The failure mode at the end of every archive
Every preservation effort eventually meets the same fork. One path is the museum: acquire comprehensively, because omission feels like loss and completeness feels like virtue. The other is the woodpile: acquire deliberately, keep what a winter might burn, and let the rest stay where it is. The museum path fails by drowning — budgets and attention exhaust themselves across an unbounded collection, and the archive degrades everywhere at once. The woodpile path fails rarely, and only by omission — a specific thing you skipped turns out to matter.
Model preservation has the same fork, with a twist that makes the museum path even worse here than usual: open-weight repositories are mostly derivatives of a modest number of originals. For any significant base model, the ecosystem produces scores of quantizations, merges, and format conversions — and, on a different axis, fine-tunes that encode genuinely new training. Collecting "everything about model X" is collecting the same information repeatedly at degrading fidelities. The museum in this domain is not just expensive; most of its halls hold copies.
The woodpile rules
The discipline that keeps a pile a pile and not a hoard is short:
- One authoritative copy per thing. The highest-fidelity representation the maker actually published — not the largest file by accident, not the most convenient quantization. Others can be re-derived; the original cannot.
- A reason for every acquisition. Vessel argument, research dependency, or plain respect — but nameable. "It was on the front page" is not a reason; it is a vector for other people's ranking decisions to become your disk usage.
- Deliberate exceptions. A quantized derivative earns shelf space when the use case is concrete (this device, this runtime, this week) — labeled as what it is, so nobody later mistakes the derivative for the original.
- Patience over speed. Acquisition is paced, resumable, polite to sources. An archive built in a hurry inherits the hurry as corruption risk.
Notice what is missing from the list: completeness. A woodpile is not judged by coverage; it is judged by whether the specific winters you expect are stocked for. That standard lets you say no — the most important word in acquisition.
Why this ends track one
You now have the argument (links are loans), the definition (four conditions of ownership), the deadline (three clocks), the payoff (vessels), and the discipline (the woodpile rules). Everything after this point is craft: reading the artifacts, judging their sources, verifying their bytes, and building the shed they live in. The philosophy is done — the hands start now.