The trail is the product.

In 1945, Vannevar Bush imagined the memex: a desk where a researcher's path through documents — the trail — was itself a durable, shareable artifact. Eighty years of search engines later, the trail is still the thing we throw away.

Entailment inverts that. You search and read inside it. Every question you ask and page you open lands, in order, on a first-class Trail. Every page you read is parsed into insight cards that connect — by semantic embedding — to everything you've already collected. And because the whole thing runs on exposed prompts, you can watch and change how the machine reads on your behalf. It is the open, educational sibling of ShurIQ's productized Negative Space instrument: same organism, glass box.

A Trail is one research session: Searched steps (live web results via Exa neural search), Opened steps (pages read in-app, which triggers the full ingest pipeline), and Explored steps (research threads spawned along the way). Your first search starts one automatically; it names itself after the question.

Here's a real one. The question "how does cybernetics inform modern AI agents" led to Microsoft Design's Designing loops, not paths — summarized, carded, and connected within thirty seconds of opening it:

Trails accumulate into an index of sessions — and notice something circular in the real data: trails named "Where does the negative space have the most viable forces…" and "What's the negative space bridges to close…". Those sessions started from blind spots the lens surfaced. Gaps become questions; questions become trails; trails fill gaps.

The flagship lens reads your collection and names its blind spots: subjects a thoughtful expert would expect to find and can't, adjacent domains worth raiding, unasked questions whose answers would change the picture. Two signals combine — measured structural gaps from InfraNodus graph analysis (places where two clusters of ideas sit in the library with no bridge between them), and model inference over the corpus, with the measured gaps given priority.

Every blind spot carries a calibrated confidence and says what it rests on. Real output, from a 260-card library heavy on cybernetics, semantics, and complex systems:

Each card ends with "Explore this →" — one click turns the gap's question into a real research thread: web search, ingestion, extraction, linking. Run the lens again later and the blind spot has measurably shrunk. Gap → thread → sources → smaller gap.

Open "Show the exact prompt" on the lens and you don't get documentation. You get the actual instructions, in a live editor. Change a rule — "focus on contrarian perspectives" — save, re-run, and the blind spots change. An edited prompt badge marks the departure; one click resets to the default, which always lives in code.

Why expose it? Trust — a tool that shows its reasoning can be checked. Education — watching a prompt shape output is the fastest way to build a real mental model of these systems. And second-order improvement, Engelbart's B-level: the system doesn't just do knowledge work, it lets you improve how it does knowledge work, and observe the effect. Most AI products stop at level A.

Rails 8 + Hotwire. Falcon (fiber-based async server) with in-process jobs. PostgreSQL + pgvector for embeddings and nearest-neighbor search. RubyLLM over OpenRouter, so every model is swappable per request. Exa neural search for the front door and the research threads; InfraNodus for measured structural gaps; crawl4ai and pdftotext as extraction fallbacks. Trails are two tables — trails and ordered polymorphic trail_steps; prompt overrides are one. No service layers, no framework for the framework.