ADR-041: An algorithm-local tuning constant belongs in the reading of the paper that measured it, not in the practice registry

Context

The import behind #139 brought 187 recommendations out of 70 readings, and the single largest class in what remained after the obvious filings was constants:

• “Set alpha = 2/(m+2) for EASGD”
• “Use rank 2 for CNNs and rank 4 for LSTMs” (PowerSGD)
• “Set tau = 48 for SGP and tau = 12 for local SGD” (SlowMo)
• “Set Moshpit grid dimensions so M^d approximates the worker count”
• “Use alpha = 0.6 and anchor momentum beta = 0.7” (Overlap Local-SGD)

Every one is correct, measured, and useful. None of them has been filed, and the reason each was passed over has been re-derived by hand each time rather than decided once. That is the thing this closes: not the question of whether they are worth having — they are, and they are already in the record — but the question of where, so that the next contributor meeting the twentieth one does not have to think about it.

The pull toward filing them is real and worth naming. They are the most concrete, most actionable sentences in any of these papers. A practice registry that holds “prefer decentralized SGD below 1 Gbps” and not “use rank 4 for LSTMs” can look like it is holding the vague half.

Decision

A recommendation whose subject is a named algorithm’s own hyperparameter stays in the reading. It does not become a practice.

The test, which is mechanical enough to apply without judgement: remove the algorithm’s name from the sentence. If what remains is an instruction, it is a practice. If what remains is nothing — “set alpha to 2/(m+2)” — then the sentence was a fact about that algorithm and its home is the note on the paper that measured it.

Two consequences that follow and are the point:

The note already holds it, findably. ## Recommendations in a NOTE is a rendered section with the strength and the conditions the reading recorded. Nothing is lost by not promoting it; a reader who has decided to run PowerSGD reaches the PowerSGD note, which is where the rank is.

A practice may still cite the constant in prose. SOTA-226 names LIT-254’s 30-epoch densification schedule and declines it; SOTA-222 names LIT-323’s anchor-momentum pair and declines it. Declining to give a number its own document is not declining to mention it, and saying why in the practice that neighbours it is how a reader learns the boundary.

What this does not cover

A constant that transfers is not algorithm-local. beta2 = 0.999 is in this record as SOTA-002 because it is what people set on a family of optimizers across a decade of workloads, not because Adam’s paper reported it. The distinction is whether the number outlived the paper, and that is a judgement, not a test.

A threshold about the world is not a tuning constant. SOTA-225 carries “below about 1 Gbps or above about 5 ms”, which is a number, is dated, and is filed anyway — because it is a fact about when one regime beats another, and the reader it addresses has not yet chosen an algorithm. The line is between a number that configures a thing you have chosen and a number that helps you choose.

This is not a claim that the class is empty. Some of the 157 will turn out to be practices wearing a constant’s clothes, and the test above is what catches them. It is a claim that the default is “no”, stated once.

Alternatives considered

• File each as a Deferred practice with a promote_when naming a second workload. This is what DP-006 would suggest for a claim with weak-but-honest evidence, and it is the wrong reading of DP-006 here: the evidence is not weak, the scope is narrow. Deferred says “probably right, not yet worth asserting”, and “use rank 4 for LSTMs” is not something the record is hesitating about. It is something the record is not addressed to.
• Add a scheme for them. A CONST or PARAM family alongside SOTA, LIT, THEORY and NOTE. Rejected because the NOTE scheme already is it: one document per reading, with a recommendations section, which is exactly where a paper’s own settings belong. A fifth scheme would move them without changing what they are.
• Decide per paper, as now. What produced the #139 backlog, and the cost is not the thinking — it is that the thinking leaves no trace, so the hundredth constant costs what the first did.