Aggregate evidence
Counts, methods, synthetic examples, architecture, decisions, limitations, and the planned evaluation.
How the system fits together
A compact reference for the model roles, data contracts, checkpoints, public/private boundary, and reproducibility decisions behind the project.
Architecture, data flow, checkpoints, security boundaries, and a practical glossary.
01 / Architecture
Model work stays private; documentation stays static
Private inputAd media + transcript
Locally held, streamed only through review tools.
Private inferenceBase Inkling
Produces expressive targets or matched comparison arms.
Private trainingLoRA on Tinker
Remote rank-32 adapter and checkpoints.
Public outputThis static journal
Aggregate evidence, synthetic examples, and no provider calls.
The local comparison application can call Base Inkling, the trained sampler checkpoint, and eventually Fish Audio. This public Cloudflare site cannot. It has no server logic, secret binding, D1 database, R2 bucket, or provider endpoint.
02 / Data contracts
The narrow interface makes the result reviewable
| Stage | Consumes | Produces | Primary gate |
|---|---|---|---|
| Selection | Private ad media | Frozen cohort identity | Explicit human train/do-not-train decision |
| Preparation | Selected media | Audio and verified transcript | Readable media and transcript integrity |
| Teacher | Audio, script, task context | Global and segment expressive direction | Structure and wording preservation |
| SFT | Prompt-target pairs | Updated LoRA adapter | Deterministic render and completed optimizer steps |
| Comparison | Matched prompts and model identities | Base/trained scripts | Symmetric settings and full arm preservation |
| Fish render | Two expressive scripts, one voice | Blind A/B audio | Same voice and synthesis settings |
Errors are records, not inconveniences to erase. A malformed teacher output, a wording mismatch, or a failed provider arm remains visible and cannot be silently converted into a success.
03 / Glossary
The terms that caused the most confusion
Base Inkling
The unchanged general audio-language model. It can receive audio or text and generates text. It is the parent over which an adapter is trained.
LoRA
A compact trainable adapter that modifies behavior while leaving the base weights fixed. “Fine-tuned model” here means Base Inkling loaded with this adapter.
Sampler checkpoint
The remote artifact used to load the trained adapter for inference. The public documentation names the role but never publishes the concrete identity.
Training-state checkpoint
The resumable state used to continue training, including optimizer progress. It is not the normal artifact for a comparison request.
Teacher label
The expressive target produced from source evidence before review. It becomes training data only after the required gates accept it.
Renderer
The downstream text-to-speech model. Fish Audio S2-Pro is the planned renderer; Inkling itself does not synthesize the final ad voice.
04 / Public boundary
What this site intentionally does not contain
The production build is scanned before deployment. It rejects provider-token patterns, private checkpoint schemes, absolute home paths, authorization headers, environment assignments, known private identifiers, media, databases, archives, and model weights.
Operational state
Source ads, audio, transcripts, URLs, identities, credentials, local paths, checkpoints, raw logs, and generated comparison media.
Credentials are read from a server-side environment only when private tools run. The public browser receives no model key, voice reference, checkpoint identity, or inference capability.
05 / Reproducibility
Reproduce the method without reproducing the private data
- Collect media you have the right to process and define a reviewable selection criterion.
- Freeze a cohort before labeling; record exclusions rather than deleting history.
- Create word-preserving expressive targets and review them against source audio.
- Prove the training path with one row before running the complete reviewed set.
- Store serving and resumable checkpoints as different artifact roles.
- Freeze untouched scripts, run Base Inkling and the adapter with matched settings, and preserve failures.
- Render paired text through one Fish voice, listen blind, lock preference, then reveal.
- Publish the decision rule and the failures alongside any result.
Reproducible does not mean public raw data
The experimental structure can be inspected and repeated while private creative, licensed media, personal voices, and service credentials remain protected.