What does a healthy PM-to-research relationship look like when research timelines are open-ended?

Ashfen Systems builds Quillrace, an AI tool that helps researchers search and discover the scientific papers relevant to their work. Linus Marrast is the AI PM who owns Quillrace's search and ranking roadmap. Ludovica Panzeri leads the three person applied research team exploring how Quillrace ranks papers that are too new to have earned real citations yet.

How to answer this, stage by stage

Nobody is grading whether you can say "communicate early and often." They are grading whether you can name the actual shape of a cadence that survives contact with a team whose work has no ship date yet.

Let's learn

This is the gap the cadence exists to close. Not a lack of talent in the room. A lack of a road for what the room already knows.

Quillrace reads a researcher's search and ranks the papers that match it. Before Ashfen touched the ranking model, a materials scientist searching for recent work in her own niche field had to scroll past pages of older, heavily cited papers, then cross check preprint servers by hand to see if anything newer had come out. That took her close to forty minutes a week, done on top of her actual research.

Quillrace's ranking leaned hard on citation graphs, which papers get cited alongside which others. That worked well for established work. For a cold start paper, one that's actually the best match for a search, it still landed outside the top twenty results sixty eight percent of the time, measured across one quarter of real search logs.

Neither extreme is a person failing at their job. Both are what a missing structure hands people instead.

Ashfen's research group started at one person, Panzeri alone, doing work close enough to shipping that a weekly "what can we demo Friday" check in fit fine. Once two more researchers joined and the work turned into open exploration, that same weekly ask started to cost something. Two of the three researchers began saving their update slot for whatever looked closest to done, and quietly leaving out the runs that hadn't gone anywhere yet.

Bavo Trevisan, one of the researchers, said it plainly in a meeting: they'd started deciding what to tell Marrast based on what sounded finished, not on what was actually true that week. Marrast heard it, and swung hard the other way. He cancelled the standing meeting and told the team to work however they needed to and ping him when something was ready.

In month two under that new, contact only when needed default, Panzeri's team ran an experiment clustering papers by cross field similarity instead of citation graph alone. It failed to beat the baseline overall, so by the team's own ship metric it was a dead end. But it surfaced the sixty eight percent number above as a side finding, and traced the cause: the ranking model had almost no click signal to learn from on papers this new, so it fell back on citation weight, which a brand new paper cannot have yet.

Five months sit between the finding and the sync that would have carried it. Nothing in that gap was anyone deciding to wait.

That finding sat in the team's shared notebook. Nobody outside the three researchers read it, because there was no standing place to bring it. In month seven, a rival product called Hallantide shipped a feature that boosted freshly posted papers closely matching a search, solving close to the same problem Panzeri's team had already found the shape of. Marrast heard about it from a customer call, went looking for anything Quillrace's own research had on cold start ranking, and found the sixty eight percent number sitting in a notebook entry dated five months earlier.

What it costs at its worst: Ashfen spent five months not shipping a fix it had effectively already found, while a rival shipped first and some researchers using Quillrace started trying Hallantide instead. Worse than the lost months, if this kept happening, research would learn that surfacing an in progress finding gets it either ignored or turned into a demand, so eventually nothing travels at all, not even the big ones.

What I would leave alone: the engineers keeping Quillrace's citation graph pipeline running don't need this softer cadence at all. That work already has a clear deliverable, ingest lag under a set number of hours, and a real deadline each sprint. Handing them an open ended "what surprised you" question instead of a deadline would be exactly the wrong move there.

The lesson: the opposite of too much pressure isn't no contact. It's a smaller, steadier kind of contact that doesn't ask for the one thing exploratory work can't yet give: a date.

Now here is the same thing as a story

The short version above is what you'd actually say in the room. Read this one for why the fix had to be a cadence and not just a resolution to "check in more."

Ludovica Panzeri has run applied research teams for eleven years, and she can tell within a week of starting a direction whether it's going to be a real answer or a detour that teaches you something anyway. Both kinds matter to her. Only one kind used to get told to anyone outside her own three person team.

When she joined Ashfen, it was just her, working close enough to the roadmap that a Friday check in with Marrast made sense, a quick demo, a number that moved, done. For nearly a year, that was fine. Then Ashfen brought on Bavo Trevisan and a third researcher to help chase a real, open problem: Quillrace's ranking was quietly bad at surfacing new work, and nobody yet knew why.

The weekly Friday meetings kept happening, unchanged, because nobody had thought to change them. For a while the team tried to keep up, bringing whatever partial result they had. Then, week by week, the partial results got a little cleaner before they showed up, a little more finished looking, a little further from what had actually happened that week.

Then came the meeting where Bavo said it out loud, not angry, just tired: they'd started deciding what to tell Marrast based on what sounded done, not on what was true. Marrast sat with that for a day, felt genuinely bad about it, and did the thing that felt like the opposite of the mistake. He cancelled the standing meeting. Told the team, work however you need to, ping me when something's ready.

Same failed experiment, same real finding inside it. What changes is only whether there's a road for it to travel on.

For a while, that felt like relief on both sides. Research stopped performing for a weekly audience. Marrast stopped hearing about work that wasn't ready. Nobody noticed, because nothing about "nobody noticed" looks like a problem while it's happening.

In month two of the new arrangement, the team ran an experiment clustering papers by cross field similarity rather than citation weight. It lost to the baseline on the team's own eval set, a clean failure by their own ship bar. But buried in the write up was a side finding: papers posted in the last ninety days, with fewer than five citations, were landing outside the top twenty results sixty eight percent of the time when they were genuinely the best match. The model had almost no click history to learn from on anything that new, so it fell back on citation weight, which by definition a brand new paper can't have.

Panzeri's team wrote that up carefully, the way they wrote up everything now, for themselves. Nobody outside the room read it. There was no meeting left to bring it to, and nobody wanted to be the one asking Marrast for time on a direction that had, by their own scoring, already failed.

Month seven, on a call with a customer weighing Quillrace against a competitor, Marrast heard the name Hallantide for the first time in that context. Hallantide had just shipped a feature that boosted freshly posted papers closely matching a search, close kin to the exact problem shape Panzeri's team had already found. He came back from that call unsettled enough to go digging through the team's own experiment notebook, half expecting to find nothing useful.

He found the sixty eight percent number, dated five months earlier, with the cause already traced. Not a rough idea. A specific, checkable finding, sitting exactly where good research puts things: written down, precise, and completely unreachable by anyone who needed it.

The old decision sat in that single meeting, the one where Bavo said the quiet part out loud. Cancelling the standing check in felt, in that moment, like the responsible response to a real problem. It was the right diagnosis and the wrong prescription. The team had grown from one person doing near shippable work to three people doing genuinely open exploration, and the old cadence had never been redesigned to match that, only removed once it started to hurt.

Run the same month two again, with the sync Marrast eventually built. Panzeri's team's failed clustering run still fails the eval bar. Nothing about the finding is different. But at the very next sync, at most fourteen days out, someone says: this run didn't work, but here's something strange, new papers are basically invisible no matter how good a match they are. Marrast writes three sentences that evening on what that might mean for ranking, sends it back to the team, and asks if he's reading it right. They tell him where he's overreaching. A validated fix, checked against a held out set of real queries, ships within the quarter, months before Hallantide's launch, not two months scrambling after it.

What I'd tell myself, sitting in that room the day Bavo spoke up: the mistake wasn't asking for updates. It was asking for the wrong thing, a date, instead of asking for the one thing research can always honestly give you, whatever surprised them.

SPARK, mapped onto a relationship instead of a screen

Not a script for sounding collaborative. SPARK is what forces you to name the one concrete decision, the cadence itself, and prove it survives both ways a research relationship actually breaks.

The whole answer in one picture. A small fixed ritual, one honest question, a memo that travels back to the people who earned it.

Same finding, same team, same failed run. The only thing that changes the outcome is whether the road exists.

A cadence that never asks for these three things is what actually earns research's honesty.

The recap, one line per letter: situation is a lab whose real findings had no standing road to product, payoff is research staying honest about failure without fearing what it costs them, anchor is the fixed sync and the memo that travels back rather than up, risk is either extreme costing months on a different problem, and keep out draws the line at a date, never at the question itself.

And if you want to be sure it really works, try it somewhere else

Same five letters, a veterinary diagnostics company instead of a paper search tool, and this time the open ended research isn't about ranking. It's about which lesions a scan model has never learned to see.

Corvusveil Diagnostics builds Petronix, an AI tool that reads veterinary X-rays and flags anything a vet should look at closer before signing off. Fletcher Scanlon owns Petronix's model roadmap. Corvusveil's imaging research group is exploring why small, early stage soft tissue lesions get missed on breeds whose scans looked different from most of the model's training data.

Same shaped ritual, a scan instead of a paper. The finding still has to travel through a fixed road, not luck.

Mapped onto SPARK: the situation is a research group whose early, breed specific findings had no standing place to land before Scanlon designed one. The payoff is a research team that keeps flagging what's hard about a training gap the moment they see it, not once they've built a whole new model around it. The anchor is the same biweekly sync, one question, a memo that goes back to research. The risk is the same both ways, ask for a demo too soon and research starts smoothing its early findings, go quiet too long and a real training gap sits unaddressed while a clinic somewhere signs off on a missed lesion. Keep out draws the same line: no ship date before the fix is checked against a held out set of scans from breeds it wasn't trained on.

Swap the trigger and it still runs.
Speed: an interviewer caps you at ninety seconds. Skip straight to the anchor, say the one question and the memo, and what it protects against.
Cost: no budget for a standing weekly ritual across every research thread. Run the sync biweekly instead of weekly, the finding still travels inside two weeks instead of five months, and the PM's time cost is half.
The model got better, for real: say Quillrace's baseline ranking doubles in overall accuracy. The cadence still matters, because a better baseline still learns from click history, and a brand new paper still has none to learn from. The cold start problem doesn't shrink just because the average got better.

Where people run it wrong.
They let the sync quietly turn into a status meeting again, asking "so what's shippable," which brings the whole old failure mode back in a new outfit.
They skip the memo step and just nod along in the room, so nothing gets written down and the same finding gets rediscovered later with no record it was ever said the first time.
They ship a fix straight from one dramatic near miss without checking it against a broader held out set, and quietly hurt ranking on the ordinary, well established queries nobody was watching.

How to use it live. Before answering a "what does a healthy research relationship look like" question cold, ask yourself one thing: what's the smallest standing structure that lets an honestly unfinished finding travel, without anyone being asked to promise something the work can't promise yet. Naming that structure, not a value word like trust or openness, is usually exactly what the question is listening for.

Flashcards (tap any card to flip it)

Check yourself Score: 0 / 0