AI Training, Testing and Validation Data Transparency
Effective: August 11, 2026
Version: 1.2
Service release covered: EVELI Casting Beta, repository revision reviewed August 11, 2026
This page describes data Eveli, Inc. uses to train, test, validate or fine-tune the EVELI generative-AI service. California's disclosure rule treats testing and validation as part of “training” for this purpose. EVELI therefore includes its orchestration and quality-assurance materials even though EVELI does not pretrain or fine-tune the underlying image model.
1. System identification
| Field | Disclosure |
|---|---|
| System/service and public version | EVELI Casting Beta, August 11, 2026 release |
| Intended purpose | Generate synthetic casting faces and consistent multi-view talent anchor images from structured creative selections |
| EVELI-developed components | Prompt orchestration, internally structured casting taxonomy, prompt templates, anchor dependency workflow, generation safety/spend gates, credit enforcement, private-asset storage and automated tests |
| Third-party inference routes | fal.ai \fal-ai/nano-banana-pro\ and \fal-ai/nano-banana-pro/edit\; fal.ai identifies the underlying model as Google's Gemini 3 Pro Image. \fal-ai/nano-banana\ is present as an allowed fallback identifier but no current role opts into it |
| EVELI model training or fine-tuning | None. EVELI does not pretrain or fine-tune the underlying image model for this release |
| Data cutoff for this disclosure | Repository and public provider information reviewed through August 11, 2026 |
2. EVELI-controlled data used for testing and validation
| Data set | Source/owner and purpose | Approximate size and type | Labels and processing | Rights/personal information | Dates and synthetic content |
|---|---|---|---|---|---|
| Casting taxonomy and prompt library | EVELI-authored structured creative taxonomy used to map user selections into prompt phrases and to validate that required anchor variables resolve correctly | Approximately 478 structured taxonomy rows and 10 prompt templates; text, categories, mappings and metadata | Category/value labels, applicability labels, status labels, anchor framing and prompt-variable mappings; normalized and compiled into source-controlled TypeScript | Copyrightable EVELI-authored text/code is included; no consumer personal information is intended | Compiled during EVELI development through August 11, 2026; includes descriptive synthetic-character attributes |
| Generated synthetic QA combinations | Generated by EVELI's dry-run and test code from the taxonomy/prompt library to exercise mapping, validation, credit and workflow behavior | Variable per test run; structured prompts, synthetic casting states, expected status/results and local placeholder URLs | Automatically assembled, compared against expected validation and state-transition results, and discarded or retained with source tests | No customer content is required; test combinations are fictional/synthetic | Used during development through August 11, 2026; synthetic data is included |
| Bundled dry-run visual set | Sixteen face image files listed in \public/faces/manifest.txt\ and nine role-based anchor images used for local dry-run display and visual workflow checks | 25 image files plus filename/name or role labels | Cropped/sized visual placeholders and role mapping; not used to update model weights | Image copyright/provenance is not recorded in the current repository manifest. EVELI therefore treats the set as rights-status unverified and potentially depicting identifiable adults. The set must be replaced with documented EVELI-owned or licensed synthetic assets before public California release if it remains part of release validation | Bundled before the August 11, 2026 review; whether every face image is synthetic is not documented |
EVELI does not use aggregate consumer information as an EVELI training, testing, validation or fine-tuning dataset for this release.
3. Private customer content
Private customer Inputs, references and Outputs are not used to train or fine-tune a generalized model and are not part of EVELI's validation corpus by default. A live request may use the customer's permitted reference images transiently to generate that customer's requested Output. That inference use is service delivery, not an authorization to add the material to a reusable evaluation or training set.
EVELI has not launched a customer-contribution program. If one is introduced, it will use separate, asset-specific consent and license records and this page will be updated before any contributed content enters a reusable dataset. Identifiable real-person, minor and confidential-client assets will remain ineligible.
4. Third-party model information
| Provider/model | Public information | EVELI modification/evaluation data | Retention and training control | Last verified |
|---|---|---|---|---|
fal – Features & Labels, Inc.; \fal-ai/nano-banana-pro\ | https://fal.ai/models/fal-ai/nano-banana-pro | EVELI supplies an assembled prompt and generation parameters; EVELI does not fine-tune the model | fal.ai's documented default stores request input/output JSON for 30 days unless \X-Fal-Store-IO: 0\ is sent. EVELI requires that header before live customer use | August 11, 2026 |
fal – Features & Labels, Inc.; \fal-ai/nano-banana-pro/edit\ | https://fal.ai/models/fal-ai/nano-banana-pro/edit | EVELI supplies an assembled prompt, generation parameters and permitted reference-image URLs; EVELI does not fine-tune the model | Same payload non-retention requirement; temporary provider media must use a short expiry and restricted ACL before live customer use | August 11, 2026 |
| Google Gemini 3 Pro Image, reached through fal.ai | Identified by fal.ai as the underlying Nano Banana Pro architecture | No direct EVELI training or fine-tuning relationship; EVELI evaluates only the behavior of the routed service | Governed through fal.ai's service terms, DPA and subprocessor chain; private customer content is not licensed by EVELI for generalized training | August 11, 2026 |
5. Release updates
EVELI will update this page before releasing a substantial modification in California that changes the model route, introduces fine-tuning, adds a reusable evaluation dataset or begins an optional contribution program. EVELI retains the source-controlled inventory and a snapshot of each published version.
Questions: legal@eveli.ai
Primary source: https://leginfo.legislature.ca.gov/faces/billTextClient.xhtml?bill_id=202320240AB2013