What Wakemark is, and what it is not

The shape of it

You send a generation request. Wakemark routes it to a provider — either one you name explicitly, or the best candidate for a capability you asked for — submits it with your provider key, tracks it to a terminal state, persists the output byte-for-byte, records a SHA-256 of those bytes, detects any C2PA manifest the provider embedded, and writes every step to an append-only event ledger.

What you get that a direct provider integration does not give you is the record. The job, its inputs, which provider and model actually ran it, every retry and failover, the digest of every byte stored, and the provenance of the file are one queryable history rather than a scatter of provider dashboards.

The ledger is the truth

The events table is append-only and a job's status is a projection of it — a cached fold, not a separate source of truth. Terminal states never transition. The fold is pure and converges under duplicated, reordered and dropped provider deliveries, which is the failure mode that actually kills hand-rolled media pipelines: a webhook that arrives twice, or arrives before the submit call returns.

There is one sanctioned exception, and it is evented: when a retention deadline passes, content-bearing fields inside event payloads are redacted in place and the redaction is itself recorded as an event. Event rows are never deleted or reordered. See Retention.

What Wakemark is not

Not a model provider

Wakemark trains nothing and hosts no weights. Every generation runs on somebody else's model, on your account, under your key. When a provider embeds a C2PA manifest, the provider made it — Wakemark preserves and reports it, and never creates one.

Not a reseller of inference

You bring your own provider keys and pay the provider directly. Wakemark charges for orchestration, custody and provenance — never a percentage of what you spend with fal, Replicate or anyone else. There is no markup on routed spend because there is no routed spend to mark up.

Not a compliance product

Wakemark is designed to support EU AI Act Article 50 marking and logging workflows. It does not establish compliance with the AI Act or any other law, and the obligations remain yours. See Article 50 for what the audit export does and does not attest.

Not a training corpus

Customer content is never used to train, mine or build datasets — not aggregated, not anonymised, not ever. What Wakemark does learn from is content-blind: latency, cost and reliability per provider and model, which touches none of your bytes.

Not synchronous inference

Jobs are asynchronous by design. There is no chat or token streaming, no media transcoding, and no hedged requests — firing two providers in parallel and keeping the fastest is normal for text gateways and unacceptable here, because in generative media every request costs real money.

Why bring-your-own-key changes the posture

Because inference is billed to your provider account, Wakemark never sits between you and your spend. That has a pleasant consequence for trust: the parts of the system that could be tempted to over-collect — usage, prompts, outputs — have no billing reason to exist. What Wakemark meters is bytes it stores and jobs it admits, both of which it can see honestly without reading a prompt.