Daniel Kitchen
Researcher.
The Researcher decides what’s worth asking. Every question it raises is ranked by what the answer would actually change — never by how many questions it can produce.
Phrasing a question is easy. Knowing which one is truly worth asking is the entire hard part. Given an objective and everything already known, the engine finds the real gaps, over-generates candidate questions without restraint, throws out every one that fails a hard quality gate, and ranks what survives by how much an answer would genuinely move the decision. It hands those questions off completely — it never answers them itself. That’s the Pipeline’s job, and its job alone.
Context in, ranked questions out.
- Knowledge state — what’s genuinely known versus still open gets modeled in full before a single question is asked.
- Gap detection — finds what’s truly worth asking about: the unanswered, the contradictory, the thin, the missing link, wherever it actually sits.
- Perspective framing — induces genuinely distinct lenses so the questions carry real breadth, never five rewordings of the same one.
- Generation — over-generates structured candidate questions without holding back, one full batch per perspective and gap.
- Quality gates — hard pass/fail filters, applied completely — grounded, answerable, presupposition-clean, single-focus — followed by a full semantic de-dup.
- Value scoring — every survivor is scored on information gain and relevance together, then a genuinely diverse top-k gets selected.
It ranks by value, not volume.
Every surviving question earns exactly two scores: how much the full spread of plausible answers would genuinely shift belief, and how directly that answer serves the real objective. Give it the actual decision you’re making and it damps, completely, any question whose answer wouldn’t change your choice — it refuses to chase something merely interesting when it isn’t actually useful.
Selection is fully diversity-aware: it picks the top few by value while penalizing anything too close to a question already chosen or already asked. What you get is a short, genuinely broad, high-value set — never five paraphrases dressed up as five questions.
A separate model does the judging.
A question is only ever eligible once it clears every single hard gate. The model that writes the questions never grades its own work — the judge is a fully separate call, because self-evaluation is inherently biased and the system refuses to pretend otherwise.
Ask, read, ask sharper.
On its own the engine runs one fully principled pass. Hand it a genuine way to answer — the Pain Point Pipeline — and it iterates completely: it asks, reads the real evidence that comes back, and asks sharper the next time. Together, the two of them are a self-driving research loop in the truest sense.
Built to stand entirely alone. The engine knows nothing whatsoever about the system around it — its entire contract is five small interfaces, with that decoupling proven by tests. It runs inside AJO, but depends on none of it. Code is private; this page is the record.