Researcher.
It decides what's worth asking. Given an objective, it generates the research questions a pipeline should pursue — ranked by what their answers would change.
Phrasing a question is easy. Knowing which one is worth asking is the hard part. Given an objective and what's already known, the engine finds the gaps, over-generates candidate questions, throws out the ones that fail a hard quality gate, and ranks what's left by how much an answer would move the decision. It hands those questions off — it doesn't answer them. That's the Pipeline's job.
Context in, ranked questions out.
- Knowledge state — model what's known versus still open before asking anything.
- Gap detection — find what's worth asking about — the unanswered, the contradictory, the thin, the missing link.
- Perspective framing — induce distinct lenses so the questions have breadth, not five rewordings of one.
- Generation — over-generate structured candidate questions, one batch per perspective and gap.
- Quality gates — hard pass/fail filters — grounded, answerable, presupposition-clean, single-focus — then semantic de-dup.
- Value scoring — score survivors by information gain and relevance, then select a diverse top-k.
It ranks by value, not volume.
Every surviving question gets two scores: how much the spread of plausible answers would shift belief, and how directly that answer serves the objective. Give it the decision you're actually making and it damps questions whose answers wouldn't change your choice — it won't chase something interesting but useless.
Selection is diversity-aware: it picks the top few by value while penalizing anything too close to a question already chosen or already asked. You get a short, broad, high-value set — not five paraphrases of the same question.
A separate model does the judging.
A question is only eligible if it passes every hard gate. The model that writes the questions never grades its own — the judge is a separate call, because self-evaluation is biased.
Ask, read, ask sharper.
On its own the engine does one principled pass. Hand it a way to answer — the Pain Point Pipeline — and it iterates: it asks, reads the evidence that comes back, and asks sharper. The two together are a self-driving research loop.
Built to stand alone. The engine knows nothing about the system around it — its entire contract is five small interfaces, decoupling proven by tests. It runs inside AJO, but depends on none of it. Code is private; this page is the record.