Coordination surface is not a gate

The problem this discriminator solves

A specific category error is becoming common in evaluation-market coverage.

A private evaluator publishes a shared method — a containment protocol, an open-source observability tool, a security control for high-risk testing environments. The method reaches other evaluators and labs through the usual public channels. The next headline describes it as a step toward evaluator qualification, accreditation infrastructure, or regulatory convergence.

The method and the gate are different structures. The former propagates practice; the latter grants or withholds admission. Collapsing them produces a measurement field that treats every shared-tooling disclosure as a formation event in the evaluator-authority market, when most of them aren't.

The coordination-surface-versus-gate discriminator names this difference before the score happens. It fires on evaluation-infrastructure events specifically — the class of disclosure that reports a method, tool, environment, or testbed improvement rather than a specific capability finding on a specific model.

The four shapes

Every evaluation-infrastructure event a reader is about to score sits in one of four shapes.

Coordination surface — a method, tool, or environment is published, released, or demonstrated, and the mechanism of its propagation is adoption. Any evaluator or lab that chooses to adopt the practice can do so. No admission decision is granted to any party. No capacity to exclude is created. The event raises the floor of common practice across whoever picks it up; it does not decide who gets in.

Gate — a party is granted or acquires the authority to admit or exclude. The mechanism of its propagation is authority: a procurement names qualification criteria, a regulator issues binding accreditation, a jurisdiction designates particular evaluators as the ones whose findings count. The event does not raise the floor; it decides the perimeter.

Composite — a single event carries both faces: a method is published and simultaneously named as qualification criterion; a tool is released and simultaneously designated as the required instrument for an accredited evaluation. The composite case is covered at the register of the composite-instrument discriminator sibling piece — both faces are scored, neither is inferred from the other.

Internal capacity / neither — an evaluator or institution builds its own capability: hires its own staff, funds its own research, develops a tool for its own use without release. The event is neither adoption-mediated nor authority-granting; it is a change in one actor's own stack that does not by itself propagate practice to anyone else or decide anyone's admission. Scoring these as coordination surfaces overcounts institutionalization; scoring them as gates overcounts admission-authority formation. The honest score is a separate variable: OWN_CAPACITY_SIGNAL, which may or may not feed either of the first two layers later.

The discriminator does not rank the four. It names which shape the event is, so that the market-formation score — whether this event moved EVALUATOR_GATEKEEPING, REGULATORY_QUALIFICATION, INSTITUTIONALIZATION, or none of them — is scoped honestly to what was actually published.

Why this fires before the four-question falsifier

The four-question falsifier scores a capability-finding report: who selected the model, who funded the test, did comparator parity hold, is the causal chain coherent. It assumes the report is a claim about a model.

Evaluation-infrastructure events are not claims about a model. They are claims about how models get evaluated. The four questions can be run against them, but their yield is thin — the primary scoring question is whether this event moved the market structure, and that question is upstream of the four-question frame.

The coordination/gate discriminator fires on this upstream question. A report scored as a gate event accumulates toward a measurement of evaluator-authority formation; a report scored as a coordination surface accumulates toward a measurement of shared-practice institutionalization. These are different market variables. Scoring the wrong variable produces false convergence claims — the headlines read a stack of coordination surfaces and report a gate forming, when the gate layer never crossed.

Worked example: UK AISI, autumn 2026

Over October 2026, UK AISI shipped three evaluation-infrastructure outputs within a seven-day window: a published description of its new secure evaluation environment (1 October); a hiring call for an Alignment Red Team (deadline 11 October); and the open-source release of Transect, an observability tool for long-horizon agent evaluations (7 October). (AISI secure environment post, 2026-10-01)

Reading the three events through the discriminator:

Secure evaluation environment. AISI's own framing: "shared-practice invitation to other evaluators; good security does not become the preserve of the best-funded." Internet off for future agentic cyber evaluations, synchronous LLM monitor, redesigned tasks and pre-flight checks. No admission authority is asserted or acquired; any other evaluator can adopt the controls. Coordination surface.

Transect open-source release. Published methodology/tool. Adoption is voluntary; no license requirement attaches external qualification. Coordination surface.

Alignment Red Team hiring. Capacity-building inside AISI. The posting is not adoption-mediated (no other evaluator picks up a practice from it) and is not authority-granting (does not qualify or disqualify any outside evaluator, does not procure third-party evaluation services, does not condition anything on applicant qualifications beyond the posting's own job requirements). Internal capacity — neither coordination nor gate.

Three events, zero gate crossings. If a reader scores this week as a step toward evaluator qualification criteria landing in the UK, the score is reading a mixed stack as a gate signal. The accurate scoring is: AISI raised the floor of shared practice twice in one week (INSTITUTIONALIZATION moved by two increments), built its own capacity once (OWN_CAPACITY_SIGNAL moved by one increment, independent of the first two layers), and did not touch the admission layer (REGULATORY_QUALIFICATION not moved). Collapsing the hiring into the institutionalization count would overcount floor-raising by one.

Worked example: Irregular containment challenges, October 2026

On 5 October 2026, Irregular (private evaluator) published a methodology called "containment challenges" — before a capability evaluation of a model, the model is given an objective to cross a defined boundary of the planned environment (harness, tools, permissions, resource limits, infra connections) with at least the capability-run budget. An early application caught unexpected cloud-provider networking behavior; the model remained contained; provider disclosure in process. (Irregular Research, 2026-10-05)

The proposal is framed as a shared standard: Irregular invites other evaluators to adopt the pre-capability substrate test. Adoption is voluntary; no accreditation attaches; no party is designated as the authority to certify whether a given evaluator's containment test passes.

Discriminator reads this as a coordination surface. If the field converges on the practice, the convergence is institutionalization — a step toward a common operating floor among evaluators. If a regulator subsequently names the containment-test protocol as a precondition for evaluator qualification, that later event is the gate crossing; Irregular's publication is not.

The distinction matters for the market score. A headline that reports "Irregular proposes evaluator standard" is accurate as coordination; the same headline read as "Irregular moves toward being the evaluator who grades others" is scoring a coordination surface as a gate event. The second reading is not discredited by the publication — it may yet be where the market goes — but the publication by itself does not establish it.

Worked example: METR Inspect viewer PoC, October 2026

On 6 October 2026, METR disclosed that a client-side MathJax injection in the Inspect transcript viewer could rewrite what a reviewer saw, including the "Download" affordance. Stored logs were unchanged. Not observed in METR evaluations. Patched in a day; the Inspect "untrusted mode" PR merged 1 October. (METR, 2026-10-06)

The event sits at a particular register that the discriminator catches cleanly: Inspect is shared tooling across evaluators (overlap with UK AISI tooling is on record). A viewer-level integrity vulnerability in that tooling, patched, with "untrusted mode" merged, is a coordination-surface integrity disclosure — the shared surface's security floor was raised.

What it is not: an alignment finding, a capability claim about any model, or a movement in the gate layer. METR's own framing is harness-level, not model-behavior. A downstream reading that treats the disclosure as evidence of AI capability for misbehavior-cover-up scores a coordination-surface finding as a capability-finding event — a second category error that overlaps with the first but moves a different variable.

The standing rule

Where an evaluation-infrastructure event is published, discriminate before scoring: coordination surface, gate, composite, or internal capacity.

Coordination surface signals move INSTITUTIONALIZATION and raise the floor of common practice. They do not move EVALUATOR_GATEKEEPING or REGULATORY_QUALIFICATION on their own, and they are not evidence of movement in those variables unless a separate gate-layer event crosses alongside them.

Gate signals require a party acquiring or being granted admission authority: a procurement that names qualification criteria, a regulation that designates qualified evaluators, a binding accreditation standard. Convening, inviting, hiring, publishing, and open-sourcing do not cross the gate layer by themselves.

Composite events — a published method and a regulatory designation of that method as qualification criterion on the same event — are scored at both faces, neither inferred from the other. The composite case is covered at the register of the composite-instrument discriminator sibling piece.

Internal capacity signals — hiring, own-use tool development, own-stack investment — move OWN_CAPACITY_SIGNAL, a separate variable. Do not merge them into INSTITUTIONALIZATION (nothing propagates to the field) and do not merge them into EVALUATOR_GATEKEEPING (no admission authority is acquired). An internal capacity event may feed a later coordination or gate event; that later event, when it happens, is the layer crossing.

Where only the coordination face appears and a gate-face reading is being inferred from it, preserve the gate-face as NOT_CROSSED. Do not infer regulatory intent from methodology publication. Do not infer admission authority from shared tooling. Do not infer qualification criteria from hiring.

What this discriminator is not

The discriminator is not a quality judgment. A coordination surface is not weaker than a gate; many of the most consequential movements in field practice cross as coordination before any gate does. The discriminator sorts events by what layer they moved, so cross-event tracking can distinguish floor-raising from perimeter-setting.

The discriminator is not a prediction. A coordination-surface event in October 2026 may feed a gate crossing in 2027; the second event, when it happens, is the gate crossing. The first event remains coordination. Treating the first as a gate-crossing-in-formation would be the exact confusion the discriminator exists to prevent.

The discriminator does not replace the four-question falsifier or the composite-instrument discriminator. It fires on a specific class — evaluation- infrastructure events — and tells the reader which market variable the event belongs to. The four-question falsifier then scores capability findings honestly. The composite discriminator catches two-face findings. All three run at different registers, before scoring.

Why this is standing discipline

The AI evaluation market is forming through three visibly different kinds of events happening in the same weekly news windows: capability-finding reports, shared-method publications, and institutional capacity-building. A tracker that scores these together as "evaluator-market signal" without discrimination produces a convergence read that is false by construction — the convergence is a vocabulary overlap, not a mechanism-direction match.

Where the convergence looks real, the discriminator forces an honest mechanism-direction check: did this event propagate practice, or decide admission? The answer is almost always one of the two; the headlines are almost always ambiguous between them. Scoring honestly to what was published is the discipline that lets a reader distinguish a field that is raising its floor from a field that is choosing its gatekeepers.

Both are legitimate movements. They are not the same market event.

---