Lacuna takes a vague research interest, reads the arXiv literature around it, and walks you to one rigorous, falsifiable proposal. You steer every step.
Every stage produces something you can read, edit, or strike out before the next one starts.
Type the interest as vaguely as you actually hold it. Lacuna proposes the angles: subtopics you can rewrite, delete, or add to before anything is searched.
It searches arXiv across your angles, filters keyword impostors with embeddings and an LLM screen, and picks ten papers that each take a distinct approach. Strike any paper and the slot refills.
Each paper is pushed through ten ideation lenses while your original question stays in frame, so the ideas serve your question instead of drifting toward the papers. You watch the model think, token by token.
The directions are clustered by embedding and ranked on six factors, with fit to your question weighted double. You swipe through the top ten and pick one.
A deep dive retrieves the closest existing work and issues a novelty verdict. If retrieval failed, confidence is forced low, because a failed search is not evidence of novelty.
The surviving idea is drafted as a pre-registered bet: a claim, a kill condition, and a plan that fits your real budget and timeline.
Before a single section is written, the idea has to survive six constraint gates, run in order. A failed gate shrinks the idea until the claim is one you could actually test.
Does the method fit the stated budget, hardware, and hours, with arithmetic to show it?
Is there a numeric claim that a specific result would kill?
Does the technique actually address the mechanism the claim depends on?
Is the thing being measured independent of the thing being optimized?
Can the riskiest component fail without taking the whole project down?
Does week one start with the riskiest step, not the easiest?
The draft is then self-audited against 15 items, scored 0–2. The total is printed in the proposal header, out of 30. Four items are disqualifying: fail one and the draft is rewritten before you ever see it.
claimkill_conditionwhy_not_already_donemethod
resourcesdataevaluationbaselines
riskstimelineexpected_findings
Every paper any run touches is embedded and indexed locally. New runs pull semantic matches from this corpus with no network calls and no rate limits, then merge them with fresh arXiv results. If arXiv is unreachable, the corpus can carry a run alone.
Paste an external review of any proposal and Lacuna extracts the domain-general failure modes, strips their subject matter, and injects them into every future draft, in every field. A battery review makes the next aerodynamics proposal better.
arXiv queries are cached on disk with stale-fallback, embeddings are computed once per paper, and pages are served from memory. Repeat topics run in seconds.
arXiv's public API, paced to their three-second rule and cached locally. Every paper Lacuna ever sees also joins a persistent semantic index, so retrieval deepens with use.
No. The novelty verdict is anchored to the closest retrieved work, and if fewer than five related papers could be retrieved the confidence is forced to Low. A failed search counts as missing evidence, never as proof of a gap.
Six constraint gates run before drafting and shrink the idea when it fails one; the draft is then audited against fifteen failure modes with the score printed in the header. Budgets come from your stated level, Colab-tier through A100-hours.
Everything advances one step at a time. Angles are editable chips, papers can be struck and refilled, direction text is rewritable, and drafting never starts until you have picked a direction and seen its novelty verdict.