Day 17 — Critique, Verify, and Route

Today Cutroom produces useful judgment: where an edit loses clarity, where on-screen text contradicts speech, or where a call to action arrives without supporting setup.

Ask over segments, not raw video

Send a bounded set of SegmentEvidence values and request structured findings:

pub struct CritiqueFinding {
    pub category: FindingCategory,
    pub statement: String,
    pub segment_ids: Vec<String>,
    pub evidence_ids: Vec<EvidenceId>,
    pub severity: Severity,
    pub suggested_edit: String,
}

Validate every reference and require the suggested edit to preserve factual claims. A generated finding with an unknown segment is invalid, even if the prose sounds plausible.

Route expensive work

Run cheap signals first. Escalate a segment to a vision-language model only when OCR confidence is low, motion is high, the transcript and screen text conflict, or a frozen eval shows that escalation helps.

all segments ─▶ cheap specialists ─▶ confidence/policy
                                      ├─ accept evidence
                                      └─ expensive multimodal model

Measure recall, citation accuracy, p95 latency, decoded bytes, accelerator seconds, and provider cost. Routing is successful only when it moves the measured frontier.

Run the funnel

cd rust/rust-ai-engineering
cargo run -q -p mosaic-ml --example video_generation_funnel
cargo run -q -p mosaic-evidence --example multimodal_adversarial_lab

Then add a low-confidence segment that should escalate and a high-confidence segment that must not.

What you built

Cutroom reduced a rich modality into inspectable, aligned evidence before asking for judgment. Rust made clocks, bounds, specialist contracts, queues, and citations explicit; AI contributed perception and critique where deterministic code could not.

Finish with Project 4 Revision →.