Champion–challenger comparison answers whether a candidate scores acceptably beside the frozen production model. It does not answer how you hand that candidate a growing share of live reject, advisory, or setpoint influence without discovering the cost on 100% of the line.
This note is percentage-based canary inference rollout for plant models: traffic splits, promotion gates, and kill criteria. It is not concept-drift theory, not training–serving skew diagnosis, and not recipe MoC model-freeze alone—though a canary without those disciplines is theater.
A working pattern looks boring on purpose. Start with a small, named percentage of eligible decisions—often single-digit—on a scoped product family or line. Log canary and champion outcomes with the same feature clock. Watch paired metrics the plant already prices: false-reject rate, escape proxies, latency against the control cycle, and disagreement clusters by SKU. Hold the percent until the gate window completes; do not “bump to 50% because Tuesday looked fine.”
Gates that belong on paper before traffic moves
Write promote and kill thresholds before the first percent goes live. Examples plants actually use: sustained false-reject lift beyond a fixed delta, latency breaches that miss the control cycle, spike in human overrides, or disagreement density on a protected family. Rollback must be one action—route 100% back to the champion artifact—with an owner and a time box, not a meeting invite.
Safety-critical hard interlocks stay out of percentage experiments. Soft sensors and quality advisors are the usual canary surface; anything that can hurt people or equipment needs a different assurance path.
Where canaries fail in plants
Teams confuse shadow mode with canary: logging without influence never stresses the write path. Others canary on the easy mix and promote into the hard mix. Some skip delayed labels and declare victory on agreement with the champion—two wrong models can agree. A few leave the canary percent running for months as an unofficial A/B test with no MoC.
Percentage gates do not replace offline evaluation or champion–challenger shadow runs. They sit after both. Plants that keep a dated canary plan, a kill list, and a one-step rollback learn how much live traffic a model has earned. Plants that only keep the latest pickle file learn on scrap.