A project health score you can argue with
Most health indicators are a coloured dot nobody trusts. Here is the exact arithmetic behind ours, and why no language model touches the number.
Every project tool eventually ships a health indicator. Most of them are a red, amber or green dot with no explanation, so the first time it disagrees with a manager it gets ignored forever. We wanted the opposite: a number you can take apart.
Five signals, fixed weights
The score starts at 100 and each signal subtracts points. The weights are published in the product, not buried in a blog post:
- Overdue ratio — 30 points. Full penalty once 40% of open tasks are past due.
- Blocked dependency chains — 20 points. Full penalty at 25% of open tasks blocked or waiting on an unfinished blocker.
- Velocity trend — 20 points. Compares points closed in the last 14 days against the 14 before. No history means no penalty.
- WIP overload — 15 points. The share of people carrying more than four tasks in flight.
- Silent tasks — 15 points. Open tasks with no activity for five or more days.
Why the model never computes it
A language model asked to "score this project" will produce a confident number with no reproducible basis, and it will produce a different one tomorrow from the same data. So the model is handed the finished arithmetic and asked to write two sentences of English. It cannot invent a figure because it is never asked for one.
The health score is the first project metric my team has not immediately dismissed, because you can see exactly which five numbers produced it.
The part people actually use
Underneath the score we rank the signals by how many points they cost and turn the top three into concrete actions: re-date these four overdue tasks, unblock this chain, rebalance this person. A score without a next step is just a mood.