ADR-015: What the field thinks is a second axis, and this record's endorsement is not it

ADR-NNN: What the field thinks is a second axis, and this record’s endorsement is not it

Context

statuses.yaml opens with the reason this record has a status vocabulary at all: “A status every record shares is indistinguishable from no status at all.” 133 of 144 practices are Active. The vocabulary is being exercised, barely, and the reason is not laziness — it is that the one axis is being asked to carry facts that are not the same fact.

Take mHC. Today it is simultaneously:

adoption	shipped at 1.6T parameters in DeepSeek-V4 — one lab, production
consensus	two independent groups attacked its central constraint within a month, by different arguments, with opposite remedies
our position	Proposed

Those are three orthogonal readings and status: holds one. “Standard at one lab, contentious everywhere else” is a real and common state of a practice, and this record could not say it.

The deeper point, and the one that prompted this: advancement in a field is largely a social process. A paradigm shifts when the community behaves as though it has, not when a decisive experiment lands. A record whose only axis is its own endorsement is therefore structurally unable to distinguish we are not convinced from the field is not convinced — and those call for opposite actions. The first is a reason to go read something. The second is a fact to write down.

ADR-014 is where this bites hardest. It requires a provisional practice to say what would settle it, and the rule for phrasing one — name a kind of result, not a count — reads as though only evidential kinds qualify. Two of the eight conditions written under it are already social: SOTA-120 waits on a report that states which choice it makes, and SOTA-141’s first clause waits on a frontier recipe adopting the schedule and saying so. The mechanism was broader than its own rule.

Decision

A second axis, consensus:, backed by a scheme-local vocabulary (LU-ADR-076) and defaulted rather than required:

[luria.schemes.SOTA.fields.consensus]
vocabulary = "consensus"
default    = "unassessed"

Six values, in consensus.yaml: unassessed, unreplicated, contested, emerging, converged, universal.

unreplicated and contested are the distinction that motivated this. One group with one result that nobody has examined is not the same as a claim credible people are publicly arguing about, and Proposed covers both. SOTA-122 is the first; SOTA-136 is the second; they read identically today.

The axes are orthogonal, and the corners are the useful part.

• Deferred + converged — SOTA-120. AdamW is the default optimizer, so decoupled weight decay is what almost everyone runs; what is unmeasured is whether anyone chose it. High consensus, low editorial confidence, and the two-axis version says exactly that.
• Active + contested — this record being ahead of the field and saying so. Worth stating as legitimate, because the alternative reading of the Kuhnian point is that an anthology should wait for consensus, which would make it a lagging indicator by construction. Being early is part of the value; being early without flagging it is the failure.

unassessed is the default and most of the corpus carries it. A consensus reading is a judgement about a community and needs grounds. Ten practices are assessed here, each with a consensus_note: citing what the record actually holds; the other 134 say unassessed, which is true and countable. A value invented to fill a column would be worse than the admission.

Branched conflict

The second half of the same question: the record could not distinguish an agreed trunk from the forks contending off it.

The relations ADR-011 added to the literature are now declared on practices too — extends: and compared_against: — and walked into a second chain. The two layers answer different questions and can disagree, which is why they are separate chains rather than one: papers fork where the designs differ, practices fork where the recommendations do.

With the consensus axis rendered beside the status (dmarx/luria#173), the residual line now reads:

[SOTA-137](../practices.d/SOTA-137.md) — freely learned mixing            (Superseded, unassessed)
  [SOTA-136](../practices.d/SOTA-136.md) — doubly-stochastic mixing       (Proposed, contested)
alongside: [SOTA-133](../practices.d/SOTA-133.md) — attention residuals   (Active, emerging)

That is the shape the record had already worked out in prose and could not state. LIT-181’s standing section says the constraint is “the least well-defended part of the design, disputed by two groups within a month, while the thing all four papers agree on — widen the residual stream, constrain the mixing somehow — is untouched.”

The trunk practice that paragraph describes is deliberately not filed here. LIT-181 says the record should probably carry it eventually, and converting a hedge into an assertion is an editorial act, not a mechanical one. The structure now exists to hold it the day someone decides.

Alternatives considered

More statuses. Contested, Emerging alongside Active and Proposed. Rejected on two grounds. The status vocabulary is closed at five upstream (ADR-003), and more importantly it would re-commit the original error: mHC would still need one word for three facts, and the word would now be even harder to choose.

A numeric score — 0–5, or a fraction of adopting labs. Rejected as false precision. Nobody can defend 3 over 4, and a number invites arithmetic on a judgement.

Derive it from the graph — count production adoptions, count citing groups. Rejected, and it is the tempting one, for the reason ADR-014 already gives: a derived value is necessarily a count, and counting is the error that produced that decision. Two critiques and two endorsements is not a wash; it depends entirely on what they say.

Put it on LIT instead. A paper is not accepted or disputed — a claim is. Two papers can agree on a design and disagree on its constraint, which is precisely the residual case. The axis belongs where the claim lives.

The cost, stated plainly

A consensus value ages with nothing to trigger a re-read. Unlike a promote_when: condition, which names the evidence that would reopen the question, a consensus reading just quietly rots as the field moves. This is the real risk and there is no mechanism here that addresses it. Two things make it survivable for now: the default means only assessed entries can rot, and each assessment carries a consensus_note: saying what it rested on, so a reader can tell whether the grounds still hold. If the count of assessed practices grows past what one person can re-read, this needs a staleness report of its own.

Consequence for ADR-014

Amended in place (version 2): its phrasing rule now says explicitly that a threshold of adoption is a kind of result. The requirement is unchanged — before it existed, four of 144 practices stated a condition at all, and the failure was silence, not over-precision.