The Wake · field notes
Notes from the control plane.
Writing on the unglamorous parts of generative media — provenance, failover, custody, compliance deadlines, and what models actually cost — from the team building the layer above the providers.
You didn't train a model. You're still a provider under Article 50.
Call an image or video model over an API and ship the output under your own name? Article 50(2)'s marking duty is very likely yours — not the model vendor's.
9 min readRead the post →
A log is not a mark: what the EU's Code of Practice actually requires
The final Code makes logging an optional layer that 'alone is not sufficient.' What actually marks AI content — and where Wakemark's ledger honestly sits.
8 min read
When a marked file becomes a prompt: Measure 1.2 and the provenance chain
The Code says preserve marks even when marked content becomes the input to a new generation — a cross-provider lineage problem no single vendor can see.
7 min read
EU AI Act Article 50: the compliance clock on generative media is already running
Article 50 applies from 2 August 2026. What machine-readable marking means for image, video, and audio pipelines — and the evidence auditors will ask for.
7 min read
C2PA Content Credentials survive exactly as long as you never touch the bytes
One re-encode, one thumbnail, one optimizing CDN — and the provenance manifest is gone. How to design a media pipeline that keeps Content Credentials intact.
6 min read
Your outputs have a half-life: what media providers keep, and for how long
Replicate deletes outputs after about an hour; fal keeps them roughly a week. If your product links to provider URLs, that's a countdown, not storage.
5 min read
Cross-provider failover: the 429 always comes at launch
Rate limits and provider outages hit exactly when traffic spikes. Why failover inside one aggregator isn't enough — and why input errors must never retry.
6 min read
The model you shipped on will be gone in a year
Runway's Gen-4 Aleph lasted about twelve months. Model APIs retire faster than roadmaps — capability-based routing is how a product outlives any one catalog.
5 min read
Spend caps that actually stop spend
Billing surprise is the #1 complaint about every generative media platform. Per-job cost attribution, hard caps, and why an alert is not a limit.
5 min read
How to choose a text-to-image model (without a Twitter thread)
FLUX dev vs pro vs ultra, gpt-image-1, Z-Image, schnell-class models: a working taxonomy for picking text-to-image models by job, not by hype.
7 min read
Text-to-video in production: notes from the queue
Kling, WAN, and MiniMax can hold a shot now. What still breaks in production text-to-video: queue variance, prompt drift, and cost per usable second.
6 min read
Put a ledger under it: event sourcing for AI pipelines
When legal asks which model made an asset eighteen months ago, status=succeeded is not an answer. Why generation history belongs in an append-only ledger.
6 min read
Your provider keys are money now: BYOK without the foot-guns
A leaked media-provider key burns GPU dollars by the hour. Where provider keys should live, what envelope encryption buys, and what to demand from platforms.
5 min read
The infrastructure behind the writing
Everything covered here — capability routing, cross-provider failover, byte-exact custody, the audit ledger — is what Wakemark ships. One API above the providers, with receipts.