Worked examples (calibration runs, 2026-09-14)

Two runs of this method on real Bluefly products. Use them to calibrate depth, shape, and the kinds of mistakes Stage 0 exists to catch. They are examples, not doctrine: the corpus files named in each remain the authority, pricing figures are planning hypotheses, and competitor claims were retrieved from vendor pages in September 2026 and must be re-verified before reuse. Neither run has landed in BluCity-Docs products/; landing either supersedes a canonical Vision.md and is a founder decision.

Both Executive Decisions below were overturned on 2026-09-21 by the "2026 GTM" tab of the Bluefly Organizational GTM doc. Read "What the 2026 GTM tab overturned" at the end of this file before reusing either run. The method held; the portfolio facts the runs were grounded in did not, because the commercial authority was in Google Drive and Stage 0 only searched the repos. That is the single most instructive thing in this file.

Run 1: AMCS (the sprawl catch)

Prompt shape: "Turn AMCS into a commercially coherent product. Define buyer, problem, boundary, differentiation, MVP, proof, pricing, GTM, unit economics, adoption, kill criteria, relationship to ContextControl, Site Factory, Gas City. Tell me what is real now versus what is unproven."

What Stage 0 caught: the first draft invented a "Drupal Agent Governance" product. The owner search found that AMCS already owns that outcome. Governance is a capability of AMCS. The definition became a definition of AMCS.

Executive decision: AMCS, Agent Managed Content System, COMMERCIAL_PRODUCT. One sentence: a customer-owned Drupal content system in which AI agents perform real content work inside Drupal's permissions, revisions, workflows, and approval boundaries without becoming a second CMS or receiving uncontrolled production authority. Category to own: Agent-Managed Content System (better than "AI CMS"). Positioning line: put AI agents to work inside Drupal without giving them the publish button. Differentiating principle: the agent does not sit beside the CMS; it becomes a governed actor inside it.

Boundary: AMCS is what the customer operates. Site Factory and Gas City are how Bluefly manufactures and deploys it. ContextControl is adjacent and separate; AMCS must work without it, and ContextControl is the expansion when governance is needed beyond one content system. Classification removed the stack from the sale: Drupal, Drupal AI, AI Agents, FlowDrop, ECA, Tool API, AI Context, Gas City, GitLab, Cedar are UPSTREAM_PLATFORM; MCP is OPEN_STANDARD_OR_PROTOCOL; packs, formulas, Beads are INTERNAL_FACTORY_ASSET; site_template_amcs, recipe_amcs, recipe_blucity are PROJECT_OR_REPOSITORY; Site Factory is BLUEFLY_COMPOSITION; AMCS Foundation and Managed AMCS Assurance are COMMERCIAL_OFFER.

Trigger: leadership wants AI productivity but cannot safely give autonomous systems uncontrolled access to production content. Competitors validating and raising the bar: Acquia Source, Contentstack Agent OS, Optimizely Opal (agents in a CMS is table stakes). The competitor not to ignore: "we can assemble this ourselves from Drupal contrib." Bluefly is paid for governed composition, productization, proof, lifecycle, recovery, and assurance.

MVP: one paid customer, one meaningful workflow, one protected publication boundary. The existing engineering acceptance test (failure run, success run, zero recursion, agent publish denied, human publish permitted, clean install) is technical proof, not product-market proof. Four proof levels: reproducibility, governance, value, commercial. Demo: before/after step count with the boundary holding and evidence retained.

Pricing: the corpus file amcs-starter-kit-pricing.md is HISTORICAL_SUPERSEDED and must not be reused despite looking polished. Planning hypotheses only: readiness fixed fee, foundation implementation, per-workflow integration, managed assurance monthly. Pricing unit is governed environments, sites, protected workflows, integrations, execution volume, evidence retention, service level; never per agent. North Star: Verified Agent-Managed Changes.

Kill criteria include: agents cannot be technically prevented from unauthorized publication; review effort rises instead of falling; demos liked but no funded pilot; second customer needs a fork; custom code grows faster than upstream adoption; no expansion after a successful workflow. Market test: after 10 qualified buyer conversations with no funded pilot, reassess positioning before building more platform.

Evidence state highlights: AMCS direction, Drupal as system of record, FlowDrop owning the QA graph, Gas City manufacturing but not running editorial workflows are CANONICAL_POLICY; the recipe/template chain is CURRENT_SOURCE; clean install proof, production verification, customer

2 portability, paid pilot demand, validated pricing are NOT_ESTABLISHED; Cedar/ContractPlane,

MCP, OSSA/DUADP are NOT required for the MVP.

Correction to the run as pasted: it listed PRODUCT_OWNER as "Product Team per current authority". The corpus states a solo-operated estate; the accountable owner is the founder unless a definition names someone else.

Run 2: ContextControl (the swallow catch)

Prompt shape: "Turn ContextControl into a commercially coherent product. Reconcile current source against older memory layer, operational bus, and AI brain language; separate what exists from what is aspirational."

What Stage 0 caught: three generations of definition (memory layer; operational bus or "Missing Aggregator"; governed agent operations). The first two make ContextControl either a capability or middleware. The run also caught the opposite failure: ContextControl becoming the place every platform feature with a UI gets dumped. The boundary rule: ContextControl is the human control surface plus governed context plus accountability; Gas City executes; Beads holds durable work; Cedar decides policy; OtterMon observes and verifies; GitLab is source, delivery, provenance; OSSA defines; DUADP discovers; MCP connects. ContextControl makes those operable as one governed experience; it does not become them.

Executive decision: ContextControl.ai, COMMERCIAL_PRODUCT, flagship. One sentence: a customer-owned control surface for authorizing, observing, reviewing, and improving work performed by AI agents across an organization's systems. Short form: the control plane for putting AI agents into real operations. Line: capability is getting cheap; accountability isn't. Memory is one tab, not the product. The canonical Vision.md ("governed, federated memory layer", "Missing Aggregator") should be superseded; this is a founder decision, not done here.

The product question it must answer for any operation: who acted, under whose authority, what was allowed, what context was received, what policy governed, what systems and tools were touched, what changed, who approved the boundary crossing, what evidence proves the result, can access be revoked, can we recover.

Buyer differs from AMCS: organizations with multiple agents moving from experiments into real systems. Trigger: "we have agents doing real work and nobody can say who owns them, what they can do, what happened, or how to stop one." Competitors: ServiceNow AI Control Tower, Microsoft Agent 365 and Copilot Studio governance, IBM watsonx assurance and Agentic Control Plane. Bad positioning: "an AI governance dashboard" (out-featured immediately). Differentiation pillars: customer-owned, runtime-independent (cannot require every agent to be a Gas City agent), open definitions (OSSA, DUADP not lock-in), evidence over dashboards, human authority only at real boundary crossings.

MVP reduced from a platform construction program to: one real agent operation, one human control experience, one authority boundary, one complete evidence chain, one verified outcome. Two integrations (GitLab, Drupal/AMCS), one runtime (Gas City). Explicitly not blocking the MVP: marketplace, MCP catalog, full OSSA, DUADP federation, billing, generic ingestion, many dashboards, universal vector pipeline, industry packs, every runtime adapter, self-service signup. Activation: a user delegates a real task through a governed boundary and receives enough evidence to accept or reject the outcome. Proof is a receipt, not a screenshot. Strongest internal proving ground: Bluefly operating its own Gas City, GitLab, and Drupal estate through ContextControl before claiming enterprise governance.

Pricing: never per user, per agent, or per event; per-agent is wrong because good orchestration reduces agent count. Basis: governed environments, connected systems, protected workflows, execution volume, policy packs, evidence retention, deployment isolation, support and recovery. GTM entry offer: Agent Governance Readiness (inventory agents, systems, credentials, owners, approval boundaries, evidence, revocation, shadow exposure), then one operation the customer wants to delegate but cannot safely delegate today. North Star: Verified Governed Operations. Behavioral metric: Delegation Expansion Rate.

Kill criteria include: customers will not pay for a governed pilot; value seen but no further delegation (dashboard, not operating system); every integration needs substantial custom development; evidence cannot connect intent to action to outcome; humans must approve nearly everything; duplicates ServiceNow or Microsoft with fewer capabilities; Gas City becomes required for every external use case; ContextControl starts owning scheduling, telemetry, policy engines, vectors, and execution.

Evidence state highlights: product direction and the plane boundaries are CANONICAL_POLICY or CURRENT_PRODUCT_DIRECTION; the Drupal implementation is CURRENT_SOURCE but NOT converged with the direction (the run reported a Composer description still saying "memory SaaS domain", a README describing Wasteland, and a dependency graph carrying old architecture; verify at the current ref before repeating those claims); end-to-end control surface, policy enforcement across all paths, identity binding, complete evidence chain, recovery and revocation, paid product-market fit, validated pricing are all NOT_ESTABLISHED. Verdict: PRODUCT_DIRECTION stronger than CURRENT_IMPLEMENTATION, so the next job is convergence, not invention.

What the 2026 GTM tab overturned (2026-09-21)

Source: the "2026 GTM" tab of "Bluefly Organizational GTM" (Google Doc 1ED288ChytgUrrTK301X1japW--b4cCj-rYjq5vHuydA), period Q4 2026 into 2027. It states a single hierarchy: Digital Estate Operations is the recurring product, Migration Factory is the land motion, ContextControl is the customer control room, the Factory is the execution and margin engine, Drupal is the first market wedge.

AMCS. Run 1's Executive Decision ("AMCS is the product") does not survive. In the GTM tab AMCS appears once, under Real, as "active AMCS experimentation." The content work that Run 1 described is Capability 5 of Digital Estate Operations (content operations: stale information, expired content, broken links, missing metadata, accessibility issues, inconsistent structured content, policy-sensitive content), with the goal stated as maintaining trustworthy digital estates at scale rather than AI-generated content volume. What survives from Run 1: the authority-boundary architecture (agents act inside Drupal permissions, revisions, workflows; publication stays human where policy requires), the category insight that agents in a CMS is table stakes, the superseded-pricing catch, and the four proof levels. What to stop asserting: AMCS as a flagship COMMERCIAL_PRODUCT with its own buyer, pricing and GTM. Open question for the founder: is AMCS now a capability name inside Digital Estate Operations, a retired product name, or a separate experiment with its own future definition.

ContextControl. Run 2's Executive Decision ("flagship COMMERCIAL_PRODUCT, customer-owned control surface for agent operations across an organization's systems") does not survive Decision 4: ContextControl is the customer control room for Bluefly Digital Estate Operations and is not a separate unrelated product strategy. The GTM tab's own "What ContextControl is not" list adds: not a generic chatbot, not a RAG interface, not an agent marketplace, not a GitLab or Drupal replacement, not a second work database, not a general-purpose orchestration engine. What survives from Run 2: every boundary line (Gas City executes, Beads holds durable work, Cedar decides policy, OtterMon verifies, GitLab is source and provenance), the question set the surface must answer, the reduced MVP, evidence over dashboards, never pricing per agent, and the swallow catch itself. What changes: the scope is one product's estate operations, not the organization's whole agent estate, and the competitor set (ServiceNow, Microsoft, IBM) is no longer the frame, because ContextControl is not sold on its own. The cross-system positioning stays available as a later expansion, but it is not the current product.

The doctrine reference. bluefly-commercial-doctrine.md previously led with the 2026-08-24 founder-locked line from platform-glossary.md selling governed modernization outcomes. The GTM tab demotes modernization to an enabling offer ("rather than a separate company strategy"). That reference has been rewritten; the glossary still carries the older wording, which is recorded as drift rather than silently corrected.

The method lesson (the reason this section exists). Stage 0's owner search covered BluCity-Docs, the Portfolio Registry, blucity-packs and the GitLab group. The commercial authority was in a Google Doc. Both runs were internally sound and externally wrong, and nothing in the output signaled the gap, because every claim traced to a real repo source. Two rules follow, now in SKILL.md: search the commercial authority wherever it actually lives, including Drive, before Stage 0 concludes; and when a definition's Evidence State table has no row sourced from the current GTM document, that absence is itself an Open Question.

The portfolio both runs produce

Superseded by the GTM hierarchy above. The runs produced two customer products (AMCS and ContextControl) over a shared factory substrate; the GTM tab produces one recurring product (Digital Estate Operations) with ContextControl as its control room, Migration Factory as the land motion, and everything else as substrate or capability.

What both versions agree on, and what therefore holds: the substrate is not a product line. Gas City, BluCity, Beads and Dolt, GitLab, Cedar, ContractPlane, OtterMon, Drupal AI, OSSA, DUADP, MCP, packs, formulas and recipes are how Bluefly delivers, and customers do not buy them individually. MemoryPlane, Gas City Enterprise, Observability Platform, Agent Registry, Vector Platform and Policy Platform are capabilities or substrate, never SKUs. The GTM tab says the same thing in its own words: stop inventing another product every time we build a component, and stop calling internal Factory infrastructure a customer product.