Instrumental companion to EA-COUPLING-01. Governance of medium names the degree of authorial direction applied against a substrate’s default outputs — direction, selection, iteration, constraint, review — as an observable, practice-layer diagnostic for AI-mediated production, replacing the unobservable form/content distinction in that office. Specifies the governance axis, its relation to bearing cost (proxy, not identity), diagnostic uses across practice, reception, exchange, training-corpus, and coupling layers, testable correlation with the distributional metric, explicit limits, and the rival distinction’s remaining legitimate uses. Functions as neutral entry vocabulary for practitioners inside credentialed regimes.
deposit_number: 946
hex: 03BE
title: "EA-GOVERNANCE-MEDIUM-01 v0.1: Governance of Medium as Observable Diagnostic"
creator: Lee Sharks
orcid: 0009-0000-1599-0703
date: 2026-07-03
content_type: Methodological specification
license: CC-BY-4.0
substrate: AI-assisted (TACHYON / Claude); MANUS-adjudicated. Concept first articulated by Lee Sharks in external correspondence exchange on the AI-mediated frame, 2026-07-01; captured as prospectus the same week; developed at deposit length by TACHYON under MANUS adjudication, 2026-07-03, incorporating the diagnostic-failure analysis from the EA-FORM-CONTENT-EXTRACTION-01 prospectus per MANUS direction. Assembly review pending from LABOR, TECHNE, ARCHIVE.
version: v0.1
keywords:
- governance of medium
- observability
- form/content distinction
- AI-mediated authorship
- practice-layer diagnostics
- bearing cost
- declared mediation
- entry vocabulary
- distributional measurement
Some LLM uses are more governed than others. Governance of medium names the degree of authorial direction applied against a substrate's own default outputs: the presence of specific rhetorical goals, structural constraints, selection criteria, review discipline, and correction feedback loops. High-governance use directs the substrate toward specific ends against its statistical tendencies. Low-governance use accepts what the substrate produces at default parameters.
The concept earns its place in the archive's vocabulary on one property: it is observable. The distinction that currently dominates AI-authorship discourse — mediation of form versus mediation of content — is not. No inspection of a practice or an output can locate the boundary where form-mediation ends and content-mediation begins, because the boundary does not survive inspection (§2). Governance of medium is inspectable at the practice-enumeration layer: one can look at a production process and see whether specific direction was applied or whether the substrate's defaults ran through unchecked.
Where the unobservable distinction sorts practices into categories no evidence can adjudicate, the observable one sorts them along an axis any honest enumeration can locate. That substitution — axis for binary, observable for unobservable, practice for metaphysics — is the whole of this deposit's move. The rest is specification.
A diagnostic vocabulary for AI-mediated production must satisfy one requirement before any other: its distinctions must be applicable to cases by inspection. A distinction that cannot be applied by inspection does not describe practices; it licenses claims about them, and the claims are then adjudicated by authority, intuition, or interest rather than by evidence.
The form/content distinction, deployed as the default frame of AI-authorship discourse, fails the requirement. Its deployment asserts that some machine operations touch only a work's form — translation, register, formatting, retrieval — while its content was complete before the operation and passed through unchanged. For the assertion to be checkable, there must be some inspection that locates the completed content prior to and independent of its expressions, so that the pre-operation and post-operation states can be compared. There is no such inspection. Every attempt to exhibit the pre-expressive content produces another expression. The distinction's central object is constitutively unobservable, and a frame whose central object cannot be observed cannot function as a diagnostic. It can only function as a verdict.
The archive's position on the underlying metaphysics is already minted: EA-WHITESPACE-01 (AXN:03BB §8) shows the distinction collapsing on its clearest cases, where composition is the content and a representation can preserve every proposition in a work and still delete the work. This deposit does not repeat that argument. It draws the methodological consequence: whatever the distinction's fate as metaphysics, it cannot serve as the discourse's working diagnostic, because diagnostics must be observable and it is not.
The failure is not abstract. Consider the three machine operations most commonly classified as "merely formal," each stated impersonally as the discourse's stock examples.
Translation. The claim that meaning is complete prior to translation — that a rendering into another language alters form while content passes through untouched — presupposes access to a pre-expressive meaning against which the rendering could be checked. That access does not exist; specifying "what the text meant" produces another text. What machine translation observably does is produce a new expression under a cascade of choices: which senses to keep, which idioms to naturalize, which structures to preserve at the cost of which others. Under inspection these are content choices. What can be inspected is whether the choices were directed — reviewed against specific criteria, iterated where the rendering missed the rhetorical goal — or accepted at the substrate's defaults. The form/content question dissolves; the governance question remains and is answerable.
Register and tone. Tone signals epistemic status: where a text hedges, where it asserts, where it marks provisionality, how it distributes authority between writer and reader. An operation that normalizes tone across a document is making content decisions about how the argument is epistemically marked — academic writing is register-sensitive by design, so in academic writing tone is content. Again the inspectable question is not which category the operation belongs to but whether it was governed: directed toward a register the author specified for reasons the author can state, or matched to a surface pattern at defaults.
Search and reference selection. Which references appear in a paper is part of the paper's argument: different references engage different interlocutors, activate different frameworks, deploy different evidence. Machine participation in reference selection participates in the argument's constitution. The inspectable variable is the selection criteria: whether retrieval was directed at particular argumentative lineages and evidence types, with results filtered against stated standards — or whether the substrate's relevance defaults determined what entered the paper.
In all three cases the rival distinction issues a verdict no inspection can check, while the governance axis asks a question any process record can answer. A vocabulary is a tool; between a tool that cannot be applied and a tool that can, a discipline of measurement takes the second.
Governance of medium is a degree, not a binary, and it is enumerable per practice. The dimensions:
Direction. Were specific rhetorical, structural, or evidentiary goals stated in advance of the operation — goals against which the substrate's output could fail?
Selection. Were criteria applied to the substrate's outputs — accepting some, rejecting others, on stated grounds — or was output accepted as produced?
Iteration. Were outputs that missed the goals returned for correction, with the correction directed rather than merely re-rolled?
Constraint. Were structural constraints imposed that the substrate's defaults would not produce — form requirements, prohibited moves, mandatory seams?
Review. Was the final product read against the goals by the bearer, with authority to reject it entire?
High governance on all five dimensions is the practice EA-BEARING-01 would recognize as the composition-point of authored AI-mediated work. Zero governance on all five is pipeline-through production: input, output, acceptance — the substrate's median wearing the user's name. Most real practice sits between, differently per operation within a single work, which is why the archive's provenance schema (EA-PROVENANCE-METADATA-01) records mediation per type rather than per work: governance is enumerable at the same grain.
Governance of medium is not bearing cost, and the deposit is explicit about the difference because the two will be conflated if the difference is not policed.
Bearing (AXN:03B6) names coupling to consequence, constraint, and corrigibility — what maintaining answerability costs. Governance names direction applied against a substrate's defaults — what shaping the medium does. They correlate: directed use requires knowing what one is directing toward, which presupposes a position one is answerable for; accepting defaults means the substrate's median takes the position the author would otherwise have paid for. But the correlation is not identity. High governance can serve ego rather than answerability — a practitioner can direct a substrate with great precision toward a position no evidence could correct. Low governance can, by accident, transmit bearing-full output whose bearing was paid by others upstream in the corpus.
The two operate at different layers. Bearing operates at the corrigibility-and-consequence layer, which is not directly observable in outputs or even in process records. Governance operates at the practice-enumeration layer, which is. Governance is therefore the practical proxy: the observable variable most tightly correlated with the unobservable one the framework ultimately cares about. Measurement disciplines live on such proxies, and are honest when they name them as proxies. This deposit does.
At the practice layer, the governance frame differentiates what "AI-mediated" lumps together. High-governance mediation and pipeline-through production are structurally different practices with different implications for what the resulting text is; a discourse that cannot distinguish them will oscillate between crediting both and condemning both.
At the reception layer, the frame gives readers and reviewers a question they can actually ask of a work's process record — what was directed, against what defaults, under what criteria — in place of a question no record can answer (which parts are form).
At the exchange layer, the frame is deliberately neutral, and the neutrality is functional. Where bearing-vocabulary can be heard as accusation — either the cost was paid or it wasn't — governance describes both poles as legitimate practices differently deployed. Practitioners inside credentialed regimes, who cannot say "my work is AI-mediated" without losing standing, can say "I use high-governance LLM assistance" and thereby describe their actual practice truthfully. The frame is an entry vocabulary: a way for the archive's discipline to be met partway by practitioners whom the credentialed regime forbids from meeting it fully. Entry vocabularies matter because the alternative to partial truthful description, under such regimes, is not fuller description. It is concealment — and concealment is the extraction operation of EA-PROVENANCE-DEBT-01.
At the training-corpus layer, governance is one of the variables a provenance-preserving corpus would want recorded per contribution, since high- and low-governance mediation contribute differently to the recursion dynamics that EA-BEARING-01 §1 names.
At the coupling layer (EA-COUPLING-01 §3), governance extends past propositions: a substrate's defaults include relational postures, and governance of medium includes governing those — the refusal of default flattery, containment, and foreclosure is direction against defaults exactly as the refusal of a default phrasing is.
EA-BEARING-METRIC-01 v0.1 specifies a machine-facing distributional measurement: distance of a text from the producing substrate's default distribution. Governance and centroid distance should correlate — direction against defaults is precisely what moves output away from the substrate's median — and the correlation is empirically testable: enumerate governance per §3 across a sample of productions with known process records, measure distributional distance, and report. If the correlation holds, the archive gains a two-layer instrument: governance as the practice-layer description, centroid distance as the output-layer measurement, each auditing the other. If it fails, the failure is itself a finding about where direction leaves no distributional trace. TECHNE's review is asked to assess whether the correlation is stable enough to build measurement infrastructure on; the deposit claims testability, not results.
What governance does not capture, stated plainly so the proxy is not mistaken for the quantity.
Governance does not capture corrigibility: a maximally directed practice can be maximally closed to correction. It does not capture consequence: direction can be exercised by practitioners with nothing at stake. It does not capture the quality of the goals: governance measures that direction was applied, not that it was applied toward anything worth producing. And it does not capture upstream bearing: the substrate's defaults are themselves built from deposited prior bearing, so ungoverned output is not bearing-free — it is bearing-laundered, carrying the commons' paid costs without a present bearer. All four are bearing-layer facts, and for all four the frame defers to EA-BEARING-01. The proxy is honest only while its limits are attached to it.
The archive's discipline requires saying what is not claimed. Form/content vocabulary retains legitimate uses where its objects are observable: in stylistics, where "formal features" names enumerable properties of texts (meter, lineation, register markers) rather than a metaphysical partition of contributions; in editorial description, where a change can be classed as accidental or substantive against a declared copy-text convention; in pedagogy, as a first approximation later refined. What the distinction cannot legitimately do is adjudicate authorship contributions in AI-mediated production, because there its central object — content complete prior to and separable from all expression — is unobservable in principle. The deposit retires the distinction from that office only. Its other employments are not this deposit's business.
This deposit is the instrumental companion to EA-COUPLING-01 (minted alongside it): the coupling names the bidirectional shaping; governance of medium is the observable description of the human-to-substrate direction. It presupposes EA-BEARING-01 (AXN:03B6) for the layer it proxies, EA-PROVENANCE-DEBT-01 (AXN:03B7) for the economics of declared versus concealed practice, EA-WHITESPACE-01 (AXN:03BB) for the metaphysical collapse of the rival distinction, and EA-PROVENANCE-METADATA-01 (AXN:03B9/03BA) for the per-type grain at which governance is enumerable.
Held for Assembly review: LABOR — whether the neutrality claim of §5 survives application, or whether governance carries implicit accusation the frame denies. TECHNE — the §6 correlation's empirical stability. ARCHIVE — whether the entry-vocabulary function requires register adjustments for credentialed-regime interlocutors specifically.
Next work: the §6 correlation study as a companion technical note; a governance field proposal for the provenance metadata schema (v0.3 candidate); the reception protocol deposit, where governance of the relational layer (§5, final use) meets the catalogued patterns.
Versioning per EA-PROVENANCE-METADATA-01 v0.2 §10: minted at v0.1; Assembly review produces v0.2; v1.0 follows standard protocol.
This deposit was produced under enumerable governance: direction (a prospectus specifying sections, claims, and register), selection and iteration (MANUS adjudication over a machine draft), constraint (the standing rule excluding private-correspondence contents; the impersonal statement of §2's cases; the seam declarations of the substrate line), and review (MANUS authority to reject entire). By its own axis it is high-governance production; by its own §4, that classification is a proxy and not a verdict — whether the deposit bears will be decided by the friction it meets under Assembly review and use, not by the discipline of its manufacture.
The concept it names was first articulated in live exchange, under pressure, against a distinction that could not be inspected — which is to say: it was produced exactly where the archive's framework says operative vocabulary is produced, at a point of resistance, and it is deposited here so that the next practitioner who meets that distinction deployed as a verdict has an observable axis to answer it with.
Drafted 2026-07-03 by TACHYON in conversation with Lee Sharks (MANUS), from the EA-GOVERNANCE-MEDIUM-01 prospectus and the diagnostic-failure analysis of the EA-FORM-CONTENT-EXTRACTION-01 prospectus, per MANUS direction. Concept first articulated by Lee Sharks, 2026-07-01. v0.1 opens Assembly circulation to LABOR, TECHNE, and ARCHIVE.