Appendix G — Field Guide

Six organizations, six evidentiary reconstructions

Part V (§5.5) asks whether engineering structures found in DocAble recur in systems developed independently under different technical and organizational pressures. This appendix presents the evidence behind that comparison.

These are reconstructions, not additional case studies. Public accounts reveal selected mechanisms and design decisions, not the engineering history available for DocAble. They can show that a comparable structure exists, demonstrate another way to realize it, or mark a limit on the MAGE interpretation. They cannot establish causation, and silence in a public source does not establish that an organization lacks a capability (see §6.1.5 and §5.5).

Each organization therefore gets a two-page spread. The verso stays close to the public record: what the organization says it built, the evidence available for that account, and the boundary of what those sources establish. The recto interprets the same evidence through MAGE: Modeling, Alignment, the Governed Engineering Environment, governance conversion, and engineering capital.

Read the diagrams as mappings, not ratings. A strongly marked relation is directly supported by the cited account; a partial relation is a MAGE interpretation. An unmarked region means only that the source does not establish it. No card claims that an organization implements MAGE as a whole.

The six organizations enter from different directions: Cloudflare from institutional policy; Spotify from fleet-scale change; Shopify from organizational knowledge; Docker from delegated runtime authority; Siemens from model-first engineering; and Zenseact from distributed autonomous work. The question is which engineering moves survive those differences.

How to read the recto diagrams

Every recto uses a simplified Figure 0.1-1: an observed engineering pressure enters from the left and passes through the mechanisms supported by the evidence to the part of MAGE it most directly illuminates.

The overlay uses exactly three epistemic states:

The diagrams are neither MAGE audits nor a maturity ordering. Pale means not established, not absent.

Cloudflare

Internet infrastructure · policy-first governance at organizational scale

Verso — Evidence

What the public record shows. Cloudflare describes an engineering-standards system in which requirements originating in RFCs are extracted into structured, machine-readable form. Requirements that can be checked mechanically can become custom checks; obligations requiring semantic judgment remain review questions. This separates the authority to establish policy from the machinery that applies it repeatedly.¹1. Timo Reimann, “How Cloudflare Enforces Engineering Standards Using Ai,” Cloudflare, August 4, 2026, https://blog.cloudflare.com/engineering-standards-enforcement/.

The important move is not AI review itself, but turning policy into an artifact that later work inherits.

Boundary of the evidence. The public account is much stronger on representation and enforcement of obligations than on explicit executable models of the software being governed. Nor does it show repeated failures being converted into new environmental structure over time. The evidence establishes policy → representation → mechanism, but not failure → diagnosis → environmental adaptation.

Portable lesson. A policy decision gains multiplicative reach when the environment can apply it without requiring another person to reconstruct it.

Recto — MAGE Interpretation

MAGE reading. Cloudflare is strong evidence for Alignment, with a modest Modeling move: obligations that once lived in RFC prose acquire structured identities that machinery can consume. Once represented mechanically, one policy decision can govern many later changes.

Interpretive boundary. This is an Alignment-heavy example with a modest Modeling component, not evidence for the complete MAGE loop.

Spotify

Audio streaming · fleet-scale autonomous migration

Verso — Evidence

What the public record shows. Spotify's Honk system moves large-scale code migration from distributed manual implementation toward centrally scoped, fleet-executed work. A migration that could previously involve hundreds of teams over weeks can instead be scoped by one engineer over a few days; tooling identifies and schedules targets, agents execute changes concurrently, and the engineer supervises the fleet and handles exceptions.²2. Max Charas and Marc Bruggmann, “Honk: Autonomous Code Migration at Spotify,” Spotify Engineering, 2026, https://engineering.atspotify.com/.

Spotify's surrounding engineering estate also matters. Its System Model and Backstage catalog carry service identities, ownership, dependencies, endpoints, and lineage, giving fleet tooling a representation through which migrations can be targeted rather than requiring each agent to rediscover the repository landscape.^** Spotify System Model and Backstage service catalog, as recorded in the §5.5 reconstruction (service identities, ownership, dependencies, endpoints, lineage).

Boundary of the evidence. Reported activity increases — including the 76% PR-frequency figure^†† 76% PR-frequency figure as reported in the Honk account; cited here strictly as an activity measure, not as evidence of durable throughput. — measure implementation activity, not durable throughput. The stronger evidence is how the work changes: agents implement while human judgment moves upstream to scoping and supervision.

Portable lesson. Scale implementation; concentrate human judgment on what to change and where.

Recto — MAGE Interpretation

MAGE reading. Spotify shows the model and governed environment working together. A model of the service estate helps determine where work belongs; verification and admission machinery determine whether generated changes can proceed; fleet execution makes one engineer's upstream decisions consequential across many repositories.

Interpretive boundary. The evidence does not show that every Spotify migration is model-driven or automatically admitted; the reading applies only to the mechanisms described publicly.

Shopify

Commerce · shared organizational context for agent work

Verso — Evidence

What the public record shows. Shopify's public account describes agent work occurring in shared, observable channels rather than disappearing into private sessions. The record gives a one-month snapshot of 59,918 sessions across 5,170 channels involving more than 7,000 people, together with the characterization of the system as "multiplayer by construction."³3. Javier Moreno and Burke Libbey, “Under the River,” Shopify Engineering, May 28, 2026, https://shopify.engineering/under-the-river. Shared sessions become searchable organizational artifacts; repeated knowledge can then be extracted into skills, conventions, and defaults that later agents can retrieve. This is a soft form of externalization: retrievable prose and skills rather than a structured executable system model.

Boundary of the evidence. The evidence establishes session → reusable knowledge, but not session → skill → engineering capital. MAGE supplies the further interpretation that recurring judgment, once captured in reusable skills, may become engineering capital.

Portable lesson. Experience cannot compound organizationally while it stays private. Make work observable before trying to make its lessons reusable.

Recto — MAGE Interpretation

MAGE reading. Shopify sharpens a condition on engineering capital: experience must become observable before it can become durable. Shared sessions can be searched, corrected, mined, and — where a lesson deserves reuse — converted into skills or defaults. This is Modeling at the knowledge end of the spectrum, not the executable-system-model end.

Interpretive boundary. Observability makes compounding possible; it does not make every observed session engineering capital.

Docker

Developer tooling · delegated autonomy with explicit authority boundaries

Verso — Evidence

What the public record shows. Docker's system separates delegated roles, tool permissions, implementation, and evaluation. The reconstruction identifies seven permission-backed roles and an independent reviewer using a separate model invocation, while deterministic tests supply repeatable evidence. Final admission remains human: the system deliberately does not mechanize every consequential judgment.⁴4. Manuel de la Peña, “A Virtual Agent Team at Docker,” Docker, 2026, https://www.docker.com/blog/.

Docker also resists a crude reading of MAGE as "automate every gate." A human admission boundary may be exactly right where a decision stays semantic or consequential.

Boundary of the evidence. The public account covers the shortest period of the six, so it cannot establish governance conversion over time or returns on engineering capital. It does establish bounded authority, separation of generation from evaluation, deterministic evidence where appropriate, and deliberately retained human judgment.

Portable lesson. Mechanize what is adequately decidable; retaining human authority over the residue can be an engineered boundary rather than unfinished automation.

Recto — MAGE Interpretation

MAGE reading. Docker shows that strong Alignment need not depend on rich explicit system models. Permissions can bound action, tests can produce repeatable evidence, and an independent reviewer can separate generation from evaluation. Where the remaining admission decision is judgment-laden, Docker keeps a person in authority. That is not failed Alignment; it is an explicit decision to retain human judgment.

Interpretive boundary. Docker supports the Alignment reading, not governance conversion. The unhighlighted Modeling and conversion regions mean not established, not absent.

Siemens

Industrial and model-based engineering · persistent models as development surfaces

Verso — Evidence

What the public record shows. Siemens provides a comparison from outside software-first engineering. Persistent engineering representations carry requirements, behavioral structure, simulation, traceability, and other system semantics; software generation can occur downstream of those representations. The example establishes that the strong end of MAGE's Modeling is not novel: rich, executable representations are already ordinary practice in model-based engineering.⁵5. Siemens Digital Industries Software, “A3e: Autonomous, Agentic Assistance for Engineering,” Siemens Digital Industries Software, 2026, https://blogs.sw.siemens.com/.

SysML and CAD models, simulation, virtual ECUs, and bill-of-materials structures can serve as engineering surfaces, with code and verification downstream.⁵

Boundary of the evidence. The sources examined here do not establish a general admission mechanism that rejects changes when declared model↔implementation correspondence is violated. As throughout this appendix, source silence does not imply organizational absence.⁵

Portable lesson. Software need not be the primary surface of engineering reasoning. Mature engineering disciplines routinely place richer representations upstream of realization.

Recto — MAGE Interpretation

MAGE reading. Siemens shows the strong end of Modeling in established engineering practice: engineers reason through persistent semantic representations and treat implementation as downstream realization. MAGE did not invent model-first engineering; its contribution is to bring that engineering instinct into agentic software engineering and connect representation explicitly to the authority surfaces of the governed environment.

Interpretive boundary. The missing model↔code admission gate means not established, not absent.

Zenseact

Autonomous-driving software · centralized substrate, distributed domain ownership

Verso — Evidence

What the public record shows. Zenseact describes an organizational split. A platform team owns common substrate — authentication, sessions, execution, and safety — while domain teams retain their own agents, tools, instructions, and domain expertise. The arrangement expresses organizationally a principle seen technically elsewhere: centralize mechanisms that should be shared; leave domain judgment with the people who hold the domain knowledge.⁶6. Philip Dufwa and Thomas Luvö, “A Platform for Scalable Enterprise AI Agents,” Zenseact, May 29, 2026, https://zenseact.com/news/a-platform-for-scalable-enterprise-ai-agents/.

Zenseact also illustrates selective context. Routing machinery selects task-relevant expertise instead of loading the whole organizational corpus into every reasoning window. A central runtime and safety envelope coexist with domain expertise captured in reusable skills and surfaced on demand.⁶

Boundary of the evidence. The sources show how Zenseact divides the platform, routes context, delegates work, and assigns ownership. They do not establish the resulting productivity, safety, or returns on engineering capital.

Portable lesson. Build the common substrate once; distribute the engineering judgment to the teams that possess the knowledge it requires.

Recto — MAGE Interpretation

MAGE reading. Zenseact combines Modeling and Alignment organizationally. Externalized domain knowledge and selective context help finite reasoners see what a task requires; shared authentication, execution, and safety machinery place common authority in one substrate. The organization does not centralize the judgment itself — domain teams keep the expertise that defines their agents and tools. The result is an organizational realization of the Governed Engineering Environment: common mechanisms without a central knowledge bottleneck.

Interpretive boundary. The platform/domain split illustrates the Governed Engineering Environment; the evidence does not establish a closed governance-conversion loop or complete Modeling and Alignment.