The Optimization Journey

The five stages a gateway climbs from connected to evidence-earned optimization, and what unlocks each one.

The Optimization Journey

A gateway does not arrive able to make measured claims. It earns them, and the journey is the five-stage narrative of that: Observe, Compress, Measure, Audition, Promoted.

The dashboard renders it as a strip, and there is a journey pill in every Inference Platform page header -- that pill is the journey's entry point, since it has no navigation item of its own.

The stages

StageStateWhat it requires
1. ObserveGateway connected, traffic flowing, nothing altered.A connected proxy gateway.
2. CompressLossless compression is on.Transforms enabled.
3. MeasureA holdout control arm is running and the gateway can produce judge evidence, so savings become measured instead of estimated.A non-zero holdout percentage, and a parity channel.
4. AuditionContent-dropping transforms run in the background -- projected, never served -- accumulating would-have-saved evidence.Shadow evaluation enabled, and retrievable originals enabled.
5. PromotedEvidence has cleared and those transforms serve live traffic, with retrievable originals behind them, per workload type.Earned evidence, per team, workload type and billing regime.

A team reaches stage 5 when every one of its active gateways is at stage 4 and at least one billing regime has promoted workloads.

What is blocking you

The dashboard names the blocker rather than showing a percentage. The reasons are a fixed vocabulary, so the answer is always actionable:

BlockerWhat to do
No proxy gatewayInstall one. A direct source has no request path to optimize.
Enable compressionTurn transforms on for the installation.
Enable holdoutSet a non-zero holdout percentage.
Needs a parity channelThe gateway cannot produce judge evidence. Either content uplink is on and a hosted judge can score, or the in-environment judge is armed.
Needs a judge keyThe parity judge has no credential.
Enable shadowTurn on background evaluation of content-dropping transforms.
Enable retrievable originalsTurn on the reversibility net those transforms depend on.

A gateway is parity-capable when a content channel exists -- either the content uplink or a judge running beside the gateway -- and a holdout is running. Without both, nothing ever promotes. On a content-free deployment the in-environment judge is the residency-safe way to get there, and it is armable from the dashboard while the judge credential and the content uplink stay environment-only.

There is a one-click preset that advances a gateway through the journey's enabling settings as a correct combination -- compression, holdout, shadow evaluation and retrievable originals together -- applied explicitly through the audited configuration path.

Defaults stay off. The profile is consent, not a default. Nothing in this journey happens to your traffic because you installed a gateway.

Why promotion is earned rather than toggled

A content-dropping transform is dark until trailing holdout evidence in that billing regime says it may serve, for a specific team, workload type and regime combination. It is reversible by construction: evidence degrades, the transform demotes automatically.

Two independent gates must both allow a transform for it to serve:

  • The promotion gate -- per team, workload type and billing regime, decided by evidence.
  • The release stage -- per transform, deciding whether it may run at all before any team's evidence is consulted. It fails safe: a configuration it cannot parse degrades to dark.

Promotion earns an entitlement, not an immediate full rollout. Where a transform has a release stage, going live runs through that stage's canary staging rather than switching a population's whole traffic at once.

Your team can additionally require an explicit admin approval for evidence-earned promotions. Approval is consent, never eligibility -- the evidence gate stays the sole authority, and demotion is automatic in both modes.

Reversibility

A promoted content-dropping transform is recoverable rather than destructive. Originals it dropped are parked in a short-lived store inside the deployment, which the model can retrieve on demand. That is the net which makes the whole loop safe to run on production traffic, and it is why enabling it is one of the stage-4 requirements rather than an optional extra.

Reading the savings it produces

See Savings for what a measured claim means, why a subscription dollar figure is always an estimate, and why a sub-percent share of effective context is not a small result.

On this page