What the Architecture Promises—and Where It Can Break
OpenMontage's agent, instruction layers, pipeline contracts, and feedback loop form a coherent design, but they do not all provide the same kind of assurance. This chapter connects those ideas into a practical way to judge a run: identify what is enforced by a contract, what depends on the agent following a procedure, what is only a heuristic, and what comes from a changing or self-reported document. That distinction is the difference between inspectable flexibility and overconfidence.
The question is not whether OpenMontage has safeguards. It clearly does. The question is what each safeguard can establish. A checkpoint can prove that a gated state was written correctly. It cannot prove that a person looked carefully at every creative choice. A pipeline manifest can define the intended stages. It cannot prove that every provider named in it is available today. A review score can expose a likely slideshow. It cannot decide whether an unusual edit is artistically justified.
A reliable run is a stack of different kinds of claims
The first category is the persistent contract: a rule that is represented in a structured interface and, in important cases, enforced when state changes. Pipeline manifests define stage order, required inputs, outputs, review criteria, and approval defaults. Canonical artifacts give those stages a shared vocabulary. The checkpoint implementation reads the manifest, validates predecessor conditions, and rejects a completed gated stage when approval evidence is absent. The composition path also carries a selected render runtime forward and can block an unapproved runtime change. These mechanisms do not make a production good, but they constrain what counts as a valid transition.
The second category is the procedural rule: an instruction that tells the agent how to make a good decision within that contract. The agent is told to read a stage skill before acting, use the live capability report at preflight, present meaningful options, announce consequential provider choices, preserve the approved promise, and review the resulting artifact. The executive-producer procedure says to compare actual results with cross-stage expectations and send work back when a downstream fact changes an upstream decision. These rules are essential because the contract cannot encode every creative judgment.
They are also more dependent on compliance. A manifest can require a review, but the quality of that review still depends on the agent examining the right field or asset. The project documents this dependence rather than claiming that the agent is a deterministic controller. Python provides the tool and persistence layer; the coding assistant interprets the instructions. A readable rule is easier to inspect and revise than hidden orchestration, but it remains a rule that an interpreter must follow.
The third category is the heuristic. OpenMontage uses heuristics where a formal proof of quality is impossible or too expensive. A delivery promise can classify whether an edit contains enough real motion for a motion-led brief. A slideshow-risk scorer can measure repetition, decorative visuals, weak shot intent, purposeless motion, and typography overuse. A post-render review can sample frames, inspect audio levels, and compare a transcript with a script. Those checks turn vague concerns into evidence that a reviewer can act on, but they do not transform taste into a number.
This distinction prevents a common category error. If a risk score says that a scene plan resembles a slideshow, the score is not a verdict that the video has no artistic value. It is a signal that the plan has characteristics associated with a weak delivery of the stated promise.