Decision models

A decision model is a small, fast AI service that answers simple multiple-choice questions, such as "which colour look fits this footage: warm, cool or filmic?". Some workflow decisions are simple enough that such a model could answer them instead of a full agent, which would be quicker and cheaper. ReelBolt's decision models let you test that idea safely first and only then, if the numbers support it, let the fast model take over some of those decisions.

This feature is optional, advanced and switched off by default everywhere. Most installations never need it.

The three decision model modes: Off, Shadow and Gate

Mode	What happens	Changes your videos?
Off	Nothing. No decision model is asked anything. This is the default.	No
Shadow	The agent makes the decision as usual. The decision model is asked the same questions on the side, and ReelBolt records whether it agreed and how confident it was.	No, purely an observation
Gate	The decision model is asked first. If it is confident enough on every question for that step, its answer is used and the agent is skipped. If it is unsure about even one question, the whole step goes to the agent as normal.	Yes, when the model is confident

The usual path is: run in Shadow for a while, look at the results on the calibration page, and only switch to Gate for the decisions where the model has proven reliable.

A decision model can never cause a step to fail. If it is not configured, is slow (it gets a few seconds by default), or answers nonsense, ReelBolt simply falls back to the agent.

Where ReelBolt can use a decision model

Decision models are only used at a few specific points:

• the Colorist agent (choosing a colour-grade look),
• the MusicSupervisor agent (choosing music intensity, ducking and fit),
• the colour grade room, to decide whether the room needs to meet at all,
• a Conditional step set to decide its branch with a decision model,
• a ReviewLoop step's optional early stop.

Everything else in ReelBolt is unaffected.

Turning decision models on in ReelBolt

Two things are needed:

1. A Decision provider. An administrator adds an inference provider with the Decision (Jev / logprob) capability and makes it the default. Suitable kinds are TypeSafe, a self-hosted decision server (OpenJev), Azure OpenAI, OpenAI-compatible services and DeepSeek. Anthropic and Google Gemini cannot be used. See Inference providers.
2. A mode. For the Colorist and MusicSupervisor agents, the mode is set by whoever runs the installation, in its settings (COLORIST_DECISION_GATE_MODE and MUSIC_SUPERVISOR_DECISION_GATE_MODE, each Off, Shadow or Gate). The other three uses are switched on in the settings of the individual workflow step. Neither the dashboard nor the assistant can change these modes.

Privacy note. When a decision model is a cloud service, the questions and some context about the step are sent to that service, in addition to the main AI service the agents already use. In Shadow mode that context can include transcript text and on-screen text from your footage. A self-hosted decision server keeps this on your own network. Bear this in mind before turning it on.

Reading the Decision Calibration page in ReelBolt

Administrators can open Admin → Calibration (/app/admin/calibration), the Decision Calibration page. It is read-only: it shows what the decision models have done so far and what would happen at different settings. It changes nothing. Open How this page works at the top for a built-in explanation.

The key ideas:

• Observations: how many decisions were recorded.
• Agreement: how often the decision model gave the same answer as the agent.
• Confidence: how sure the model said it was. A trustworthy model agrees more often when it says it is more confident; the reliability chart shows whether that holds.
• Accept-at and Minimum margin (the two sliders): what "confident enough" means. A decision is accepted only if the model's top answer is at least Accept-at sure and beats the runner-up by at least the minimum margin. The defaults are 0.85 and 0.25. Move the sliders to see how many decisions the model would have handled alone at other settings.
• Would handle alone / Would escalate: the share of steps the model would have taken over, or sent back to the agent, at the slider settings.
• Tokens saved: an estimate of the AI usage that would have been saved at the slider settings. It is a projection, never a measurement.

Each agent gets a card with an advisory badge: Not enough data (fewer than 30 observations), Keep Shadow, or Safe to try Gate at X. These are suggestions based on simple rules, not guarantees, and none of the default thresholds has been tuned for your footage. Use Ask assistant for a plain-language recommendation; the assistant can read this report but cannot change any setting.

If the page says No decision observations recorded, no workflow has run with Shadow or Gate switched on yet.