The index · rubric v2 · August 2026
How we decide whether a tool is worth replacing.
The verdicts
Five, not two.
Two verdicts leave a hole exactly where the largest line item sits: nobody replaces their cloud provider, and “keep” is not useful advice about a bill that size.
- Replace now
- Worth replacing, and it pays for itself fast enough to go first. Both conditions, not either.
- Replace next
- Worth replacing on the score, but the payback is slower — so it is phase two, not phase one.
- Optimise
- You cannot replace it, only stop overpaying for it. This is where infrastructure lands by default.
- Repatriate
- Infrastructure where owning the hardware plausibly beats renting it. Never quoted with a figure — see the limitations.
- Keep
- Do not touch it. Either a gate fired, or it scored below the line. Roughly half of a typical stack.
Step 1 · gates
Four gates run before anything is scored.
A gate is a statement about what a tool is, not how much it costs — so no score can override one. This ordering is deliberate: it is what stops a large enough bill from arguing you out of a structural fact.
| If the tool is | Verdict | Why |
|---|---|---|
| Network effects | Keep | Rebuilding what everyone outside your company already uses is an own-goal. Slack, Stripe, Google Workspace. |
| Regulated surface | Keep | Tax tables and filing formats change yearly. The maintenance outruns the saving, permanently. |
| Data moat | Keep | You are paying for their dataset, not their software. Rebuilding gets you the software only. |
| Infrastructure | Optimise or Repatriate | Never Keep and never Replace. Which of the two depends on the infrastructure profile you give us. |
| Cannot pay back | Keep | Below roughly $6,000 a year the arithmetic cannot work whatever the score says, because any build has a floor cost. |
A gate cannot be overridden by a score. It can be overridden by you — you know things about your own stack that a catalogue does not, so any verdict can be set by hand and the plan records that it was.
Step 2 · the score
Four factors, weighted.
What is left after the gates gets a 0–100 replaceability score. This replaced a set of absolute-dollar thresholds, which produced cliff edges — $7,999 was a Keep and $8,000 was a Replace, with nothing about the business different either side of the line.
- Share of your stack
- 40%weight
- How little you use
- 25%weight
- How far you have bent it
- 20%weight
- How loosely it is wired in
- 15%weight
| Class | Utilisation | Customisation | Integration |
|---|---|---|---|
| workflow | 0.45 | 0.65 | 0.45 |
| commodity | 0.6 | 0.3 | 0.35 |
| network | 0.85 | 0.1 | 0.85 |
| regulated | 0.7 | 0.25 | 0.7 |
| infra | 0.8 | 0.4 | 0.9 |
| data-moat | 0.7 | 0.15 | 0.3 |
Step 3 · the split
The score decides whether. Payback decides when.
One number cannot answer both questions, and trying made the model wrong in a specific way.
- Worth replacing at all
- 45+score
- Goes first
- ≤ 12 mopayback
- Longest payback we would back
- 3 yr
The floor
Why some tools are Keep whatever their score.
Not a threshold we picked — one that falls out of the arithmetic.
Any replacement has a floor cost: even a small tool takes real work to replace. Divide that floor by the longest payback we would recommend and you get an annual spend below which no honest recommendation exists — about $6,000 a year on current figures. Under that line the answer is Keep, and it is Keep for a reason we can show you rather than a number we chose.
Limitations
What this rubric cannot currently tell you.
Published with the same prominence as the method, because openness that omits the caveats is just marketing with tables in it.
- 01
The weights are reasoned, not calibrated
The four weights and both thresholds are arguments, not findings — they have not been tuned against real deal outcomes, because we do not yet have enough of them. They will move, and when they do the change and its reason get published.
- 02
Every catalogue price is an unverified estimate
The list prices behind each tool have not been checked against a source yet. That is why the teardown asks you to correct them and why it says which figures are ours rather than yours — and it is the reason the cost benchmarks are not published.
- 03
Build costs and paybacks come from a placeholder rate
The build estimate is 0.62× annual spend with a $18,000 floor, and infrastructure optimisation assumes 25% recoverable. Both are illustrative pending real delivery costs. Every payback figure the tool produces inherits that uncertainty.
- 04
Repatriation is never given a figure
A Repatriate verdict tells you to model it, not what it would save. Cloud exit economics depend on your utilisation curve, your egress, your commitments and your appetite for running hardware — a number from a webpage would be fiction, and it is the largest line on the page, so it would be expensive fiction.
- 05
It scores tools, not organisations
The rubric knows nothing about your team’s capacity, your security review queue, or the renewal you signed last month. A verdict is an input to that judgement, not a replacement for it.
Citing this
Now run it on your own stack.
Four minutes, no email. Every verdict shows the score or the gate behind it, so you can check the rubric above against your own result line by line.