how we work it out
the evidence we lean on, with the studies named, and the places where we are still guessing.
the idea
jollygig is built on one idea: you can plan and run things well without knowing how they will turn out, as long as you are honest about what you do not know. that goes for us too. below is the evidence we lean on for each part of the product, and the places where we are still guessing. where a claim rests on a study, the study is named in brackets so you can look it up.
team health
the team, not the person
we read the team, not individuals. psychological safety is a belief the whole team shares rather than a trait of one person, and it predicts how well teams learn and perform (edmondson 1999; frazier et al. 2017). the mood of a team is a group property too (barsade & gibson 2007). burnout research treats burnout as mostly a feature of the job, not the person (maslach & leiter 2016; bakker & demerouti 2017). a review of trials in doctors found that changes aimed at the organisation worked better than changes aimed at the individual (panagioti et al. 2017, jama internal medicine). so our action plans point at conditions, and we say "tend to work better" because that is as far as the evidence goes.
the six areas
a check-in asks which of six areas is on your mind. we chose them by reading two well-used models: the job demands and resources behind burnout research (demerouti et al. 2001; bakker & demerouti 2017), and the six areas of worklife, which are workload, control, reward, community, fairness and values (maslach, schaufeli & leiter 2001). psychological safety sits inside "relationships" (edmondson 1999). the mapping is our reading of those models. we have not tested it.
why answers have no name, and why five
an answer is stored with no name, no email and no user id, and we show nothing until at least five people have answered. every breakdown needs five of its own. the idea comes from k-anonymity (sweeney 2002), and five is the default threshold in tools such as microsoft viva glint. managers do not answer their own team's check-ins and are left out of the count, because someone who knows their own answer could subtract it. five is a convention that makes guessing unreliable. it is not a guarantee: if everyone in a group gives the same answer, that answer is still visible.
short, with a reminder
a check-in is three questions, because a single question can work well for a broad feeling (wanous et al. 1997). by default one reminder goes a day before a check-in closes, and a team can change that or turn it off. follow-ups raise response (dillman, smyth & christian 2014). the reminder goes only to people who have not taken part yet, worked out from a separate record of who took part that is never joined to what anyone said. we make no claim about how often to ask, because we have not found a study that settles it.
how a share becomes "active now" or "emerging"
when a check-in closes, a small bayesian model (gelman et al. 2013) works out how likely it is that a pattern is common in recent check-ins, counting newer check-ins more than older ones. it also wants clear evidence before a label changes, so a pattern does not flicker between "active now" and "emerging". the share is a share of check-ins, not of people.
scheduling
people guess short
people are reliably optimistic about how long tasks take (buehler, griffin & ross 1994), and the ranges they give are too narrow, so the real answer lands outside them more often than they expect (soll & klayman 2004; jørgensen et al. 2004). that is why a p50 is an expectation, about as likely to be beaten as missed, and the p95 is the better date to promise. we show both. the product does not correct for the optimism in the numbers you type in. once five activities have finished, the forecast panel shows how often the actual landed inside the range and offers a stretch based on it.
three points, three shapes
each activity takes three numbers: a best plausible case, a most likely case and a worst plausible case, the kind of estimate three-point methods have used since the 1950s (malcolm et al. 1959). we offer three shapes for turning them into a spread, because none fits every kind of work: a beta (pert) shape by default, a triangle, and a lognormal. the beta treats your best and worst as hard limits, and formulas built on percentiles estimate better than the classic pert one (keefer & verdini 1993). real overruns have long tails, and a lognormal fits that kind of duration better than the other shapes (mohan et al. 2007). in the lognormal, best and worst mean roughly the 5th and 95th percentiles, not hard limits.
tasks share causes
treating every task as independent makes the finish look narrower than it is, because tasks share causes: the same person, the same vendor, the same unknown (van dorp & duffey 1999; touran & wiser 1992). correlation barely moves the average finish and mostly fattens the late tail, which is where the p95 lives. so we ask about four things that tend to link two tasks: a shared person, a shared source of uncertainty, the same technical area, and one estimate being anchored on another. the numbers behind those four ticks are a rule of thumb, not a measurement.
which tasks matter
the criticality index is the share of simulated runs in which a task sat on the critical path (van slyke 1963). a companion measure, how closely a task's duration moves the finish, is also standard (williams 1992). we use them to show where attention pays off, never to grade anyone: the spread belongs to the work.
both together
when a workspace is linked to a team and that team shows signs of strain, the forecast panel offers a what-if: widen the pessimistic estimates. you choose to apply it, the widening is shown next to the results, and you can undo it. your own estimates are never changed, and it never names a person. there is evidence that teams under sustained time pressure deliver less reliably (kuutila et al. 2020). no study we know of says how much to widen an estimate, so we do not pretend to know.
what we have not validated
- the persona names and the emotion words are ours. we have not checked them against a standard scale, and a persona is a label for a pattern of what people chose, not a measured trait.
- the check-in cadence defaults are a choice, not a finding. the same goes for the balance of positive and negative words.
- the model's constants are judgement calls: how much weight its starting guess gets, how fast older check-ins fade, and the levels where a label turns on and off. we plan to characterise them by simulation. the same people answer each check-in, so the model is probably more confident than it should be.
- the correlation weights are a heuristic with no empirical basis. the way we link tasks also understates the chance of many going wrong at once (embrechts, mcneil & straumann 2002).
- the score we show for what drives a forecast is our own combination of the two standard measures above.
help us check it
if you are a researcher and want to help us check any of this, write to hello@jollygig.com.