A node’s column is its longest distance from a node nothing feeds. Sources sit left, consumers right, and a cycle is reported rather than drawn — a dependency graph with a loop in it is a finding, not a rendering problem.

What it assumes

That you can name every piece. The value of the exercise is mostly in discovering that you cannot: the map people draw from memory is reliably missing one vendor and one spreadsheet.

That “fragile” is a judgement you are willing to write down. Marking a handoff dashes its line and lists it in prose beneath the map, so the claim carries a reason and not just a colour.

What it does not model

Latency, volume, cost, redundancy or failure probability. It answers one question — what is downstream of this — and a map that tried to answer the others would answer none of them legibly.

Each score is multiplied by its criterion’s weight, and the products are summed. A blank cell removes that criterion from the option’s total and from the maximum it is measured against, so an unknown does not quietly count as a zero — an option nobody has evaluated should not lose on that basis.

Because options can be scored against different numbers of criteria, the totals are reported as a percentage of each option’s own available maximum. No option is called the leader unless it is clear of the next one by more than five points; below that the matrix is telling you the decision is a tie, and printing a winner would be inventing precision the scores do not have.

What it assumes

That the criteria are independent. They are not: time to first release and ongoing maintenance usually trade against each other, and scoring both high for the same option is a sign the scores are aspirational rather than observed.

That a weighted total is a prompt, not a verdict. The useful output is the moment the leader changes when you move one weight by one point — that tells you which disagreement the decision actually rests on.

What it does not model

Cost in money, timelines, staffing, or the risk that the vendor is acquired. It also cannot model the thing that decides most of these arguments in practice, which is whether the team wants to build it.

One multiplication, stated in full so you can check it against your own spreadsheet rather than trusting the box above.

attendees × (minutes ÷ 60) × loaded hourly cost × occurrences per year

Attendee hours are reported separately because they are the figure a manager can act on. A cost can be argued away as sunk; four hundred hours cannot.

What it assumes

That everyone invited attends, for the whole meeting, and that their time has one price. All three are wrong in detail and close enough in aggregate: a half-attending director and a fully-attending engineer roughly cancel out.

That the default $95 loaded hourly cost is a placeholder and nothing more. It is not a benchmark, it is not sourced from a salary survey, and the tool does not pretend otherwise — replace it with your own organisation’s figure before quoting any number from this page.

What it deliberately does not model

Preparation and follow-up, context switching either side of the meeting, the value of the decisions the meeting produces, and the cost of the meeting not happening. The first two make the real cost higher; the second two are the reason a high number is not automatically an argument for cancelling.