The Judgment Layer: When to Trust, Adapt, or Stop

OpenMontage is best understood as a governed production system whose final question is not “Did a tool return a file?” but “Does the evidence support the deliverable that was approved?” The judgment layer connects the agent, instruction contracts, live capability choices, budget, checkpoints, provenance, and post-render review into a practical decision: trust the run, adapt it with approval, or stop because the promised result is not supportable. That is the standard that turns flexibility into something a user can responsibly use.

The answer is conditional. OpenMontage does not provide one permanent guarantee for every provider, pipeline, or machine. It provides mechanisms for making a decision visible and revisable. A production can be trustworthy even when it uses a modest local path, as long as the promise is honest and the evidence is sufficient. A production can be untrustworthy even when it uses a premium model, if a fallback, runtime swap, missing source, or unresolved quality issue is hidden.

Trust requires alignment, not perfection

A user should trust a run when five things line up.

First, the approved intent must be clear enough to evaluate. The selected pipeline and production plan should say what kind of video is being promised: motion-led, source-led, data-explanatory, teacher-oriented, screen-based, hybrid, or another supported treatment. The point is not to force every brief into a category. The point is to prevent a later stage from changing the objective without acknowledging it.

Second, the live capability envelope must support that intent. The registry's provider menu and runtime checks should show what is configured on the current machine, not what the repository happens to document. A missing credential, unavailable binary, absent Node dependency, or insufficient local hardware is not a minor implementation detail if it changes the path the user approved. A provider score can help rank available options, but availability and fit remain separate questions.

Third, the production state must be coherent. Required predecessor stages should be complete, gated stages should carry actual approval, and the current artifact should match the decisions that led to it. Checkpoint history and decision logs matter here because they show whether a provider, voice, style, source, or runtime changed during the run. A current file without its decision history can be difficult to interpret, especially when a revised path costs more or produces a different visual character.

Fourth, the budget and provenance must be acceptable for the intended use. Estimate, reserve, and reconcile give the user a way to see expected and actual cost, but prices and quotas remain time-sensitive. A free or local path can be valid without being effortless. It may require hardware, installation, storage, or different quality expectations. For retrieved media, source URLs and license declarations should travel with the asset. They improve traceability, but they do not settle every rights question.

Fifth, the actual output must survive layered review. Pre-compose checks should not find a delivery-promise violation, a missing asset, or a composition that is likely to feel like a slideshow. Post-render review should establish that the file is technically real, contains the expected media, and does not show obvious black frames, silence, clipping, broken subtitles, transcript divergence, or an unapproved runtime change. These checks do not need to be perfect to be useful. They need to be read at the level they support.

Trust therefore means “the important relationships are aligned and the remaining uncertainty is understood,” not “the system guarantees quality.” A motion-led project can pass technical review and still need creative revision. A documentary montage can have complete provenance and still contain a weak sequence. A local explainer can be an honest success even if it does not look like a premium cloud-generated trailer. The promise sets the comparison.

Adaptation is legitimate when the decision remains visible

Production rarely follows the first plan exactly. Adaptation is not a failure when it preserves the approved purpose or receives explicit approval for a changed purpose. The key question is whether the change modifies an implementation detail or changes what the user is being asked to accept.

Some changes are usually within the same promise.