Don’t write around the gaps.
When a question gets hard, most research agents retreat into a survey. Basis was built to keep going. When a run cannot continue, it reports precisely where and why it stopped.
I The Failure Mode
Hard questions rarely fail in an obvious way. Ask a capable research agent to settle something genuinely difficult and it will usually widen the question until it can answer an easier one. A requested proof comes back as a taxonomy of proof strategies. The mechanism you asked about becomes a review of mechanisms. Twenty fluent pages arrive on schedule, and none of them engage the original question directly.
The trouble is that this failure reads as success: the prose is confident, the citations are real, and the document compiles. Only a reader who already knows the field would notice that the question itself has been written around, and such a reader would not have needed to ask.
II The Stance
Basis learned its anti-goal the hard way. An early course correction, stop writing around proof gaps, put one sentence into the design charter. Every decision since has had to answer to it:
Basis “must not turn depth into ‘write a long paper around gaps.’”
In practice, the sentence sets the standard for completion. A run is finished when the question closes, or when it can state specifically what stands in the way: a lemma that fails, an estimate that will not tighten, or a citation that cannot be verified. A run that cannot close names its blockers rather than padding the manuscript until the gap becomes hard to see.
III In Practice
A Basis run works in cycles. Each cycle proposes, tests, and revises the developing argument, and only supported claims survive into the manuscript. The manuscript is versioned LaTeX, checkpointed every cycle, so you can see exactly when each claim entered and what it displaced.
The process takes time. Expect 6–22 cycles for meaningful progress on a hard question, and expect to read some dead ends. Because the manuscript is checkpointed every cycle, an approach that was tried and later cut remains in the draft history, in the cycle where it fell. We preserve the failed routes deliberately: they show that the surviving claims were contested.
You set the cycle cap and the budget cap in dollars before launch. The budget cap is a hard stop, and a run that reaches it halts. Runs are persistent: you can pause a run, resume it, rewind it to an earlier cycle, or continue it next week, and the record travels with it.
IV What We Won’t Claim
We do not guarantee correctness, novelty, or completeness. Every Basis deliverable is a hypothesis prepared for your judgment: structured so it can be attacked, cited so it can be checked, and versioned so you can see where each claim entered. It requires your review rather than replacing it.
Some runs end in a state we call Completed with blockers, and the run page reports that status plainly. We keep the blunt label because the blocker list is often the most useful thing a run produces: it marks where the question resisted. You will not find success rates or benchmark scores on this page, because we have not published any.
V The Ask
Basis is in public beta. It is STEM-centric, tuned for LaTeX, and unfinished. Some of our own runs end with blockers, and the record shows exactly where. Basis will not hand you twenty confident pages written around your question and present them as an answer.
Run your question and get either a closed answer or a precise account of what stands in the way.
Create accountPublic beta · email verification required
Sign in with email or Microsoft; some organizations require admin approval.