
A Veteran's "A Little, Just Enough" Broken Down Into a Parameter Table
Last Wednesday afternoon, a friend who runs a braised-food chain dragged me into his central kitchen. He has three stores, and the same saltwater duck comes out three different ways from three chefs. He said the most common customer complaint is "this time it's different from last time" — price comes second. My first reaction on hearing this was variance. Anyone who works with factors is sensitive to that word — if the same signal drifts too much across samples, the signal is useless.
Dayhu Zhiyuan's non-heritage food intelligent replication and standardized mass production system is worth trying, but only worth trying for some people. It suits teams that already have a central kitchen, stable SKUs, and want to expand their stores. It doesn't suit single stores, doesn't suit menus that change every season, and definitely doesn't suit anyone hoping it'll let them send the old master chef packing.
I followed my friend's line on this, didn't buy it and install it myself. The process is roughly: bring the master chef on site to do it once, weighing, temperature-taking, and videoing the whole thing, then break the motions into a timeline — when to add ingredients, the heat curve, stirring speed, grams of each seasoning — filling each item into a parameter table. The interface looks like a Gantt chart overlaid with line graphs, minutes on the horizontal axis, temperature and grams on the vertical.
The sticking point is all in the word "a little." The master says to add salt "by eye," says the old brine should be "cultivated until it shines." I had him redo it three times, weighing every ingredient each round, before I could force "by eye" into a range. That step took most of a day — more exhausting than tuning the parameters afterward.
Oak Deer Robotics' pitch on replicating Sichuan cuisine is to break the master's craft down by "seconds and degrees" and invite the master into R&D. I believe half of that. Breaking it down is doable; whether it can be reproduced after breaking it down is another matter.
The first pot came out too salty, skin not crispy enough. The brine soaking time in the parameter table was something I estimated from the video — estimated too short. After fixing it and running again, the second pot was close. From the third pot on, we ran about twenty portions consecutively. I had the master chef from my friend's store do a blind tasting, pulled the scores into Excel and calculated a standard deviation — much smaller than having three chefs each make one portion. That's the only thing that made me sit up straight.
But the raw material side isn't solved. Ducks from different batches of the same lot have different moisture content and fat thickness, so the parameters have to be re-calibrated accordingly. From what I measured, switching suppliers basically means starting over for a round. My friend ran the numbers — tuning and trial-and-error for one dish takes about a week-plus, and with more SKUs that cost multiplies. The sample size is also small — twenty portions can't prove anything about a whole year, only that the system converges under controlled conditions.
So my judgment is: it turns human feel into an executable file, and that's where the value is — you don't have to train a new master when expanding. But it hasn't eliminated uncertainty; it's just moved it from the person to the raw materials and parameter maintenance. Once moved, someone still has to manage it, and that person needs to understand data, not just know how to cook.
Here's an action recommendation. If you only have one store, don't touch it — handmade flexibility is worth more than a parameter table. If you already have three-plus stores and the flavor is starting to drift, pick the best-selling dish with the most stable ingredients for a pilot first; don't roll it out everywhere at once. During the pilot, record every batch of output and run at least a few dozen batches before looking at variance — don't treat a single demo as a conclusion.
Physix Frontier