Not every successful generation deserves a permanent place on the shelf. A grid that looks strong in the moment can still be the wrong thing to keep. After twenty posts and more archived adapters than I care to count, the question I end up asking is simpler than most evaluation frameworks: is this workflow actually worth keeping?
This final Field Note is the keep-or-kill test I run on every MiniMax H3 Turbo and LoRA recipe once the initial excitement fades. It is less a scoring system and more a set of practical pressures that force an honest answer.
Five Questions That Cut Through Sentiment
The first question is whether someone else can run the recipe from the card alone. If the answer is no, it is still a personal habit wearing a recipe’s clothes. Documentation fixes this more often than additional training data or a higher strength value. A card that needs the original experimenter to interpret it has not yet earned a place on a shared shelf.
The second question is whether the recipe still clears a frozen probe set. A pass rate that has slipped below the threshold I set when the recipe was first published is not a temporary dip. I do not average away the bad cells because the rest looked fine. A recipe that has failed its own probe set three months in a row is already archived in everything but name.
Cost is the third pressure. Here cost means attention as much as compute. A recipe that saves three hours of client revision but burns four hours of maintenance every month is not a net win. The monthly review forces that comparison into the open. The recipes that survive are the ones whose value still exceeds their carrying cost.
The fourth question asks whether the failure modes are known and logged. Unknown failures are more dangerous than known ones. I would rather keep a narrower recipe whose failure modes are fully documented than a broader one that still surprises me under ordinary prompts. Surprises belong in the experiment log, not in client work.
The fifth question is the sharpest: would I reach for this recipe again under real deadline pressure? Sentiment wants to keep things that once impressed a client. Deadline pressure is a cleaner filter. If I would choose a different method when the timeline is short and the stakes are real, the current recipe is already on borrowed time.
Question | Keep signal | Kill signal |
|---|---|---|
Can someone else run it from the card? | Yes | Needs original operator |
Does it still clear the probe set? | Above threshold | Chronically below |
Is the cost still worth the value? | Clear surplus | Attention sink |
Are failure modes known? | Documented | Still surprising |
Would I choose it under deadline? | Yes | Would avoid it |
What the Rubric Actually Protects
These five questions are fed by everything that came before them. Identity cards, dataset audits, strength sweeps, structured prompts, and the Monday team sequence all supply the evidence. A workflow that never accumulated those artifacts almost never survives the first question. The earlier posts in this series are not decoration. They are the measurement infrastructure that makes a keep-or-kill decision possible instead of purely emotional.
I run a light version of the five questions every month for any recipe that saw production use. Full quarterly cleanouts catch the ones that have been sitting idle. Model or host changes trigger an immediate re-probe of every active card. The cadence is not glamorous. It is the reason the Test Kitchen has never become a museum of former successes.
The hardest recipes to archive are the ones that once produced a grid I was proud of. Sentiment argues for keeping them “just in case.” Measurement makes the decision clearer. When the probe set has failed three months running and the hand-off test still requires the original operator, the card is no longer a living tool. It is nostalgia. Archiving it is not a failure. It is the only way the remaining shelf stays small enough to trust.
What “Worth Keeping” Looks Like in Practice
A kept recipe is one I would choose again under real deadline pressure, knowing its cost and its known failure modes. It is one that another person can run from the card alone. It is one whose probe pass rate is still above the threshold I set when I first published it. Anything that fails those tests is either revised until it passes or moved to the archive with a short note explaining why. That note is often more useful to future me than the original recipe ever was.
I have archived more adapters than I have published. That ratio feels healthy. A kitchen that keeps everything eventually cannot find anything worth running when the pressure is on. The archive is not a graveyard. It is the record of what we learned not to repeat.
Run the test. Save the settings. Show the failure. Then decide whether the workflow still deserves a card on the shelf. The kitchen stays useful only if it stays selective.
Twenty posts in, the pattern is stable: define the target, change one variable, record the result, admit the failure, and keep only what still earns its place. MiniMax H3 Turbo and LoRA are the current instruments. The method outlasts any single model version. If you take one habit from this site, take the refusal to confuse a good screenshot with a good system. Everything else in the Test Kitchen builds from that refusal.
No comments yet.