post(blog): "Deciding What to Buy" — a purchase decision framework

bytesofpurpose-blog/blog/2026-08-04-deciding-what-to-buy.md

@@ -0,0 +1,176 @@
+---
+slug: deciding-what-to-buy
+title: "Deciding What to Buy"
+kind: framework
+sidebar_label: "Deciding What to Buy"
+description: "A framework for making a purchase decision on evidence instead of spec sheets: rule things out before you score them, grade every claim by how you learned it, and never quote an asking price as a market price."
+authors: [oeid]
+tags: [decision-making, evaluation, best-practices, research]
+date: 2026-08-04T10:00
+draft: true
+---
+
+The listing said the printer could run PEEK-CF. I looked up the machine's own spec sheet: the
+nozzle tops out at 280 °C. PEEK needs somewhere around 400 °C and a heated chamber. Not a stretch,
+not a "your mileage may vary" — physically impossible on that hardware.
+
+That one line took ninety seconds to check and told me more than the rest of the listing combined.
+Not because the material mattered to me — it didn't — but because the seller had just demonstrated
+they'd never run the machine.
+
+<!-- truncate -->
+
+I set out to answer one question — *is this used printer worth it?* — and ended up building a
+framework, because the reasoning kept being the same reasoning and I kept rebuilding it by hand.
+This is that framework. The printer is just the worked example; the method is what transfers.
+
+## Why the normal approach fails
+
+The normal approach is: open four tabs, compare spec sheets, pick the one with the best numbers.
+
+It fails because **spec sheets are marketing**, and the things that actually determine whether you
+get what you wanted appear on none of them:
+
+- What the replacement part costs when the common failure happens — and whether it's a $30 part or a
+  $180 sealed assembly that contains the $30 part
+- How long spares will exist, which almost nobody publishes and which is therefore itself a signal
+- What it actually sells for used, as opposed to what people are asking
+- Which axis is slightly off, which the manufacturer will never tell you and the owners will
+
+Meanwhile the specs that *are* published are chosen to be flattering, quoted "up to," and measured
+under conditions nobody names.
+
+## What I actually did
+
+Six steps. The order matters more than any individual step.
+
+**1. Pin the requirements before looking at any product.** Not specs — the *job*. What's the biggest
+thing you need to make? Where will it live? Who's going to run it? A person knows what they want to
+do. They don't know what build volume, tolerance, or duty cycle that implies, and asking them to
+translate is asking them to do the hard part themselves.
+
+Write these down and stop. Everything downstream is built on them, and adding one later is expensive
+— more on that below.
+
+**2. Rule things out loudly, before scoring anything.** Some failures aren't low scores, they're
+disqualifications: it doesn't fit, it can't cut that material, there's no ventilation path in that
+room, the electrical service doesn't exist, it's end-of-life inside your ownership window.
+
+The temptation is to score those candidates anyway and let them lose on points. Don't. A
+disqualification buried inside a weighted average reads as *merely weak* rather than *impossible*,
+and a table is very good at making impossible look like 2.4.
+
+**3. Score the survivors against written anchors.** Not "5 = excellent accuracy," which says nothing.
+`5 = ±0.05 mm held across a batch, under load.` Anchors written as observable outcomes mean the same
+thing across candidates and across years. Anchors written as adjectives mean whatever you felt like
+that afternoon.
+
+**4. Grade every score by how you learned it.** This turned out to be the most valuable single idea,
+and it came from a frustration: **nobody publishes comparable test data.** No reviewer had run all
+three candidate machines on the same file. Every comparison I could find was assembled from
+different materials, different profiles, different rooms, different operators.
+
+So each cell carries a grade:
+
+| Grade | Means |
+|---|---|
+| **A** | Independently measured — somebody put calipers on it |
+| **B** | Reviewed but not measured — described, no figure |
+| **C** | Owner report — a forum post, a seller claim, one anecdote |
+| **D** | Inference — reasoned from specs or a sibling model |
+
+And it goes **inline, in the cell**, not in a footnote. A footnote is a place where honesty goes to
+be skipped. Inline, a table full of Ds looks like a table full of Ds, which is exactly what it is.
+
+The rule that makes this work: **a criterion you couldn't verify gets a D with a note saying what you
+checked — never a blank, never a quiet guess.** A recorded gap stays visible. A skipped one silently
+narrows the comparison down to whatever happened to be easy to find.
+
+**5. Weight by use case, not in the abstract.** The same eleven criteria produce different winners
+depending on whether you're selling parts, prototyping, or doing this on weekends. Set the weights
+from what the person told you in step 1. If two weighting profiles never change the ranking, they're
+the same profile and one of them is decoration.
+
+**6. Compute cost over three years, separately from the score.** Purchase price plus consumables plus
+the wear part plus the thing nobody budgets for — in this category, ventilation and the electrical
+work — minus realistic resale. If the winner on points is also the most expensive to run, one of your
+inputs is wrong. Go find out which.
+
+## What the research turned up
+
+Three findings that changed how I'd do this in any category.
+
+**The vendor's own support forum beat every review.** The single most decision-relevant fact in the
+whole comparison came from a manufacturer's support forum, not from a review: owners reporting a
+nominal 150 mm feature coming out at about 149 mm on stock profiles. The reviews described the
+machine as accurate. The forum said *which axis was off*.
+
+This is now near the top of my evidence hierarchy — above spec sheets, above independent reviews.
+Search the model name plus failure words: `accuracy`, `fails`, `warranty`, `won't`, `broke`. Then
+read the *replies*. How fast and how well the vendor answers **is** the support-quality score,
+observed instead of claimed. It's the highest-value search available and it's the one everybody
+skips.
+
+**Asking prices ran about 87% above sold prices.** On the sample I collected — small, and I'll say so
+rather than dress it up — the median asking price sat roughly 87% above the median confirmed sale.
+The forum "comps" I found for the machine I was looking at ($5,250 in 2022, $3,500 in 2023) were
+*never confirmed as sales at all*. Someone asked. That's all we know.
+
+This is why browsing a marketplace gives you a badly inflated sense of what something is worth: the
+listings that never sell stay up forever, so what you're actually browsing is a museum of prices
+nobody paid. Sold and asking are two different quantities and they must never be averaged together,
+drawn with the same mark, or quoted in the same sentence without saying which is which.
+
+**One check killed more candidates than price did: the ceiling.** A used item cannot rationally be
+worth more than the cheapest *new* item that meets all the same requirements — minus the warranty,
+the remaining support horizon, and a healthier parts supply. Work out that new number first. It's the
+anchor; everything else is a discount off it. Original MSRP is irrelevant, and a listing that leads
+with "originally $X" is anchoring you on purpose.
+
+## What changed my mind
+
+The recommendation reversed twice, and both reversals came from *searching* rather than reasoning
+from what I already knew.
+
+I started confident: buy the cheaper, better-supported current-generation machine; the used one has a
+dying consumable format. Then I found a competitor I hadn't considered that cleared the size
+requirement outright. Then I found a machine from the first manufacturer that inverted their own
+naming — the cheapest model in the line had the *largest* build volume and the *highest* speed.
+Nobody would guess that from the product names, and I certainly didn't.
+
+Related: one of the products I was comparing had a "+" in its name that I read as an upgrade tier. It
+isn't. It replaced the original and there is no cheaper base version. **Verify what a product name
+actually denotes before you put it in a table.**
+
+The lesson isn't "I was wrong twice." It's that **recalled knowledge about a product lineup is
+routinely a generation stale**, and a recommendation built on it can reverse completely on a single
+search. If your answer never changes during research, that's not confidence — it's a sign you were
+confirming rather than checking.
+
+The other reversal was self-inflicted. Late in the process I added a criterion — programmatic API
+control — and it reordered the entire ranking. One manufacturer's firmware verifies that commands
+come from authorized software, and the escape hatch requires going LAN-only, which kills the cloud
+features and the phone app. The other publishes an OpenAPI spec and runs local and cloud as one
+service.
+
+Adding that criterion at the end meant re-researching candidates I'd already finished. **Pin the
+requirements before you research, and build the matrix so it can be re-weighted rather than
+re-derived.**
+
+## What I'd tell someone starting from scratch
+
+- **Check one claimed capability against the vendor's own stated maximum.** Ninety seconds. If the
+  listing claims something the hardware can't do, every other claim in it is unverified — the seller
+  either never used the machine or copied a spec sheet.
+- **Find the cheapest new thing that meets all your requirements before you look at a single used
+  listing.** That's the ceiling, and it ends most "great deals" on its own.
+- **Read the manufacturer's support forum**, not the reviews. Search failure words and read the
+  replies.
+- **Write down what you don't know**, in the table, next to what you do. A comparison that only shows
+  the confident findings misrepresents itself, and the gaps are usually the more useful half.
+- **Say your sample size out loud**, in the same sentence as the number. Three listings is an
+  anecdote. I don't quote a range under eight.
+
+A polished decision matrix is a persuasive object, and it is very easy to build one that overstates
+what it knows. Most of the rules above exist for no other reason than to stop that — including when
+the person being persuaded is me.