finding in the same corpus

Two of the 28 pages — a code-hosting profile and a video-platform channel — were fetched successfully by the light tier, but their collected text was cookie-banner and login-wall noise only. The light tier cannot execute their JavaScript, and heavy-tier escalation never triggered for either page: the escalation heuristic keys off “empty text,” and both pages technically returned some.

What this is, precisely: a distinct architecture gap from the discovery gap above — reformulating the SearXNG queries would not have fixed it, because both pages were already found and fetched. Both findings, and what they changed about the project’s own claim →

What was verified, not just reported

The synthesis step for this same campaign was checked independently before either finding above was written down: a direct call to the delegated model’s full JSON event stream — not the summarising helper that discards tool-use events — confirmed exactly three events for the whole run (step_start, text, step_finish) and zero tool_use events. Nothing left the corpus while the report was written, on the same run that showed the corpus itself had real gaps. Two separate questions, both answered by measurement rather than trust.