In 2013 a hard drive holding the private key to a wallet containing 8,000 BTC was thrown out during a clear-out and taken to the Docksway landfill in Newport, Wales.
More than a decade of attempts to retrieve it followed. In January 2025 the High Court in Cardiff dismissed the claim to access the site, ruling there were no realistic prospects of success and that the landfill and everything in it belongs to Newport City Council under the Control of Pollution Act 1974. In March 2025 the Court of Appeal refused permission to proceed. An offer has since been made to buy the relevant section of the site outright. The landfill is scheduled to close.
Now be precise about what failed, because "he lost a hard drive" misses it.
There was a backup. He knew what it was, he knew roughly where it was, within a few thousand square metres of a known site.
What was missing was redundancy. One copy. In one object. Whose continued existence depended on nobody throwing it away.
And the second failure is the one this lesson exists for. The moment that object left his control, the recovery problem stopped being technical and became legal, and the legal system said no, three times, over twelve years.
A backup with one copy is not a backup. It is a single point of failure you have written down.
F103-02 told you what a seed phrase is. This lesson is where you keep it, and it is the last thing in the Freshman level that is purely under your control.
Two failure modes, pulling opposite ways
Most people plan for one failure: somebody steals the seed. Serious backup plans for two, of equal weight.
Theft. Someone reads, photographs, or carries away the phrase.
Loss. Fire, flood, a house move, a clear-out, dementia, divorce, a forgotten location, a landfill.
Almost every "best practice" you will read addresses one of these while quietly worsening the other. A cloud backup solves loss and creates theft. A single copy hidden brilliantly solves theft and creates the autopsy.
Every choice below is a trade, and the point is to make each one knowingly rather than to find a scheme with no downside.
The rules that are not trade-offs
These are mistakes, not choices. There is no threat model under which they are right.
No photographs. A picture of the seed is in your camera roll, therefore in your cloud, therefore in whatever backup you have forgotten about, therefore reachable by anyone who compromises that account.
No cloud documents. A note, a doc, a draft email to yourself. F106-05's autopsy is 30 million encrypted password manager vaults taken in 2022 and drained through 2024 and 2025.
No password manager that syncs. Fine for passwords. Not for a seed, for the reversibility reason in F106-03: a leaked password can be changed and a leaked seed cannot.
No screenshots. They land in the same places photographs do.
No typing the seed on a networked device, except during a deliberate restore on a device you trust. Treat keyboard input as a broadcast, because to malware it is one.
If you have already broken one of these, you have not necessarily lost anything. You have raised the risk, and F106-05's rule applies without exception: a secret that was ever exposed is compromised regardless of elapsed time. Generate a new seed, move the funds, retire the old one. An afternoon and some fees.
The choices that are trades
Material. Paper survives until it does not: fire, flood, a leaking roof, twenty years of attic humidity. Stamped metal survives essentially everything except being carried away. Note the word stamped: ink on metal fails in a fire, which defeats the entire purpose. Most serious holders use metal as primary and paper as temporary.
Copy count. One copy is the autopsy. More copies survive loss better and expose you to theft more. The arithmetic below is why this is the highest-value decision.
Locations. Two copies in one house is one copy for fire purposes. The pair must be chosen so that no single event reaches both, and "single event" includes a burglary, a flood, a house fire, and a family dispute.
Splitting. F110-03's Shamir gives you the only way to add redundancy without adding exposure, and it is the correct answer once the amount justifies the complexity.
Telling someone. Maximally theft-resistant is telling nobody, and it is also how funds die with their owner. F110 is an entire course on that failure and this decision belongs to it.
Treat each copy as surviving a decade with some probability, and assume the copies fail independently, which is what separate locations buys you.
One paper copy in one place. Call it 90 percent over ten years.
Survival: 0.90. One chance in ten of the autopsy.
Upgrade the material. Metal in the same place, say 98 percent.
Survival: 0.98. A real improvement, and note what it cannot fix: the dominant risk in the same-location case is not the material failing, it is the location failing, and a house fire takes metal you cannot find as surely as paper.
Two paper copies in separate locations, back at 90 percent each. The scheme fails only if both fail:
P(both fail) = 0.10 x 0.10 = 0.01
Survival: 0.99
Two independent paper copies beat one metal copy, and cost almost nothing.
Two metal copies in separate locations:
P(both fail) = 0.02 x 0.02 = 0.0004
Survival: 0.9996
Now the comparison that should change what you do this week. Going from one copy to two took you from a 10 percent chance of total loss to 1 percent, a tenfold reduction, for the price of a second sheet of paper and a drive to a relative's house. Going from paper to metal on a single copy took 10 percent to 2 percent, a fivefold reduction, and cost you money and effort.
The general shape: survival multiplies across independent copies and only adds within one. Redundancy is the cheapest large gain available, and it is the one people skip while shopping for materials.
One caution the arithmetic hides. Independence is the assumption doing all the work. Two copies in one house are not independent. Two copies both known only to you are not independent against you forgetting, which is why F110-06 exists.
A configuration that survives a fire
Not the only right answer. One defensible one.
- Generate offline, on the hardware wallet, per F104-03.
- Stamp two metal plates, identical, stamped rather than inked.
- Separate them by a real distance. Home safe and a relative's safe. Home safe and a deposit box. The specific pair depends on your trust map; the requirement is that no single event reaches both.
- Verify by restoring, before any meaningful funds exist. This is F104-05 and it is the step everybody skips.
- Document the locations, not the contents. A note for your executor saying where the plates are and how to use them, never what they say. F110-06 is that document.
That survives a house fire, a drive failure, a stolen laptop, a compromised email account and most burglaries. It does not survive a coordinated attack on both locations, and it does not survive you never telling anyone.
Backup is not the same as plan
The distinction the autopsy turns on, and it is the handoff into F110.
A backup is an object. Metal in a safe.
A plan says where the objects are, who is allowed to retrieve them, in what order, and what they should do when they get there.
Howells had a backup. He did not have a plan, and the moment the object left his control the problem stopped being one he could solve with technology.
Tier your effort with F104-02: the burner needs no backup at all, the daily needs one copy somewhere sensible, and the vault gets everything above. Backing up three seeds to vault standard is how people abandon the architecture in month two.
Memorising the seed is the ultimate backup, because it cannot be stolen or burned.
It is genuinely appealing and it fails in three specific ways, which is worth spelling out because the idea keeps resurfacing.
Memory degrades silently. You will not receive a warning. You will discover the gap on the day you need it, under pressure, which is the worst possible test condition.
It does not survive you, which makes it the F110 problem in its purest form. A memorised seed dies with its holder, guaranteed.
It is not theft-proof either. F110-05's autopsy is a kidnapping. Memory is exactly as extractable as a safe under coercion, and rather more painfully.
There is a legitimate version, and it is narrow: memorising a passphrase while the seed is written down. That combination means a thief who finds your metal plate gets an empty wallet, which is the plausible-deniability property from F106-04. It carries the same three failure modes and applies them to a secret whose loss you can survive if the seed itself is intact and you have written the passphrase down somewhere else too.
Which is to say: memory is a useful second factor. It is not a backup.
Design against theft and loss at once, because they pull in opposite directions and most advice quietly trades one for the other. Some rules are not trades and have no valid threat model: no photos, no cloud, no synced password managers, no screenshots, no typing the seed on a networked machine. Then make the trades knowingly, and start with copy count rather than material, because survival multiplies across independent copies and only adds within one: two paper copies in separate places beat one metal copy in one. And remember what the landfill actually teaches, which is that a backup is an object while a plan says where the objects are and who may retrieve them, and once the object is gone the problem is no longer technical.