Engineering a Hybrid AI Product Generator

1. Research objective and operating thesis

PFD is developing an agent-native creative company: a company in which routine planning, production, validation, packaging, publication preparation, monitoring and reporting can be executed by software agents, automation workers and trained model adapters. The long-term objective is a zero-traditional-employee operating model for repeatable operational work. This is an organizational design target, not a claim that machines are legal persons or that responsibility disappears. Owners and designated reviewers remain accountable for rights, finance, safety, exceptional decisions and deployment approval.

The technical question is therefore broader than “Which model can draw an image?” It is: How can a small, auditable system repeatedly transform governed business context into unique digital products, reject unsafe or repetitive outputs, minimize paid inference, learn from outcomes and preserve a verifiable chain of decisions?

2. System boundary: what PFD owns

The PFD system is intentionally hybrid. Foundation-model training is capital intensive and is not the first differentiator for this product. PFD uses qualified external text, vision and image-generation models where they are economically appropriate. The proprietary layer is the control system around those models: data governance, Style DNA, concept planning, retrieval, prompt compilation, routing, novelty measurement, visual QA, cost policy, packaging, model registry and feedback memory.

This separation is important. A model provider may improve or deprecate a model; PFD should be able to change that provider without losing its creative memory, acceptance criteria or production history. The durable product is therefore the workflow plus governed data plus evaluation, not a hard-coded dependency on one API.

3. Hybrid architecture

4. A product is a contract, not merely an image

Every run begins with a typed product contract. The contract is the interface between business intent and model execution. It prevents the planner from silently changing the requested product and makes each output reproducible enough to inspect.

{
  "run_id": "immutable-id",
  "store": "target storefront",
  "category": "licensed product category",
  "audience": ["buyer segments"],
  "deliverables": ["master", "preview", "thumbnail", "manifest"],
  "style_dna_version": "style-family@v3",
  "character_profile": null,
  "rights_profile": "commercial-eligible@v2",
  "novelty_scope": "recent 180 days + protected references",
  "paid_attempt_budget": 2,
  "quality_thresholds": {"vision": 0.86, "novelty": 0.82},
  "publication_mode": "review-required"
}

The run identifier, input versions, selected model profile, seeds where available, prompts, adapter hashes, QA scores and derived-file checksums are stored in a manifest. This does not make a probabilistic model deterministic, but it makes the decision process traceable.

5. Style DNA: learning a distribution without memorizing a work

A common failure mode in creative automation is to place a folder of archive images into a prompt and ask for “something similar.” That approach mixes authorization, retrieval and generation into an opaque operation. PFD instead proposes a layered representation called Style DNA.

Style DNA is not a substitute for dataset governance. It is a structured control surface. For a recurring named character, a separate Character Profile stores identity-preserving evidence, reference permissions and consistency measurements. Store style, category style and character identity are deliberately versioned independently so one can change without retraining everything.

6. Three adaptation levels

Level A — prompt and policy adaptation

The cheapest route uses a general model with a compiled prompt, negative constraints, examples of acceptable metadata and deterministic post-processing. It is suitable when the task is generic and identity consistency is not required.

Level B — retrieval and reference conditioning

The system retrieves rights-eligible evidence and recent-history exclusions, then passes only the minimum necessary context to the planner or generation model. This follows the broad retrieval-augmented generation principle introduced for knowledge-intensive language tasks, adapted here as a governed creative-memory layer. Retrieval is also used for negative memory: previous failures are often more valuable than successful prompts.

Level C — LoRA adaptation

When a style family or recurring character requires stable representation across many products, PFD may train a Low-Rank Adaptation (LoRA) adapter over a qualified base model. LoRA freezes the base weights and trains a smaller set of injected parameters, making adaptation materially lighter than full fine-tuning. It is still training: dataset quality, licences, overfitting, leakage and evaluation remain first-class concerns.

Promotion rule. A LoRA adapter is not production-ready because its samples look attractive. It must beat the current route on a frozen evaluation set, pass similarity and safety checks, remain within cost and latency budgets, and survive a canary deployment with rollback available.

7. Concept diversity before image diversity

Changing a seed does not guarantee a different product. Three images can be visually different while expressing the same concept, title and commercial purpose. PFD therefore creates diversity at four levels before rendering:

1. Commercial intent: avatar kit, seasonal portrait, printable asset, social campaign or marketplace preview.
2. Semantic concept: subject, action, environment, narrative and emotional register.
3. Visual plan: camera distance, pose family, composition, palette and lighting.
4. Production form: transparent asset, scene, layered kit, promotional layout or animation-ready source.

The planner produces multiple candidate concept vectors. A pre-generation novelty gate compares those candidates with recent titles, descriptions, tags and product contracts. Candidates above a similarity threshold are rewritten or rejected before the paid image call. This is the cheapest place to prevent duplication.

8. Model routing as constrained optimization

No single model is optimal for every product. The router estimates quality, cost, latency and risk for a route m given contract x. A simplified decision rule is:

C is expected end-to-end cost rather than advertised per-image price; it includes failure probability and repair. Q is predicted acceptance quality, R aggregates rights, safety and privacy risk, and A is the number of paid generation attempts. A cheaper model with a high rejection rate may be more expensive than a premium model that succeeds on the first pass.

The router profile must include provider, model version or snapshot, supported operations, licence constraints, region, latency distribution, measured acceptance rate and deprecation status. External model names are configuration, not business logic.

9. The bounded-attempt generation policy

PFD’s first operational principle is generate once, diagnose once, repair once. The default production run creates one master image. If it passes all gates, no second paid call is made. If it fails with a repairable and localized defect, Vision QA generates a concise repair specification and a second call is allowed. If that attempt fails, the run stops and enters review; it does not continue burning credits.

A retry without a changed diagnosis is prohibited. Every paid attempt requires a reason code, expected correction and remaining budget. The policy converts cost from an accidental side effect into an explicit system invariant.

10. Multimodal quality and novelty gates

Acceptance is a conjunction of gates, not an average that can hide a critical failure. A high aesthetic score cannot compensate for an unauthorized logo, an incomplete subject or similarity to a protected work.

Image embeddings inspired by vision-language representation learning such as CLIP can support semantic similarity search; perceptual hashes can detect near-identical renders and simple transformations. Neither is a legal conclusion. Threshold hits create evidence for rejection or review, while provenance and licence remain separate determinations.

11. Dataset governance and training eligibility

The archive is not automatically a training dataset. Every asset receives a machine-readable eligibility state. Only records with affirmative rights evidence may enter an adapter-training snapshot.

• Identity: immutable asset ID, creator, store, category and timestamps.
• Rights: licence type, training permission, consent where relevant, territory and restrictions.
• Quality: technical defects, moderator decision, sales signal and known failure modes.
• Relationship: derivative family, duplicate group and protected reference set.
• Split: train, validation, frozen test or excluded. Near-duplicates cannot cross splits.
• Version: dataset manifest hash, inclusion query, exclusions and reviewer.

PFD adopts the spirit of dataset documentation and model-card research: a dataset snapshot and adapter version should be accompanied by intended use, out-of-scope use, composition, evaluation results, limitations and known risks. Deleting or revoking an asset must propagate into future snapshots and may require retiring an affected adapter.

12. LoRA training and model lifecycle

The Model Registry stores a semantic version, base-model identifier, adapter hash, dataset hash, configuration, code revision, benchmark results, approved use cases, owner and rollback target. A production alias points to an immutable version. Changing that alias is a controlled deployment, not an unrecorded file replacement.

Evaluation must include both positive prompts and adversarial negatives: forbidden marks, disallowed identities, anatomy stress cases, transparent-background requirements, small-object compositions and prompts close to recent catalogue concepts. Offline evaluation is followed by canary data: acceptance rate, review time, cost per accepted master, duplicate alerts and rollback events.

13. Agent-native company architecture

The zero-traditional-employee thesis requires more than a collection of chatbots. PFD models the organization as bounded roles with permissions, inputs, outputs and service-level objectives. Agents do not receive unrestricted access simply because they can reason about a task.

Each action is an event with an actor, run ID, policy version and idempotency key. Long-running work is queued; retries are safe; side effects are isolated; a failed worker cannot silently duplicate publication or payment. The system may operate continuously, but legal accountability, ownership decisions, exceptional review and incident response remain assigned to humans.

14. Observability, evaluation and economics

The most important metric is not images per minute. It is cost per accepted, publishable and sufficiently novel product. PFD’s research dashboard is organized around four groups:

Model comparison uses paired contracts and frozen evaluation sets. Provider marketing examples are not benchmarks. A candidate route is useful only if it improves the operational objective under the same product contracts and review policy.

15. Current implementation boundary and roadmap

The companion architecture specification records a cost-bounded baseline: one master image, no paid generation for gallery variants, automated master-image QA, novelty checks, local preview production, review and packaging. The architecture is provider-independent. Model identifiers in an implementation are versioned profiles and must be replaced when providers deprecate them.

Implemented or directly implementable baseline

• structured product contract and Style DNA metadata;
• one-master, maximum-two-paid-attempt policy;
• local gallery derivatives and ZIP/manifest packaging;
• visual QA, recent-history comparison and human approval;
• provider/model profile configuration rather than model-specific business rules.

Research and engineering roadmap

• rights-aware vector retrieval across eligible archive material;
• formal Character Profiles and cross-product identity metrics;
• versioned LoRA training datasets, offline evaluation and Model Registry;
• canary routing, automated rollback and drift monitoring;
• causal analysis of review feedback and commercial outcomes;
• agent permissions, signed manifests and organization-wide audit events.

16. Limitations and responsible-use commitments

This architecture does not solve authorship law, eliminate bias, guarantee non-infringement or make aesthetic judgment objective. Embedding distance is not proof of originality. A LoRA adapter can memorize. Vision models can miss defects. Human reviewers can be inconsistent. Commercial feedback can reward repetition and therefore must not become the only optimization signal.

PFD’s proposed controls are defence in depth: affirmative data rights, separate protected-reference sets, measurable novelty gates, hard attempt budgets, model and dataset lineage, manual escalation, canary deployment, rollback and incident records. The design is informed by the NIST AI Risk Management Framework and its Generative AI Profile, especially the need to govern, map, measure and manage risk across the lifecycle rather than at the final output alone.

Accountability principle. Automation may execute a decision path, but it must not erase the identity of the policy owner, the evidence used or the person accountable for deployment.

17. Reproducibility checklist

• Immutable run ID and timestamp
• Product-contract and Style-DNA versions
• Base-model, provider and snapshot identifier
• LoRA adapter, dataset and evaluation-set hashes
• Prompt, negative prompt and repair reason code
• Paid-attempt count and end-to-end cost
• QA scores, nearest neighbours and gate decisions
• Derived-file checksums and package manifest
• Reviewer decision, override reason and publication state
• Rollback target and retention policy

18. Conclusion

PFD’s core proposition is that reliable creative automation is an engineering system, not a prompt collection. The system must know what it is allowed to learn from, what makes a proposed product different, which model is appropriate, how much it may spend, why an output passed, and how to undo a bad deployment. Foundation models supply broad capability; PFD’s differentiated value is the governed loop that turns that capability into repeatable commercial production.

The agent-native company is the organizational extension of the same idea. Routine work can be delegated to bounded agents and robots only when their authority, evidence, cost and failure behavior are explicit. The goal is not automation without responsibility. The goal is a compact company whose execution is software-defined, continuously evaluated and auditable from concept to publication.