The engine that makes the closed loop close
The closed loop described in what Growyn is can call itself closed only because every section reads from and writes to a shared substrate. Content does not know what brand voice to use until Intelligence has populated brand. Ads do not know what audience to target until Audience has been resolved. Optimize does not know what to optimize for until Analyze has measured. Without this substrate, fourteen sections become fourteen point tools that happen to share a UI; the closed loop becomes marketing copy, not architecture.
The Brain is the substrate. It is infrastructure, not a section the user opens. Every AI call across every section enters through one function (assembleContext) which reads from a structured fact store, resolves conflicts deterministically across five truth tiers, applies time-decay weighting per domain, compresses to a token budget appropriate for the routed model, and logs the assembled package for replay and audit. The model itself never resolves conflicts, never chooses its own context, never decides what the strategic objective is. Those decisions happen above the model, in code that can be tested. Below is the architecture that makes this operationally true.
assembleContext() is the single auditable entry point. Loads the decision-type manifest, runs the truth cascade across five sources, applies continuous time-decay per domain (brand voice decays slowly at ~350-day half-life; ad creative decays quickly at ~70-day half-life), pulls Fingerprint and platform learnings, compresses to the routed model's token budget, logs the full assembly to ai_context_log with a content hash.org_id, contact_id, persona_id, channel, attribution chain, and ai_context_hash linking generated artifacts back to the AI call that produced them. Retroactive enrichment is banned; resolvable identifiers must populate at insert.Inside the Context Assembler
Three sub-mechanisms run inside Layer 3, the Context Assembler. The first is the truth cascade: five sources can each provide a claim about any of sixteen knowledge domains (brand, audience, channels, content, ads, pipeline, plus ten more covering financial, demand, competition, painPoints, website, email, ai_perception, ai_visibility, positioning_gaps, and source_influence). The cascade resolves which source wins per key (not per domain), so within the same domain different facts can resolve to different sources. The next diagram documents the cascade in full.
The second is the decision-type manifest. Each decision type (content.generate_post, nurture.generate_email, ads.generate_creative, engage.respond, analyze.generate_insight, optimize.generate_recommendation, plus lower-frequency types) is registered with a declarative manifest stating exactly which domains it reads, which derived facts it queries, which Fingerprint categories it consumes, which platform learnings it pulls, the routed model tier, and the token budget. The system answers the question what did the model know when it generated this by reading the manifest plus the logged assembly. There is no other path.
The third is the audit trail. Every assembled package gets a content hash and a flat array of decision factors: every pattern, every learning, every fact that influenced this generation, tagged by truth-cascade source and post-decay confidence. The hash makes generations replayable. The decision factors make outcomes attributable: when an A/B test resolves, the system can correlate which patterns were present in winners versus losers. Six months at full fidelity, then the package JSON nulls but the hash and metadata persist indefinitely. This is the technical analog to the protocol's Proof-of-Verification: the assembly is recorded, the inputs are traceable, the output can be replayed.
The Meta-Brain compounds intelligence across organizations
A new customer joining Growyn in year three should not have to start cold. They should benefit on day one from patterns validated across the platform's prior customers. The Meta-Brain is what makes that operationally true. The mechanism is active injection, not passive benchmarking: patterns validated across multiple organizations get promoted to platform_fingerprint status and injected into every eligible organization's assembleContext for matching segments.
k-anonymity governs what publishes. Patterns must accumulate samples from a minimum number of contributing organizations before they can be injected into any other organization's Brain. Two during the current portfolio phase (the four in-house orgs: Growyn dogfooding itself, EDMA Trade for B2B services, EDMA Network for crypto presale, AFU for agriculture cooperative), ten in early external phase, twenty in mature external phase. The threshold is configured per aggregation; revenue and customer-count patterns carry higher minimums than engagement-rate patterns. Segment matching ensures relevance through six dimensions (industry, company size, funnel phase, market, audience type, monetization model). A pattern tagged for crypto-presale token-monetization 50-250-person organizations in the foundation phase only injects into other crypto-presale token-monetization 50-250-person organizations in the foundation phase.
Transparency is the policy, not a feature. When a platform learning enters a generation, the assembled prompt's === PLATFORM KNOWLEDGE === header tells the model (and the user, when they view the assembly) that the learning is drawn from cross-org data and names the contributing org count and matched segments. The user always knows when they are benefiting from the network and when they are operating on their own observed and learned patterns. The Meta-Brain's value compounds with platform scale; year-one customers' learnings become year-three customers' day-one defaults.
platform_fingerprint status and injected into every eligible organization's assembleContext. The mechanism is active injection, not passive benchmarking. Currently four in-house orgs across four verticals (Growyn dogfooding itself, EDMA Trade, EDMA Network presale, AFU agriculture cooperative).platform_fingerprint rowsplatformEligible in the aggregation registry. Groups by six segment dimensions (industry, company size, funnel phase, market, audience type, monetization model). Computes effect sizes at the platform level. Runs weekly.experiment_outcomes with effect size and p-value.active in platform_fingerprint with segment keys, post-validation confidence, contributing org count, sample count. Gated on k-anonymity, segment narrowness, statistical significance.platform_fingerprint for active entries whose segments match. Matched entries flow through the cascade at T4 (above intelligence baseline, below observed and learned). Transparent in the final prompt under a === PLATFORM KNOWLEDGE === header.retired. Historical period preserved for analysis but stops being injected. The platform does not accumulate stale defaults.Continue exploring the EDMA Launchpad
The third Growyn page in this section, Used by EDMA itself, documents the dogfooding case where EDMA Group LLC operates Growyn as an internal organization and feeds the four-org Meta-Brain through its own marketing workflows. The dogfooding is the basis for the structural credibility claim: the same system that markets EDMA's projects to the world is the system Growyn customers use, with the same Brain, the same truth cascade, the same audit trail.
For the broader EDMA Launchpad context, see what Growyn is for the product-level overview, the PoV Gate for the protocol-level analog of Brain auditability (the consensus that admits a claim to chain state only after verification), and the death scale for the diagnostic study that motivated this architecture in the first place.
