The Steward Speaks. The Steward Sees. The System Reads.

The Human and the Agent Are Fast in Opposite Directions left one question standing on purpose. It established that every direction of a Steward-system exchange carries a channel mismatch, that the mismatch is structural rather than a tooling accident, and that the side which adapts in release cycles is the side that gets redesigned. It did not say what the redesign consists of. Steward Experience mandates that the Steward's surface be held to the rigor of customer-facing product, and it has never said what shape that surface should take. This memo supplies the shape, and the shape is a rule rather than a screen.

The interface discipline that resolves Bandwidth Inversion by routing each channel between a human Steward and an agentic system to whichever medium is highest-bandwidth for the receiver — in the abstract, an interface that translates between two opposite efficiency profiles rather than standardising on one; in its current form, voice for the Steward's commands, visual state for the Steward's verification, and structured text on the ledger for the system itself.

That is the Transcoding Layer. The definition has two halves for a reason. The first half is the discipline and is meant to hold for as long as Bandwidth Inversion holds. The second half is the assignment that discipline produces today and is expected to age. Keeping them apart is what makes the term usable after the tooling changes.

Three routes, one rule

Every exchange between a Steward and the system they govern runs in one of three directions. The Steward commands the system: intent flows from the human to the agents. The Steward verifies what the system did: state flows from the agents to the human. The system records for itself and for other agents: state flows from agents to agents, and no human is the addressee. Command, verification, record. Under the Stewardship Model the Steward governs the system and handles its exceptions rather than executing its routine work, and those three routes exhaust the Steward's contact with the system at that boundary. They do not exhaust the stack — Agent Council and Layered Stewardship sit between Execution and Judgment — and they do not multiply: an escalation is a verification followed by a command, a clarification is a command answered on the verification route, a delegation is a command. Each is an instance of a route, not a fourth one.

The rule assigns each route to the channel that is highest-bandwidth for whoever is receiving. Highest-bandwidth here carries the scope #125 gave it: the medium through which the receiver takes in or expresses what the governance task requires in the least time, for that receiver and that task, not a figure in bits per second. Not for whoever is sending, and not for whoever built the surface. The word receiver is the whole discipline. A route is designed correctly when the party on its far end takes in the content as fast as that party can take anything in. Everything else — which medium, which device, which rendering — follows from asking who receives, and changes when the answer changes.

That is why the Transcoding Layer is a discipline and not an interface. An interface is a specific artefact and dates with the hardware that renders it. A rule that says route by receiver survives every generation of hardware, because it is a statement about where the translation work sits, not about what the translation produces.

The assignment, as of now

Applying the rule to the three routes with the parties as they stand today gives the second half of the definition.

The command route runs by voice. The Steward is the sender and speech is their fastest output; the system is the receiver and text is its fastest input. The mismatch is resolved on the system side, where the spoken command is converted into structured intent before any agent acts on it. The Steward speaks; the system does the work of making speech legible to itself. On both human-facing routes, that is the direction the translation burden runs.

The verification route runs by visual state. The system is the sender and structured text is its native output; the Steward is the receiver and vision is their fastest input. Again the system carries the conversion: it renders state as something to be scanned rather than read — the changed cell, the red row, the shape of a trend — because the party receiving is a human and a human confirms by seeing. The Steward sees; the system does the work of making state visible.

The record route runs by structured text: the Deterministic Logging and Proof of Action trail that every decision writes, and that the Operational Ledger compounds into a queryable knowledge asset. Here both sender and receiver are agents, so the native medium is right for both and no transcoding is needed. The route exists in the rule because it has to be named as the one place the Steward is not the receiver — which is the answer to why exposing the trail to the Steward fails. The trail is written and read by agents at agent tempo; the ledger is what it becomes; the Audit Surface is what a human is shown of either. Handing the trail itself to the Steward as if it were the verification route is the error The Steward Cannot Govern What They Cannot See measured and #125 explained.

The three assignments are a reading of the parties as they are in 2026. Models that take in speech without an intermediate text step, or that render state at a tempo no dashboard yet matches, will shift the concrete form. They will not shift the rule, because the Steward's fastest input will still be vision and the Steward's fastest output will still be speech.

The mirror this corpus already built

Arco has applied the receiver rule once before, in the other direction, without naming it as a rule.

A structured, API-first interaction layer that allows external autonomous agents to discover, evaluate, and transact with a business's services without a human intermediary — built specifically for agent inference, not developer integration.

That is the Machine-Readable Interface. Read its last clause with this memo in mind: built for agent inference, not developer integration. The MRI is designed for the party that receives it — an agent — and not for the party that builds it. That is the receiver-first discipline, applied outward, before it had a name. The Transcoding Layer applies the same discipline pointed the other way. The MRI makes a business legible to agents; the Transcoding Layer makes an agentic system legible to a Steward. Both sit on the system side of the boundary. Both are defined by their receiver. The corpus had the discipline before it had the name, and it applied it outward first because agent buyers arrived before Stewards needed to govern at market tempo.

The Declaration Layer sharpens the mirror rather than complicating it. It handles the semantic half of machine readability — telling an agent what a site's vocabulary means before the agent reads any content — while the MRI handles the transactional half. Together they address the semantic and transactional sides of the agent's legibility problem, and the shape of that answer says something about why the human's answer looks different. The agent's legibility problem is structural: it needs schema, endpoints, and anchored vocabulary, because an agent that meets an unexpected structure will resolve the mismatch on its own and produce a hallucinated result. The human's legibility problem is not structural. A Steward does not need a manifest to know what a red row means. The human's problem is bandwidth — the wrong channel for the receiver — and so the human-facing remedy is a routing rule, not a schema. Two different failure modes at the two boundaries, two different disciplines to close them.

Where the rule is load-bearing

Two of the three routes can be got wrong at some cost. One cannot. A command route that uses typing instead of voice is slow, and The UI That Runs the Business already recorded that slow prompting tools inflate intervention time. A record route that a Steward is made to read is tedious and eventually ignored. The verification route is different in kind, because when it fails the Steward does not become slow — the Steward becomes blind.

Practitioner guides have recorded this without stating why. One notes that a screen holds the user's state for them where a spoken sentence cannot be read twice; another, that a visual surface can be scanned and a mistake recovered from while speech is gone once heard. Both are the verification route failing in the field: the receiver handed the wrong channel for the one route where seeing is the function.

A Steward who commands by voice and receives confirmation only by voice cannot check what the system did at the tempo the system does it. Under the Physical Intervention Threshold that is a safety condition, not an inconvenience: the escalations with the highest Rollback Cost are the ones that require the Steward to see current state and answer before the action proceeds. The Steward's Blind Spot described what happens next — disengagement, and a Nominal MTTI reading that rises because the human stopped looking rather than because the system stopped needing them. The visual verification route is the part of the Transcoding Layer that prevents Steward disengagement from masquerading as autonomy. The next memo takes that argument on its own.

A surface that defaulted

The rule gives operators a test that can be run before any screen is designed and against any surface that already exists. Count the media. If command, verification, and record all run through the same medium — a dashboard of tables the Steward reads, a console the Steward types into, a log the Steward is expected to follow — the surface has not been designed. It has defaulted to the builder's native medium, and the builder was an engineer who thinks in structured text and shipped the trail as the interface.

A surface that passes uses at least two media on the human-facing routes, and uses them for the reason the rule gives rather than because a channel happened to be available. The Steward speaks to it. The Steward sees what it did. And the record the system writes for itself stays where it belongs, behind the Audit Surface that renders it for a human, not in front of it.

The Operator's Verdict

The Transcoding Layer is not a recommendation to add voice or to build a better dashboard. It is a rule about where translation work sits in a human-agent system: on the system side, at every route, in whatever direction makes the receiver fastest. Voice and visual state are what the rule produces when the receivers are a 2026 human and a 2026 agent. They are the assignment, not the discipline.

An operator who has shipped a Steward surface should be able to say, for each of the three routes, who receives and whether the medium is that party's fastest. If the answer is the same medium three times, the surface was inherited from the trail. If the Steward has been given the agent's native medium for verification, the Steward is being asked to read what they should be shown, and the disengagement that follows is not a training problem.

Technology changes what is possible. The receiver determines the channel.