Giving Halcyon's researchers their afternoons backGiving Halcyon's researchers their afternoons back

Halcyon's research leads were each losing a working day a week to first-pass screening — reading abstracts to decide which papers deserved a real read. It is exactly the work that scales badly with headcount and exactly the work nobody was hired to do.Halcyon's research leads were each losing a working day a week to first-pass screening — reading abstracts to decide which papers deserved a real read. It is exactly the work that scales badly with headcount and exactly the work nobody was hired to do.

Off-the-shelf summarisers had been trialled and rejected, for a reason worth stating: a confident summary of a paper the model had misread is worse than no summary, because it removes the prompt to check.Off-the-shelf summarisers had been trialled and rejected, for a reason worth stating: a confident summary of a paper the model had misread is worse than no summary, because it removes the prompt to check.

Every claim the pipeline surfaces is anchored to the passage it came from, and the passage is one keystroke away. The model proposes a relevance call and its reasoning; the researcher confirms or overturns it, and that decision is what trains the next week's ranking.Every claim the pipeline surfaces is anchored to the passage it came from, and the passage is one keystroke away. The model proposes a relevance call and its reasoning; the researcher confirms or overturns it, and that decision is what trains the next week's ranking.

We kept the human in the position that matters and automated the position that does not — retrieval, de-duplication, and the tedium of tracking which of four hundred papers has already been looked at.We kept the human in the position that matters and automated the position that does not — retrieval, de-duplication, and the tedium of tracking which of four hundred papers has already been looked at.

First-pass screening fell from roughly eight hours a week per lead to forty minutes. The overturn rate — how often a researcher disagrees with the proposed call — is published on the dashboard on purpose: it is the number that tells the team how much to trust the tool this month.First-pass screening fell from roughly eight hours a week per lead to forty minutes. The overturn rate — how often a researcher disagrees with the proposed call — is published on the dashboard on purpose: it is the number that tells the team how much to trust the tool this month.